🧠 _

Podcast: Oxide and Friends (LS 41 · TOP 1.5% what is this?)
Episode: AI in Computer Science Education
Pub date: 2026-05-10

Get Podcast Transcript →
powered by Listen411 - fast audio-to-text and summarization

AI is an existential topic for all aspects of education--for none more so than Computer Science. Bryan and Adam were joined by Kathi Fisler and Shriram Krishnamurthi, professors of Computer Science at Brown, to discuss their experimental introductory course that strongly incorporates agentic programming. What do students take away from their "smoke the whole pack" approach?

In addition to Bryan Cantrill and Adam Leventhal, speakers included Shriram Krishnamurthi, Kathi Fisler, and Will.

Some of the topics we hit on, in the order that we hit them:

• OxF s5e29: AI in Higher Education with Michael Littman
• Greenspun's tenth rule
• Andy van Dam
• Brown CS15 Tetris
• Wesleyan Tetris
• Generating Programs Trivially: Student Use of Large Language Models
• Data-Centricity: A Challenge and Opportunity for Computing Education -- Shriram Krishnamurthi and Kathi Fisler
• LLMs â­˘ Regular Expressions, Responsibly!

If we got something wrong or missed something, please file a PR! Our next show will likely be on Monday at 5p Pacific Time on our Discord server; stay tuned to our Mastodon feeds for details, or subscribe to this calendar. We'd love to have you join us, as we always love to hear from new speakers!

The podcast and artwork embedded on this page are from Oxide Computer Company, which is the property of its owner and not affiliated with or endorsed by Listen Notes, Inc.

Podcast: Book Overflow (LS 33 · TOP 5% what is this?)
Episode: Has AI changed everything about software development? - The Mythical Man-Month by Fred Brooks
Pub date: 2026-05-18

Get Podcast Transcript →
powered by Listen411 - fast audio-to-text and summarization

In this episode of Book Overflow, Carter and Nathan discuss The Mythical Man-Month by Fred Brooks!Join the Book Overflow Discord here! https://discord.gg/ZwS2fqW7ZZ -- Want to talk with Carter or Nathan? Book a coaching session! ------------------------------------------------------------Carterhttps://www.joinleland.com/coach/carter-m-1Nathanhttps://www.joinleland.com/coach/nathan-t-2-- Books Mentioned in this Episode --Note: As an Amazon Associate, we earn from qualifying purchases.----------------------------------------------------------The Mythical Man-Month by Fred Brooks--Spotify: https://open.spotify.com/show/5kj6DLCEWR5nHShlSYJI5LApple Podcasts: https://podcasts.apple.com/us/podcast/book-overflow/id1745257325X: https://x.com/bookoverflowpodCarter on X: https://x.com/cartermorganNathan's Functionally Imperative: www.functionallyimperative.com----------------Book Overflow is a podcast for software engineers, by software engineers dedicated to improving our craft by reading the best technical books in the world. Join Carter Morgan and Nathan Toups as they read and discuss a new technical book each week!The full book schedule and links to every major podcast player can be found at https://www.bookoverflow.io

The podcast and artwork embedded on this page are from Carter Morgan and Nathan Toups, which is the property of its owner and not affiliated with or endorsed by Listen Notes, Inc.

Podcast: Book Overflow (LS 33 · TOP 5% what is this?)
Episode: Your Boss Has Read This Book - Radical Candor by Kim Scott
Pub date: 2025-09-15

Get Podcast Transcript →
powered by Listen411 - fast audio-to-text and summarization

In this episode of Book Overflow, Carter and Nathan discuss the first half of Radical Candor by Kim Scott! Join them as they discuss what makes a good boss, what makes up good feedback, and more!

-- Want to talk with Carter or Nathan? Book a coaching session! --

----------------------------------------------------------

Carter

https://www.joinleland.com/coach/carter-m-1

Nathan

https://www.joinleland.com/coach/nathan-t-2

-- Books Mentioned in this Episode --

Note: As an Amazon Associate, we earn from qualifying purchases.

----------------------------------------------------------

Radical Candor

https://amzn.to/3HXRukj

----------------

00:00 Intro

02:07 About the Book and Author

03:06 Initial Thoughts on Radical Candor Part 1

13:00 The Four Quadrants Framework

21:26 Ruinous Empathy - The Career Destroyer

32:39 Bringing Your Whole Self to Work

46:36 Rock Stars vs Superstars

1:03:18 Hot Takes

1:08:49 Final Thoughts and Recommendations

----------------

Spotify: https://open.spotify.com/show/5kj6DLCEWR5nHShlSYJI5L

Apple Podcasts: https://podcasts.apple.com/us/podcast/book-overflow/id1745257325

X: https://x.com/bookoverflowpod

Carter on X: https://x.com/cartermorgan

Nathan's Functionally Imperative: www.functionallyimperative.com

----------------

Book Overflow is a podcast for software engineers, by software engineers dedicated to improving our craft by reading the best technical books in the world. Join Carter Morgan and Nathan Toups as they read and discuss a new technical book each week!

The full book schedule and links to every major podcast player can be found at https://www.bookoverflow.io

The podcast and artwork embedded on this page are from Carter Morgan and Nathan Toups, which is the property of its owner and not affiliated with or endorsed by Listen Notes, Inc.

Podcast: Stellar Work
Episode: Engineering Management In The Age Of AI
Pub date: 2026-04-27

Get Podcast Transcript →
powered by Listen411 - fast audio-to-text and summarization

Most engineering teams using AI are about 2x faster. Not 10x. The bottleneck moved, but nobody optimized the rest.Jeff Lee-Chan spent 10 years at Google working on YouTube, then seven years at Snapchat. He went from IC to staff engineer to engineering manager. Now he spends 20 to 30 extra hours a week experimenting with AI tools outside his day job.In this conversation, you'll hear:Why a 30-minute AI side project convinced him the world had changedWhat the actual AI stack looks like inside Big Tech right now (it's mostly Claude Code, Cursor, and Codex)Where the real bottleneck sits when coding speed isn't the problem anymoreWhy he thinks 4-5 custom code review bots beat a single default oneHow AI is shrinking the junior engineering pipeline and what that meansThe one mistake he made early in management that he'd warn every new manager aboutJeff Lee-Chan on LinkedIn: https://www.linkedin.com/in/jeffrey-lee-chan/Mentoring with Jeff: https://mentorcruise.com/mentor/jeffreylee-chan/More episodes & all platforms: https://stellarwork.start.pageNewsletter: https://substack.com/@stellarworkHost: Ben, founder of Stellar Work

The podcast and artwork embedded on this page are from Benjamin Igna, which is the property of its owner and not affiliated with or endorsed by Listen Notes, Inc.

Podcast: The Work Item - Real Talk on Tech's Toughest Career Choices (LS 27 · TOP 10% what is this?)
Episode: #94 - Exploiting Your Unfair Advantages - Shirley Wu (Shirley Wu Studio)
Pub date: 2025-10-20

Get Podcast Transcript →
powered by Listen411 - fast audio-to-text and summarization

Shirley Wu is back-to-back on The Work Item - we previously talked about her unexpected career in data visualizations, but for this episode we're switching things up a little bit and instead focus on building long-term career advantages based on Shirley's experience. This is an especially important topic in the era of AI, where folks have a lot of uncertainty about their career tracks and what it means to build durable moats that can survive the industry being upended by new tools and approaches to getting things done.

Shirley's experience is particularly relevant here as an independent studio owner - she's someone who has years of experience to lean on flying solo and seeing how one can establish their own reputation and image in the space.

You can find Shirley on the following sites:

🎨 Shirley Wu Studio

🦋 Bluesky

💼 LinkedIn

📸 Instagram

The podcast was produced by Den Delimarsky.

Feedback

If you haven't already, make sure to subscribe to the show and leave a review or a rating, wherever you are getting your podcast. I really appreciate your feedback and am working to make this podcast more useful for you, the listener, with every episode. Ratings and feedback make it so others can easily discover and enjoy the insights you listen to here!

The podcast and artwork embedded on this page are from Den Delimarsky, which is the property of its owner and not affiliated with or endorsed by Listen Notes, Inc.

Podcast: Developers! - mer än bara kod
Episode: 272. Vad händer när man tappat gnistan?
Pub date: 2026-04-23

Get Podcast Transcript →
powered by Listen411 - fast audio-to-text and summarization

En lyssnare har kodat sedan högstadiet, programmering var hans grej – men nu är gnistan borta. Han undrar om det är AI som tagit något ifrån honom, småbarnsåren, eller om han helt enkelt växt ifrån det.

Vi pratar om vad som händer med motivationen när verktygen börjar lösa problemen åt dig, om "kod som personlighet" och vad som händer med självbilden när den identiteten rubbas, och om det faktiskt är okej att bara… inte vara en sådan där som älskar kod längre.

Vi tittar ocksü pü färsk data frün Gallup om hur AI-adoption faktiskt ser ut bland Gen Zs och pü Gallerix-skandalen där AI-genererade motiv lades ut fÜr fÜrsäljning utan att konstnärerna visste om det.

🤓 Svårighetsnivå: 2/5

🔗 Länkar:

Gallup: Gen AI Adoption Steady, Skepticism Climbs

SVT: Efter AI-kritik – Gallerix tar bort 50 motiv

💬 Ställ en anonym fråga eller insändare som vi kan ta upp i podden!

💌 Håll kontakten med oss:

Discord

Instagram

Facebook

[email protected]

https://www.developerspodcast.com

Om du gillar podden für du gärna stÜtta oss genom att kÜpa vür merch, bli en Patreon, subscriba till podden eller skriva en recension!

★ Support this podcast on Patreon ★

The podcast and artwork embedded on this page are from Madeleine SchĂśnemann och Sofia Larsson, which is the property of its owner and not affiliated with or endorsed by Listen Notes, Inc.

Podcast: The Growth Podcast (LS 38 · TOP 2% what is this?)
Episode: How to Become a "Builder PM" with n8n, Claude Code, and OpenClaw | Mahesh Yadav (ex-Google, AWS, Meta, Microsoft; Founder LegalGraph AI)
Pub date: 2026-04-20

Get Podcast Transcript →
powered by Listen411 - fast audio-to-text and summarization

Today’s episode

LinkedIn just changed the title of its product managers to product builders.

What does it even mean to be a “builder PM”?

Well, tools only get you so far. Learning Claude Code is helpful, but means nothing if you don’t have an understanding of the underlying first principles.

That’s today’s episode.

Mahesh Yadav created one of our most popular episodes, with over 35K views on YouTube, and now he’s back. Earlier, he taught you AI agents. Today, he’s touching you how to become a builder PM:

If you want access to my AI tool stack - Dovetail, Arize, Linear, Descript, Reforge Build, DeepSky, Relay.app, Magic Patterns, Speechify, and Mobbin - grab Aakash’s bundle.

I’m giving a free talk on how to get interviews at the top AI PM companies on Thursday, April 23rd 2026 @ 9:00AM PDT. Grab your seat.

----

Check out the conversation on Apple, Spotify, and YouTube.

Brought to you by:

* Maven - Build cohort-based courses that scale

* Amplitude - The market leader in product analytics

* Jira Product Discovery - Prioritize what matters with confidence

* NayaOne - Airgapped cloud-agnostic sandbox to validate AI tools faster

* Product Faculty - Get $550 off their #1 AI PM Certification with my link

----

Key Takeaways:

1. Builder PM defined - A builder PM talks to customers, figures out what to build, and ships the first version to 10 customers without talking to any developer. The skill is knowing what to build, not knowing how to code.

2. Four agent components - Every agent that works has intelligence (model), tools (actions), memory (session context), and knowledge (your company data). Every agent that disappoints is missing at least one.

3. n8n for foundations - n8n is the best learning tool because you visually see every component of the agent architecture as separate nodes. Build your first multi-agent system and evaluation pipeline here.

4. Claude Code ate three company types - Context companies, action companies, and evaluation companies all got replaced by one agentic loop inside Claude Code. The three pieces collapsed into one tool.

5. Computer control is the real unlock - File system access plus bash commands equals full laptop capability. This is why Claude Code went from coding tool to work operating system.

6. Long-horizon jobs changed the game - AI agents went from 3-minute tasks to 3-6 hour sustained jobs in six months. This turns Claude Code from assistant to autonomous worker.

7. Continuous learning loops - Build a second agent that watches your corrections to the first agent's work. After five repeated patterns, it proposes a skill update. Your tools get better every day.

8. OpenClaw pattern - Delegation through existing channels, full machine sandboxing, model-agnostic. Not a product but a pattern that Google and AWS will copy inside their ecosystems.

9. AI PM interviews changed - At L5 and L6, product sense questions are being replaced with live building exercises and system design for AI architectures. Pull out Claude Code during the interview or you are already out.

10. Compensation trajectory - From $120K at Microsoft to $1.3M at Google over 13 years, doubling every 18 months through AI-focused switches. Left because big companies kill innovation with six-week approval cycles.

----

Where to find Mahesh Yadav

* LinkedIn

* Maven Course

Related content

Podcasts:

* Claude Code Team OS with Carl Vellotti

* OpenClaw + Claude Code with Naman Pandey

* Claude Code OS with Dave Killeen

Newsletters:

* The complete context engineering guide

* How to use Claude Code like a pro

* Practical AI agents for PMs

----

PS. Please subscribe on YouTube and follow on Apple & Spotify. It helps!

If you want to advertise, email productgrowthppp at gmail.

This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit www.news.aakashg.com/subscribe

The podcast and artwork embedded on this page are from Aakash Gupta, which is the property of its owner and not affiliated with or endorsed by Listen Notes, Inc.

Podcast: Latent Space: The AI Engineer Podcast (LS 45 · TOP 1% what is this?)
Episode: Marc Andreessen introspects on The Death of the Browser, Pi + OpenClaw, and Why "This Time Is Different"
Pub date: 2026-04-03

Get Podcast Transcript →
powered by Listen411 - fast audio-to-text and summarization

Fresh off raising a monster $15B, Marc Andreessen has lived through multiple computing platform shifts firsthand, from Mosaic and Netscape to cofounding A16z.

In this episode, Marc joins swyx and Alessio in a16z’s legendary Sand Hill Road office to argue that AI is not just another hype cycle, but the payoff of an “80-year overnight success”: from neural nets and expert systems to transformers, reasoning models, coding, agents, and recursive self-improvement. He lays out why he thinks this moment is different, why AI is finally escaping the old boom-bust pattern, and why the real bottleneck may be less about models than about the messy institutions, incentives, and social systems that struggle to absorb technological change.

This episode was a dream come true for us, and many thanks to Erik Torenberg for the assist in setting this up. Full episode on YouTube!

We discuss:

* Marc’s long view on AI: from the 1980s AI boom and expert systems to AlexNet, transformers, and why he sees today’s moment as the culmination of decades of compounding technical progress

* Why “this time is different”: the jump from LLMs to reasoning, coding, agents, and recursive self-improvement, and why Marc thinks these breakthroughs make AI real in a way prior cycles were not

* AI winters vs. “80-year overnight success”: why the field repeatedly swings between utopianism and doom, and why Marc thinks the underlying researchers were mostly right even when the timelines were wrong

* Scaling laws, Moore’s Law, and what to build: why he believes AI scaling laws will continue, why the outside world is messier than lab purists assume, and how startups can still create durable value on top of rapidly improving models

* The dot-com crash and AI infrastructure risk: Marc’s comparison between today’s AI capex boom and the fiber/data-center overbuild of 2000, plus why he thinks this cycle is different because the buyers are huge cash-rich incumbents and demand is already here

* Why old NVIDIA chips may be getting more valuable: the pace of software progress, chronic capacity shortages, and the idea that even current models are “sandbagged” by supply constraints

* Open source, edge inference, and the chip bottleneck: why Marc thinks local models, Apple Silicon, privacy, trust, and economics all point toward a major role for edge AI

* American vs. Chinese open source AI: DeepSeek as a “gift to the world,” why open models matter not just because they’re free but because they teach the world how things work, and how open source strategies may shift as the market consolidates

* Why Pi and OpenClaw matter so much: Marc’s claim that the combination of LLM + shell + filesystem + markdown + cron loop is one of the biggest software architecture breakthroughs in decades

* Agents as the new “Unix”: how agent state living in files allows portability across models and runtimes, and why self-modifying agents that can extend themselves may redefine what software even is

* The future of coding and programming languages: why Marc thinks software becomes abundant, why bots may translate freely across languages, and why “programming language” itself may stop being a salient concept

* Browsers, protocols, and human readability: lessons from Mosaic and the web, why text protocols and “view source” mattered, and how similar principles may shape AI-native systems

* Real-world OpenClaw use: health dashboards, sleep monitoring, smart homes, rewriting firmware on robot dogs, and why the most aggressive users are discovering both the power and danger of agents first

* Proof of human vs. proof of bot: why Marc thinks the internet’s bot problem is now unsolvable via detection alone, and why biometric + cryptographic proof of human becomes necessary

Timestamps

* 00:00 Marc on AI’s “80-Year Overnight Success”

* 00:01 A Quick Message From swyx

* 01:44 Inside a16z With Marc Andreessen

* 02:13 The Truth About a16z’s AI Pivot

* 03:29 Why This AI Boom Is Not Like 2016

* 06:33 Marc on AI Winters, Hype Cycles, and What’s Different Now

* 10:09 Reasoning, Coding, Agents, and the New AI Breakthroughs

* 12:13 What Founders Should Build as Models Keep Improving

* 16:33 AI Capex, GPU Shortages, and the Dot-Com Crash Analogy

* 24:54 Open Source AI, Edge Inference, and Why It Matters

* 33:03 Why OpenClaw and PI Could Change Software Forever

* 41:37 Agents, the End of Interfaces, and Software for Bots

* 46:47 Do Programming Languages Even Have a Future?

* 54:19 AI Agents Need Money: Payments, Crypto, and Stablecoins

* 56:59 Proof of Human, Internet Bots, and the Drone Problem

* 01:06:12 AI, Management, and the Return of Founder-Led Companies

* 01:12:23 Why the Real Economy May Resist AI Longer Than Expected

* 01:15:53 Closing Thoughts

Transcript

Marc: Something about AI that causes the people in the field, I would say, to become both excessively utopian and excessively apocalyptic. Having said that, I think what’s actually happened is an enormous amount of technical progress that built up over time. And like for, for example, we now know that neural network is the correct architecture.And I, I will tell you like there was a 60 year run where that was like a, you know, or even 70 years where that was controversial. And so, so the way I think about what’s happening is basically, I think, I think about basically the, the, the period we’re in right now is it’s, I call it 80 year overnight success, right?Which is like, it’s an overnight success ‘cause it’s like bam, you know, chat GPT hits and then, and then oh one hits, and then, you know, open claw hits and like, you know, these are open, these are, these are like overnight, like radical, overnight transformative successes, but they’re drawing on an 80 year sort of wellspring backlog, you know, of, of, of, of ideas and thinking it’s not just that it’s all brand new, it’s that it’s an unlock of all of these decades of like very serious, hardcore research.If I were 18, like this is a hundred, this is what I would be spending all of my time on. This is like such an incredible conceptual breakthrough.swyx: Before we get into today’s episode, I just have a small message for listeners. Thank you. We will not be able to bring you the ai, engineering, science, and entertainment contents that you so clearly want if you didn’t choose to also click in and tune into our content.We’ve been approached by sponsors on an almost daily basis, but fortunately enough of you actually subscribed to us to keep all this sustainable without ads, and we wanna keep it that way. But I just have one favor to ask all of you. The single, most powerful, completely free thing you can do is to click that subscribe button.It’s the only thing I’ll ever ask of you, and it means absolutely everything to me and my team that works so hard to bring the in space to you each and every week. If you do it, I promise you will never stop working to make the show even better. Now, let’s get into it.Alessio: Hey everyone, welcome to the Lidian Space Pockets. This is CIO, founder Kernel Labs, and I’m joined by s Swix, editor of Lidian Space.swyx: Hello. And we’re in a 16 Z with a, uh, mark G and welcome.Marc: Yes, yes. A and what, half of 16? Something like that. A one. Exactly,swyx: exactly. Uh, apparently this is the, the final few days in your, your current office.You’re moving across the road.Marc: Uh, we’re, yeah. We have a, we have some, we have some projects underway, but yeah, this is actually, oh, this is the original. We’re in actually the original office. We’re in the, we’re in the, we’re, we’re in the whole thing.swyx: It’s beautiful. Yeah. Great.Marc: Thank you.swyx: So I have to come out, uh, this is a, you know, I wanted to pick a spicy start in October, 2022.I just made friends with Roone and, uh, I wanted to give him something to sort of be spicy about. And I said, uh. Uh, it’ll never not be funny. The A 16 Z was constantly going. The future is where the smart people choose to spend their time and then going deep into crypto and not in ai. And that was in October 22nd, 2022.And Ruen says there was an internal meeting in a 16 Z to reorient around Gen ai. Obviously you have, but was there a meeting? What, what was that?Marc: I mean, I don’t, look, I’ve been doing AI since the late eighties.swyx: Yeah.Marc: So I, I don’t know, like all that, as far as I’m concerned, this stuff is all Johnny cum lately.Yeah. You, I mean, look, we’ve been doing ar entire existence. I mean, we’ve been doing AI machine learning deep, you know, deeply. We’ve been doing this stuff way from the beginning. Obviously a AI is just core to computer science. I, I, I actually view them as like quite, uh, quite continuous. Um, you know, Ben and I both have computer science degrees.Um, you know, we, we both, Ben, Ben and I actually both are world enough to remember the actual AI boom in the 1980s. Yeah. There was like a, there was a big AI boom at the time. Um, and there was a, was names like expert systems. Um, and they of like lisp and lisp machines. Uh, I, I coded in lisp. I was coding a lisp in 1989.When that was the, the language of the AI future. Um, yeah. So this is something that we’re like completely, you completely comfortable with. I’ve been doing the whole time and are very enthusiastic aboutswyx: is there a strong, like this time is different because, uh, my closest analog was 20 16 17. It was an AI boom.Mm-hmm. And it petered out very, very quickly. Um, we, it just, it just in terms of investingMarc: sort of, sort of,swyx: yeah. Investment, investment excitement.Marc: Although that’s really when the, the, the Nvidia phenomenon really, it was, I would say it was in that period when it was very clear that at, at the time it, the vocabulary was more machine learning, but it, it was very clear at that time that machine learning was hitting some sort of takeoff point.Alessio: Yeah.Marc: Well, and as you guys, you guys have talked about this at length on, on your thing, but, you know, if you really track what happened, I think the real story is, it was, it was the Alex net, uh, basically breakthrough in like 2013. That was the, that was the real knee in the curve. Um, and then it was obviously the transformer breakthrough in 17.Alessio: Yeah.Marc: Um, and then everything that followed. But, but, you know, look, machine learning, you know, there were, you know, look, uh, I mean look, I’ve been working, you know, I’ve been working with, uh, one of my, you know, kind of projects working with Facebook since 2004. Um, and on the board since 2007, and of course, you know, they, they started using machine learning very early, um, and, you know, have used it basically, you know, for like 20 years for, you know, content, you know, feed optimization and advertising optimization.And obviously many, you know, financial services. You know, many, many, many companies, many different sectors have been doing this. And so it’s like one of these things, it’s like, it’s not a, it’s not a single thing. Like it’s, it’s like, it’s like layers, right? Yeah. Um, and, and the layers arrive at different paces and, but they kind of build up.swyx: Yeah.Marc: Uh, they kind of build up over time and then, and then, yeah. And then look, in retrospect, it was 2017 was kind of the, you know, the key, the key point with the trans transformer and then. And then as you guys know, there was this really weird like four year period where it’s like the, the transformer existed and then it was just like,swyx: let’s go.Yeah.Marc: Well, but, but it was just, but, but between 2020, but between 2017 and 2021, I mean, that was the era of which like companies like Google had internal chat Botts, but they weren’t letting anybody use them.swyx: Yeah.Marc: Right. And then, you know, and then OpenAI developed Chat GT or GPT two, and then they told everybody, this is way too dangerous to deploy.Right. Yeah. You know, we can’t possibly let normal people, normal people use this thing. And then you, you guys, I’m sure remember AI Dungeon, um mm-hmm. So the o for, there was like a year where like the only way for a normal person to use GP T three was in, in AI dungeon.Alessio: Yeah.Marc: And so you, you, we would do this, you’d go in there and you’d pretend to play Dungeons and Dragons.In reality, you’re just trying to talk to talk to GPT. And so there was this, you know, there was this long, you know, and I, you know, the big, big companies, you know, big companies are cautious and, you know, the big companies were cautious. It, it, by the way, it took open ai. You know, they, they, they talk about this, it took open AI time to actually adjust, you know, kind of re redirect their researchswyx: path.I, I think, uh, let say Rosewood, right? Uh, the, the dinner that founded OpenAI was right there.Marc: Right, right. But that, that dinner would’ve taken place in 20swyx: 18Marc: 19. The formation of OpenAI Uhhuh as late as 2018.swyx: Uh, uh, sorry. Uh, no, I’m, I’m, I’m, I’m wrong. Probably It should be 20. Yeah. They just celebrated a 10 year anniversary, so it it is 2025.Yeah, so, so 2015?Marc: Yeah. 2015. Yeah. 2015. But then, uh, um, Alec Radford did G PT one in what, probablyswyx: mm-hmm. 17, 18,Marc: yeah. 17, 18. So it, yeah. For, and then, and then they didn’t really, and then GPT three was what? 2020? 2020.swyx: 2020.Marc: Because that became copilot immediately. Even open ai, which has been, you know, the leader of, of this thing in the last decade, you know, e even they had to adapt and, and, and lean into the new thing.And so. Um, yeah, I, I think it’s just this process of basically sort of wave after wave layer after layer, you know, building on itself. And then you kind of get these catalytic moments where, where the whole thing pops and, and obviously that’s what’s happening now.swyx: Is it useful to think about will there be any ai, winter?‘cause there’s always these patterns. Like, is this, in the summer is something I constantly think about because do I get, do I just like. Just get endlessly hyped and just trust that I will only be early and never wrong or right. Well, are we, will there be a winter?Marc: So there’s something about, say the following.There’s something about AI that has led to this repeated pattern. Um, and, and, and you guys know this,swyx: it’s summer, winter, summer,Marc: winter, summer, winter, summer, winter. And it goes back 80 years. Yeah. 80 years. Uh, so the original neural network paper was 1943. Right. Which is, which is amazing. Uh, that it was, it was far back that long.And then there was you, if you guys have ever talked about this on your show, but there was this, uh, there was a big, uh, there was an a GI conference at Dartmouth University in 1950. 55. 55, yeah. And they got a NSF grant to, uh, for the, all the AI experts at the time to spend the summer together. And they figured if they had 10 weeks together, they could get a GI, uh, at the other end.And they got their, by the way, they got the grant, they got the 10 weeks and then, you know, 1955, you know. No, no. A GI. And like I said, I, I lived through the eighties version of this where there was a big, a big boom and a crash. And so, so there is this thing, and there, there is something about AI that causes the people in the field, I would say, to become both excessively utopian and excessively apocalyptic.Um, and, and it’s probably on both sides of like the, the, the boom bus cycle. You, you kind of see that play out. Having said that, I think what’s actually happened is like just, and you know, and we now know in retrospect like an enormous amount of technical progress that built up over time. And like for, for example, we now know that neural network is the correct architecture.And I, I will tell you like there was a 60 year run where that was like a, you know, or even 70 years or that was controversial. And, and we now know that that’s the case. And so we, we now, you know, everything we’re building on today just sort of derives from the original idea in 1943. And so, so in retrospect, we, we now know that like, these, these guys are right.They, they, you know, they would get the timing wrong and they thought, you know, capabilities would arrive faster, or they were, it could be turned into businesses sooner or whatever, but like, they were fundamentally, the, the scientists who worked on this over the course of decades were fundamentally correct about what they were doing.And, and the, and the payoff from, from, from all their work is happening now. And so, so the way I think about what’s happening is basically, I think, I think about basically the, the, the period we’re in right now is it’s, I call it 80 year overnight success, right? Which is like, it’s an overnight success.‘cause it’s like bam, you know, chat, GPT hits and then, and then oh one hits, and then, you know, open claw hits and like, you know, these are open, these are, these are like overnight, like radical, overnight transformative successes, but they’re drawing on an 80 year sort of wellspring backlog, you know, of, of, of, of ideas and thinking it’s not just that it’s all brand new, it’s that it’s an unlock of all of these decades of like very serious, hardcore research.Um, and thinking, and look, there were AI researchers who spent their entire lives. They got their PhD. They, they worked for, they’ve researched for 40 years. They retired in a lot of cases, they passed away and they never actually saw it work.swyx: Yeah. It’s all sad.Marc: It is. It is sad. It’s sad. Knewswyx: Jeff Hinton was like the last guy.Marc: Yeah. Yeah. Well, there were the guys, uh, was a guy, Alan Newell. I mean, there’s tons of John McCarthy. You know, John McCarthy was like one of the inventors in the field. He’s one of the guys who organized the Dartmouth Conference and you know, he taught at Stanford for 40 years. Wow. And passed, you know, passed away, I don’t know, whatever, 10, 10 years ago or something.Never, never actually go. Got to see it happen. But like, it is amazing in retrospect, like, these guys were incredibly smart and they worked really hard and they were correct. So anyway, so then it’s like, okay, you know, say history doesn’t repeat, but it rhymes. It’s like, okay, does that mean that there’s gonna be another, like, you know, basically boom buzz cycle.And I, I will tell you, like, let, like in a sense, like yes, everything goes through cycles and, you know, people get overly enthusiastic and overly depressed and there’s, there’s a time, there’s a timelessness to that. Having said that, there’s just no question. Um, so the form, the foremost dangerous words in investing this time are, this time is different.Do you know the 12 most dangerous words investing? No. The four most d foremost dangerous words in investing are this time is different. Yeah. Um, the 12 most dangerous words. And so like, I’ll tell you what’s different. Like now it’s working like, like there’s just no, I mean, look, there’s just no question.And by the way, I, I’ll just give you guys my take. Like L LLMs, like from, from basically the Chad G PT moment through to spring of 25. I think you could still, I think well intention, well, and of. Form skeptics could still say, oh, this is just pattern completion. And oh, these things don’t really understand what they’re doing.And you know, the hall hallucination rates are way too high. And, you know, this is gonna be great for creative writing and creating, you know, Shakespeare and so sonnets and, you know, as, as rap lyrics or whatever, like, it’s gonna be great and all that stuff, but we’re not gonna be able to harness this to make this relevant in, you know, coding or in medicine or in law or in, you know, you know, kind of feels that, you know, kind of really, really matter.And I think basically it was the reasoning breakthrough. It, it was oh one and then R one that basically answered that question basically said, oh no, we’re gonna be able to actually turn this into something that’s gonna work in the real world. And, and then obviously the coding breakthrough over the, over basically the coding breakthrough that kind of catalyzed over the holiday break was kind of the third step in that.Mm-hmm. Where you’re just like, alright, if, if, you know, if Linus Tova is saying that the AI coding is no better than he is like. Like, that’s, that’s never happened before. That’s theswyx: benchmark.Marc: Yeah. That’s never happened before. And so now we know that it’s, it’s gonna sweep through coding and, and then, and then we, we know, you know, we know that if it’s gonna work in coding, it’s gonna work in everything else.Right. It’s just then, because that’s, that’s like, that’s like, that’s like the hardest in many ways. That’s the hardest example. And how everything else is gonna be a, a derivative of that. And then on top of that, we just got the agent breakthrough, you know, with Open Claw, which is fantastic. Which is amazing and incredibly powerful.And then we just got the, the, um, the auto research, uh, you know, the, the self-improvement. You know, we’re now into the self-improvement breakthrough. And so the, so the way I think about it is we’ve had four fundamental breakthroughs in functionality, l OMS reasoning, uh, agents, um, and then, uh, and, and then now RSI, um, and, and they’re all actually working.Um, and so I’m, I’m just, as you like, you can tell I’m jumping outta my shoes. Like, like this is, like this is it like this, this is the culmination of 80 years worth of worth of work, and this is the time it’s becoming real.Alessio: Yeah.Marc: I, I’m completely convinced.Alessio: I think the anxiety that people feel is like during the transistor era, yet Mors law, and it’s like, all right, we understand why these things are getting better.We understand the physics of it. Yeah. With ai, it’s. It’s so jagged in like the jumps where like, like you said, it’s like in three months you have like this huge jump like, and people are like, well this can keep happening. Right? But then it keeps happening,Marc: it’ll keep happening.Alessio: And so like how do you think about also timelines of like what’s we’re building?I think we always have this question with guests, which is like, you know, should you spend time building harness for a model versus like the next model just gonna do it one shot in the lead space. Right. And how does that inform, like how you think about the shape of the technology? You know, you talk about how it’s a new computing platform.If you have a computing platform, then like every six months it like drastically changes in what it looks like. It’s hard to build companies on top of it.Marc: Yeah. So, so a couple things. So one is like, look, the, the Moore’s law was what we now call a scaling law. Like Moore’s Law was a scaling law and for your younger viewers, more Moore’s Law was every chip chip chips either get twice as powerful or twice as cheap every, every 18 months.And that, and that and that, you know, that it’s gotten more complicated in the last few years. But like that, that was like the 50 year trajectory of, of, of the computer industry. And then, and then by the way, and that’s what took the mainframe computer from a $25 million current dollar thing into, you know, the phone in your pocket being, you know, a million times more powerful than that.Like that, you know, for, for 500 bucks. And so that, that was a scaling law. And then, and then, and then key to any scaling law, including Moore’s Law and the AI scaling laws is, you know, they’re not really laws, right? They’re, they’re, they’re, they’re predictions, but when they work, they become self-fulfilling predictions because they, they, they, they, they set a benchmark and, and then the entire industry, right?All the smart people in the industry kind of work to make sure that, that, that actually happens. And so they, they kind of motivate the breakthroughs that are required to, to keep that going. And, and in and in chips, that was a 50 year, that was a 50 year run. Right. And it, it was amazing. And it’s still happening in, in some areas of, of chips.I think the same thing is happening with the, the core scaling laws. The core scaling laws. In, in, in ai, you know, they’re, they’re not really laws, but like they, they are basically. There are predictions and then they’re motivating catalysts for the research work that is required to be. And, and, and, and by the way, also the investment, uh, dollars, um, uh, you know, required to basically keep, you know, keep the curves going and, and look, it, it is, it’s gonna be complicated and it’s gonna be variable and they’re, you know, there’re gonna be walls that are gonna look like they’re fast approaching, and then they’re gonna be, you know, engineers are gonna get to work and they’re gonna figure out a way to punch through the walls.And obviously that’s, you know, that’s been happening a lot, you know, and then look, there’s gonna be times when it looks like the walls have, you know, the, the, the laws have petered out and then they’re gonna, they’re gonna pick up again and surge and then, and then, and then it, it appears what’s happening to the eyes is there’s not multiple, you know, multiple scaling laws.Um, there’s multiple areas of improvement. And, and I think, you know, I don’t know how many more there are already yet to be discovered, but there are probably some more that we don’t know about yet. You know, they, like, for example, there’s probably some scaling law around, um, world models and robotics that we don’t fully understand, you know, kind of acquisition of data at scale in the real world that we don’t fully understand yet.So that, that, that one will probably kick in at some point here. There’s a bunch of really smart people working on that. Um, and so, yeah, I, I think the expectation is that, that, you know, the, the scaling laws generally are gonna continue. Yeah. The, the pace of improvement will continue to move really fast.Um. To your question on like what to build. So, uh, I’m a complete believer the scaling laws are gonna continue. I’m a complete believer the capabilities are gonna keep getting amazing, um, you know, leaps and bounds. Uh, the part where I kind of part ways a little bit with how, what I would describe as the AI purists, um, you know, which is, which I would characterize as like the people who are.In many ways, the smartest people in the field, but also the people who spend their entire life, like at a lab, um, and have, have, I would say, have very little experience in the outside world. Um, the, the, the nuance I would offer is the outside world of 8 billion people and institutions and governments and companies and economic systems and social systems is really complicated.Um, and, um, and doesn’t, you know, it it 8 billion people making collective decisions on planet Earth is not a simple process of like, just like you see this happening now. It’s like a bunch of AI CEOs have this thing, which is just like, well, there’s just this, they just all have this kind of thing when they talk in public where they’re just like, well, there’s these, these obvious set of things that so society to do.Alessio: Mm-hmm.Marc: And then they’re like, society’s not doing any of those things. Right. And it’s like, how can society not, you know, what, whatever their theory is, how can society not see x, y, Z? Mm-hmm. And the answer is, well, society is number one. There’s no single society, it’s like 8 billion people. And they like all have a voice, and they all have a vote, like at the end of the day of how they, they react to change.And then, you know, it just like, it’s just human reality is just really complicated and messy. Um, and, and, and so the specific answer to your question is like, as usual, it depends. Um, you know, it, it depends. Look, pe there’s no question people are gonna, like, there’s no question they’re gonna be companies.It’s already happening. There are companies that think that they’re building value on top of the models and then they’re just gonna get blissed by the, by the next model. There’s no question that’s happening. But I think there’s no question also that just the process of adaptation of any technology into the real and into the real messy world of humanity is, is just going to be messy and complicated.It’s, it’s not going to be simple and straightforward. It’s gonna be messy and complicated. And there are gonna be a lot of companies and a lot of products, um, uh, and in, in fact entire industries that are gonna get built to, to, to basically actually help all of this technology actually reach real people.Alessio: The amount of capital going into these companies, I mean, Dario talked about it on the Door Cash podcast and Door Cash was like, why don’t you just buy 10 x more GPUs? And he is like, because I’m gonna go bankrupt if the model doesn’t exactly hit the, the performance level. How do you think about that?Also as a risk on, you know, you guys are investors, open AI and thinking machines and world apps. It seems like we’re leveraging the scaling loss at a pretty high rate, right? Like how comfortable, I guess, do you feel with the downside scenario, like, and say like things Peter out, you think you can kind of like restructure uh, these build outs and uh, you know, capital investments.Marc: Yeah. So should start by saying, so I live through the.com crash, um, and I can tell you stories for hours about the.com crash and it was horrible. No, it was awful. It was, it was, it was apocalyptic by the way. The, a lot of the.com crash was actually at the time, it was actually a telecom crash. It was a bandwidth crash.Like the, the thing that actually crashed, that wiped out all the money with the tele, the telecom companies.swyx: GlobalMarc: crossing. Global, global, yeah.swyx: I’m from Singapore and they, they laid so much cable o over over our oceans.Marc: Actually there was a scaling law in the.com. Era. And it was literally the, the US Commerce Department put out a report in 1996 and they said internet traffic was doubling every quarter.Um, and, and actually in 1995 and 1996, internet traffic actually did double every quarter. And so that became the scaling law. And so what all these telecom entrepreneurs did was they went out and they raised money to build fiber, anticipating that the demand for bandwidth is gonna keep doubling every quarter.Doubling every quarter though is like, you know, grains of chess and the chessboard, like at some point the numbers become extremely large. Right. And, and, and it really, and really what happened was the internet. The internet by the way, continuously kept growing basically since inception. And it’s, you know, it’s, it’s continuously grown.It’s never shrunk. And it’s grown really fast compared to anything else. Mm-hmm. You know, in, in, in human history. But it wasn’t doubling every quarter as of 19 98, 19 99. And so there was this gap in the expectation of what they thought was a scaling law versus reality. And that’s actually what caused the.com crash, which was the, it they, they way over companies like global crossing way overbuilt fiber, which is sort of the, and by the way, fiber, telecom equipment, you know, so all the, all the networking gear, you know, and then, and then by the way, the actual physical data centers, like that was the beginning of the, of the, of the data center build and then, and the data center overbuild.And so you had that, but it was, it was literally, I think it was like $2 trillion got wiped out, right? It was like Jesus, it was like a big, it was. And by the way, the other, the other subtlety in it was the internet companies themselves never really had any debt. ‘cause tech, tech companies generally don’t run on debt, but the telecom companies run on debt.Physical infrastructure companies run on debt. And so the companies like Global Crossing not just raise a lot of equity, they also raise a lot of debt. So they’re highly levered. And so then you just do the thing. It’s just like, okay, you have a highly levered thing where you’re, you’re just over, you’re overbuilding capacity.Demand is growing, but not as fast as you hoped. And then boom, bankrupt. Right. And, and then it, and then it’s like they say about the hotel industry, which is, it’s always the third owner of a hotel that makes money. It has to go bankrupt twice, right? You have to wash out all of the over optimistic exuberance before it gets to actually a stable state.And then it makes money. So by the way, all of those data centers and all of those, all the fiber that they’re in use, it’s all in use today. Yeah. But 25 years later. But it, it, it took, and actually the elapsed time was, it took 15 years. It took 15 years from 2000 to 2015 to actually fill, fill up all that capacity.The cautionary warning is the, the overbuild can happen. Um, and, and, and, and, you know, you, you get into this thing where basically everybody, everybody who basically has any sort of institutional capital, it’s like, wow. It’s just, I, I don’t know how to invest in these crazy software things. For sure I can put build data centers and for sure I can buy GPUs that I can deploy, you know, compute grids and, and all these things.Um, and so, you know, if you’re a pessimist, you could look at this and you could say, wow, this is like really set up to be able to basically replicate, you know, what we went through, what we went through in 2000. Obviously that would be bad. The counter argument, which is the one I I agree with, which is the counter on, on the other side is a couple things.One is the companies that are investing all the, the companies that are investing the money are like the bluest chip of companies. And so back, back, back in the, in the do, like Global Crossing was like a, it was like an entrepreneur. It was like a, a new venture, but like the money that’s being deployed now at scale is Microsoft, and, you know, and Amazon and Google, Facebook and Facebook and Nvidia and, you know, these, these, these, and, and now you know, by the way, open ai philanthropic, which are now at like, you know, really serious size, um, you know, as companies with, you know, very serious revenue.These are very large scale companies with like, lots, lots of cash, lots of debt capacity that they’ve, they’ve never used. And so th this is institutional in a way that, that really wasn’t at the time. And then the other is, at least for now, every dollar that’s being put into anything that results in a running GPU is being turned into revenue right away.Like so, and you guys know this, like everybody’s starved for capacity, everybody’s starved for compute capacity and then, you know, all the associated things, memory and, and, and interconnected and everything else. Um, data center space. And so e every dollar right now that’s being put into the ground is turning into revenue.And, and it, and in fact, I actually think there’s an interesting thing happening, which is because everybody starve for capacity, the models that we actually have that we can use today are inferior versions of what we would have if not for the supply constraints. That’s true. Um, if Right pose a hypothetical universe in which GPUs were 10 times cheaper and 10 times more plentiful mm-hmm.The models would be much better. ‘cause you would just allocate a lot more money to training and you’d just build better models and they would be better. Um, and so we’re, we’re actually getting the sandbag version of the technology.swyx: Yeah. No. Everything we use is quantized because the, the labs have to keep the, the full versions,Marc: right?swyx: LikeMarc: we’re not even getting the good stuff.swyx: Yeah.Marc: But, but getting the good stuff, it’s, it’s just, even if technical progress stops. Once there’s like a much bigger build of like GPU manufacturing capacity and memory, you know, all, all the things that have to happen in the course of the next five or 10 years.Once it happens, even the current technology is gonna get, gonna get much better. And then as you know, like there’s just like a million ways to use this stuff. Like there’s just like a million use cases for this. Mm-hmm. Like, it, it, you know, this isn’t just sending packets across a, a thing, whatever, and hoping that people find something to do with it.This is just like, oh, we apply intelligence into every domain of human activity. And then it works like incredibly well. Yeah. Um. Here’s what I know, here’s what I know. Um, in the next three or four year, it’s like somewhere between three or four years out, basically everything is selling out. So like the, the entire supply chain is, is, is, is sold out or, or, or selling out.And so there, there’s no, like, we’re just gonna have like chronic supply shortage for, you know, for years to come. Um, there’s going to be a response from the market that’s gonna result in an enormous, you know, it’s happening now. An enormous flood of investment in a new fab capacity and ev you know, every, everything else to be able to do that, at some point the supply chain constraints will unlock, you know, at least to some degree that will be another accelerant to industry growth when that happens.‘cause the products will get better and everything will get cheaper. Um, and so, so I know that’s gonna happen. I know that, you know, the deployments, you know, the, the actual use cases are like really compelling. And then, like I said, you know, with reasoning and agents and so forth, like, I know they’re just gonna get like much, much better from here.And so I, I, I know the capabilities are like really real and serious. I also know that the technical progress is not going to stop. It. It, it is excel. It is, is accelerating. Like the, the breakthroughs are are tremendous. I mean, even just month over month, the breakthroughs are really dramatic. And so, you know, I think if you were a cynic and there, there are cynics, you can look at 2000, you can find echoes.But I can’t even imagine betting it that this is gonna like somehow disappoint and, you know, at least for years to come, I think it would be essentially suicidal to make that bet. Yeah. Um, it was that Michael Burry, uh, uh, that’sswyx: anMarc: interesting guy, huh? We’ll pick on a guy. We’ll pick, let’s pick on one guy.We’ll pick. Well ‘cause he did, he he came out with, it was, it was the, heswyx: doesn’t mind.Marc: It was the Nvidia short. Right. He came with the Nvidia short. And then if you guys probably talked about this, which is the, the analysis now that like the current models are getting better faster at such a rate that if you are running an Nvidia, if you’re running an Nvidia inference chip today, that’s three years old, you’re making more money on it today than you did three years ago because the pace of improvement of the software is, is faster than the, the, the depreciation cycle, the chip.And then my understanding is Google is running. I don’t if they’ve, I don’t know exactly what, uh, these are rumors that I’ve heard or maybe it’s public, but, um, I think Google’s running very old TPUs, very profitably. Ference. Yeah. And very profit and very profitably. Yeah. Um, and so, so it actually turns out, as far as I can tell, it’s actually the opposite of the Beery thesis is actually.He was actually 180 degrees wrong. It’s actually the, the, the, the old Nvidia chips are getting more valuable, which is something that’s like literally never happened before. Like it’s never been the case that you have an older model chip that becomes more valuable, not less valuable. And that, and again, that’s an expression of the just ferocious pace of software progress.Ferocious pace of capability payoff. Yeah. Uh, that you’re getting on the other side of this. And so I just, the idea of betting against that, like.swyx: Yeah. Yeah. Well, one ofMarc: my, it seems like an invitation to get your face ripped up.swyx: One of my early hits was like modeling the lifespan of the H 100 and h two hundreds and, and going like, you know, usually they advise like four to seven years and it was, you know, maybe you sort of realistically haircut cut it down to two to three.Yeah. But actually it’s going up and not down. Yeah. And, and uh, that’s, I mean that’s, I think that’s the dream. Uh, we are finding utilization and I think utilization solves all problems. Like, you can, you can find use, use cases for even like the poor, like even memory, we’re having a shortage. Right. And, and even like the, the shittier versions of, of memory that we do have, we are finding use cases for it.So like That’s great.Marc: Yeah.Alessio: How, how important is open source AI and kinda like edge inference in a world in which you have three years of supply crunch. Like, do you think in the, like, you know, if you fast forward like five years, like how do you think about inference, uh, in the data center versus at the edge?Marc: Well, so just to start, yeah. So I think, I think open source is very important for a bunch of reasons. I think edge, edge inference is very important for a bunch of reasons. I, I think just practically speaking, if we’re just gonna have fundamental construc, supply crunches for the next, I mean, you, you guys know if you just project forward demand over the next three years, right?Yeah. Relative to supply, one of the, its main predictions you can do is what’s gonna, what, what’s gonna happen to the cost of, of inference in the core, uh, over the next three years? And like, it may rise dramatically, right? Like, so, so what is, and then is, is, you know, like the, the, the big model competition are subsidizing heavily right now.Right? Right. And so, so what’s the, what will be the average person’s, you know, per day, per month token cost, you know, three years from now to do all the things that they want to do. And I, I don’t know, it’s gonna. I mean, I have, you guys probably have friends, I have friends today who are paying a thousand dollars a day for open claw, for claw tokens to run open claw.Right? And so, okay. $30,000 a month. Right? And, and by the way, those, those friends have like a thousand more ideas of the things that they want their claw to do, right? Yeah. And so you, you could imagine there, there’s like latent demand of up to, I don’t know, five or $10,000 a day of, of, of tokens for a fully deployed, you know, per personal agent.Uh, and obviously consumers can’t pay that, right? And so, so, but it gives you a sense of the fu of the fu of the future scope of demand, right? And so, so even, even if there’s a 10 x improvement in price performance, that still, you know, goes to a hundred dollars a day, which is still way beyond what people can pay.Mm-hmm. So there’s just gonna be like. Ferocious to me, by the way. The agent thing, the other interesting thing is I think the agent thing, so up until now, a lot of the constraints of GGPU constraints, I think the agent thing now also translates into CPU constraints. Mm-hmm. Right?swyx: CPU memory.Marc: Yes. CPU memory, right?And so, like the entire chip ecosystem is just gonna get wait,swyx: wait for network constraints, that that will be the killer.Marc: It’s all bottleneck potentially for years. And so, so I, I think that Brad, and, and I think it’s actually possible, I mean, generally inference costs are gonna keep coming down, but I think the, let’s put it this way, the rate of decline, I think may level out here for a bit because of these supply constraints.And then at some point, maybe the lab stops subsidizing so much and that, that, that again, will be, be an issue. And so there’s just gonna be so much more demand for inference than, than can be satisfied. Um, you know, kind of with the centralized model. And then, and then, you know, you guys know this, but like all the, just the dramatic, I mean just the dramatic innovations that have happened in the Apple silicon to be able to do, uh, inferences, it’s quite amazing the level of effort being put.Like the open source guys are putting incredible effort into getting, you know, this recurring pattern where the big model will never run on a pc, and then six months later mm-hmm. Oh, it runs in a pc, right? It’s like amazing. And there’s very smart people working on that. So there’s all that. And then look, there’s also, you know.There’s also like other, there’s other motivators. There’s other motivators which is just like, okay, how much trust are the big centralized model providers? You know, how much trust are they building in the market versus, you know, how much are, you know, at least for, in certain cases with some people, for certain use cases, people being like, well, I’m not willing to just like, turn everything over.So there, there, there’s all the trust issues. Um, by the way, there’s also just like straight up price optimization. There’s many uses of AI where you don’t need Einstein in the cloud. You just need like a, a a, a smart local model. There’s also performance issues where you want, you know, you want, you know, you’re gonna want your doorknob to have an AI model in it.Right. You know, to be able to, you know, do, um, you know, to be able to do access control. Um, obviously like everything with a chip is gonna have an AI model in it. Mm-hmm. And it, a lot of those are gonna be local. Um, and so, yeah. No, like I think, I think you’re gonna have ti and then you’re gonna, by the way, also wearable devices, you know, you don’t wanna do a complete round trip.You want, you know, you, whatever your smart devices are, you want it to be like super low latency. Yeah.swyx: The question, do we care who makes it? Yeah. One of the biggest news this week was the collapse of AI two, the Allen Institute. Mm-hmm. One of the actual American open source model labs. Yeah. Um, and, uh, I’m not that optimistic on, on American open source.Yeah. Like you, you guys invested in MIS trial and MIS trial’s doing extremely well outside of China. That’s about it.Marc: Yeah. We’ll see. We’ll see. I look, I, number one, I do think we care. Uh, I do think we, I do think we care who makes it. Um, I would say this, the, the, the, the previous presidential administration wanted to kill it in the us Oh yeah.They wanted to drown in the bathtub. Um, and so they wanted to kill it. So at least we have a government now that actually like, actually wants it wants it to happen. And youswyx: earned to councilMarc: and Yeah. And the new and the P pcast. Yeah. So the, the, you know, this admin for whatever other political issues people have, which are many, you know, this administration has, I think a very enlightened view and in particular an enlightened view on AI and in particular on open source ai.Uh, and so they’re very supportive. Um, my read is the Chi. The Chinese have a very, the various Chinese companies have a very specific reason to do open source, which is, they, they, they don’t fundamentally, they don’t think they can sell commercial, uh, AI outside of China right now. And or at least specifically not, not in the US for a combination of reasons.And so they, they kind of view, I think, open source AI as a bit of a loss leader against basically domestic, uh, you know, paid, paid services. And then kind of an, you know, kind of an ancillary products. You know, they’re, they’re very excited about it, by the way. I think it’s great. I think it’s great that they’re doing it.Um, you know, I think Deeps seek was like a gift to the world. Um, I think. The great thing about open source, open source, the, the, the impact of open source is felt two ways. One is you, you get the software for free, but the other is you get to learn how it works, right? And so like the paper, the paper, the paper and, and the code, right?And the code. And so, like, for example, I thought this was amazing. So open comes out with L one and it’s an amazing technical breakthrough, and it’s just like, absolutely fantastic. But of course they don’t explain how it works in detail. And then of course they hide the, they hide the reasoning traces, right?And, and then, and then, and then everybody’s like, okay, this is great, but like, who’s gonna be able to replicate this? Are other people gonna be able to do this? You know, is their secret sauce in there? And then our one comes out and it’s just like, there’s the code and there’s the paper, and now the whole world knows how to do it.And then, you know, three months later, every other AI model is, is adding reasoning. And so, so you get this kind of double, like even if the Chinese models themselves are not the models that get used, the education that’s taken place to the rest of the world, the information diffusion, you know, is incredibly powerful.So that happens and then, I don’t know. We’ll, we’ll see. You know, there are a bunch of American, you know, open source, you know, ai, uh, model companies. I mean, look, there’s gonna be tremendous, you know, there already is. There’s, you know, there’s gonna be tre there’s tremendous competition, uh, among the primary model companies.You know, there’s, depending on how you count, there’s like four or five, you know, big co model companies now that are, you know, kind of neck and neck, uh, in different ways. Um, uh, you know, and, and, and, um, you know, and then obviously Bo Bo both X and then MetAware involved are, you know, both have huge, you know, huge attempts to, you know, kind of, to kind of leapfrog underway.And then you’ve got, you know, a whole fleet of startups, new companies, including a whole bunch that we’re backing, that are, you know, trying to come out with different approaches. And then you’ve got whatever it is. I don’t know how, how many, how many, like main line foundation model companies are there in China at this point?It’s probably six. It’sswyx: five Tigers is what they call it. Yeah. Uh, Quinn is in questionable because there’s change in leadership,Marc: right?swyx: Yeah.Marc: But that, does that include, that includes like Moonshot,swyx: yes. Can deep seek, uh, uh, ZI, um, Quinn oh one is in there.Marc: Right. And then, um, and by dance and, and then you see,swyx: ance would be like the next tier ance.They weren’t as prominent. They weren’t, didn’t haveMarc: a leading. Yeah. But they, you at least, you know, ance is very inspiring and presumably they have more stuff coming and Tencent probably has more stuff coming and, and so forth. And so, so, so like, look, here, here would be a thing you can anticipate, which is there are not these markets, there are not going to be between the US and China right now, there’s like a dozen primary foundation model companies that are like at scale, at, at some level of a critical mass.It’s not gonna be a dozen in three years, right? Like, it just because these industries don’t bear a dozen, it’s, it’s gonna be three or you know, there’s gonna be three or four big winners or maybe one or two big winners. And so there’s gonna be like a whole bunch of those guys that are gonna have to figure out alternate strategies.Um, and I think like open source is one of those strategies. And so I, I think you could see like a whole, i, I, I think the questions like, who’s gonna do open source? I think that could change really fast. I, I think that, that, that’s a very dynamic thing. I think it’s very hard to predict what happens. And, and I think it’s very important.swyx: NVIDIA’s doing a lot.Marc: Well, I was gonna say. Well, exactly. And then you’re got Nvidia and then, and then, you know, just to, again, indu, there’s an old thing in business strategy, which is called, uh, commoditize Compliments. Commoditize the compliment. That’s right. And so if your Jensen is just kind of obvious, of course, you wanna commoditize the software.Yeah. And he’s, and to his enormous credit, he’s putting enormous resources behind that. And so maybe it, maybe it’s literally Nvidia and I think that would be great.Alessio: Yeah. Uh, narrative violation to European projects, uh, in the, uh, damn.swyx: I’m hosting my, uh, Europe, uh, conference soon. And I got both of them.Alessio: They got us.They got us. MarkMarc: finished. They got us, us. Well, wait a minute. Where was Peter? So where was Steinberger when he did? In AustriaAlessio: was, yeah, yeah, yeah.Marc: He was in what? He was in Vienna. Oh, he was in Vienna. And then where is he now?swyx: Uh, he’s moving to sf.Marc: Okay. Okay. Alright. Okay, there we go. And then, yeah, the PI guy, right?The PI guys are European.swyx: Yeah, they’re also, they’re buddies inAlessio: Australia. Mario’s also there. Yeah.Marc: Right. And are they, yeah, they haven’t announced yet. Any sort of change changed or have theyAlessio: No, they’re, they have a company there.Marc: Okay. Got, okay. Good.Alessio: Good, good,good.Alessio: Um,Marc: yeah, good.swyx: Anyways, I think pie and open cloud very important software things and, and I just wanted you to just go off on what you think.Marc: Yeah. So I think in co the, the combination of the two of them I think is one of the 10 most important softwares. Openswyx: Claw got all the attention, but Right. Talk about pie,Marc: pi pie’s, kind of the Yeah. PI’s, PI’s kind of the architectural breakthrough for those of us who are older. There was this whole thing that was very important in the world of software basically from like 1970 to, I don’t know, it still is very important, but like 19, from 1973 to like basically the creation of Linux, which is basically this, this thing used to call like the Unix mindset.Like so, so, ‘cause there were all these different, you know, theories. There are all these different operating systems and mainframes and, and then you know, all these windows and Mac and all these things. And then there was this, but kind of behind it all was this idea of kind of the Unix mindset. And the Unix mindset was this thing where basically you don’t have these, like, like in the old days, like, like the operating system that like made the computer industry really work, like in the 1960s mm-hmm.Was this thing called o os 360, which was this big operating system that IBM developed that was supposed to basically run everything. And it was this like giant monolithic architecture in the sky. It was like a, you know, it was like a giant castle. Um, of software. And, and by the way, it worked really well and they were very successful with it.But like, it was this huge castle in the sky, but it was this thing, it was almost unapproachable, which is like, you had to be kind of inside IBM or very close to IBM. And you had to really understand every aspect, how the system worked. And then the, the Unix sky is originally out of at and t and then out out of Berkeley, um, you know, came out and they said, no, let’s have a completely different architecture.And the way architecture’s gonna work is we’re gonna have, we’re gonna have a, a prompt and, and a, and a shell. And then, and then we’re gonna, all, all the functionality is gonna be in the form of these discreet modules, and then you’re gonna be able to chain the modules together. Mm-hmm. Yeah. And so like the, the, the op, it’s almost like the operating, operating system itself is gonna be a programming language.Um, and then that led led to the, the, the sort of centrality of the shell. Um, and then that led to sort of, uh, you know, basically chaining together Unix tools. And then that led to the emergence of these, these scripting languages like Pearl, where you, you could basically kind of very easily do this, and then the shells got more sophisticated and then, and then, and then look like, you know, that, that, that number one, that worked and that, that was the world I grew up in.Like I was, I was a Unix guy. You know, sort of from, call it 1988 to, you know, kind of all, all the way through my work and it worked really well. It, it’s in the background, um, you know, nor normal people don’t need to, didn’t need to necessarily know about it, but like, if you were doing like system architecture, application development, you, you, you knew all about it.Um, and then, you know, it’s been in the background ever since. And, you know, look, your Mac still has a Unix shell, you know, kind of in there, and your iPhone still has a Unix shell kind of buried in there somewhere. So they’re kind of in there. And then, you know, the Windows shell is kind of a, you know, sort of a weird derivative of that.But, um, you know, but look, the inter, the internet runs on Unix, um, and that smartphones, actually, both iOS and Android are Unix derivatives. And so, you know, kind of Unix did end up winning. But, but anyway, and then we just started taking that for granted. And then, and then so, so basically the, the way I think about what happened with Pie and then with Open Claw is basically what those guys figured out is, I always say the, the great breakthroughs are obvious in retrospect, right?Which is the best kind, the best kind. They weren’t obvious at the time or somebody else would’ve done them already. Um, and so there is a, like a real conceptual leap, but then you look at it sort of the backwards looking and you’re just like, oh, of course. Mm-hmm. Like the, the, to me those are always the best breakthroughs.Well, actually language models themselves are like that. It’s just like, oh, next token completion. Oh, of course.swyx: Yeah. What other objective mattered?Marc: Yeah, exactly. But, but like it, right. But she’s even saying it wasn’t obvious until somebody actually did it. Right. And so the conceptual breakthrough is real and deep and powerful and, and very important.And so the way I think about pie and olaw is it’s basically marrying the, the language model mindset to the un to the Unix, basically shell prompt mindset. And so it’s, it’s basically this idea that what, what, so what is an agent, right? And as, as, and as you know, like many smart people who have been trying to figure out what an agent is for, for, for decades, and they’ve had many architectures to build agents and the whole thing.And it turns out what is an agent. So it turns out what we now know is an agent is the following. It’s, so it’s a language model. And then above that, it’s a ba, it’s a bash shell. Um, so it’s a, it’s a Unix shell, and then it’s, and then the agent has access, uh, has access to, to the shell. And, you know, hopeful, hopefully in a sandbox, maybe in, maybe in a sandbox.So it’s, it’s the model. Um, it’s the shell. Um, and then it’s a fi, it’s a file system. Um, and then the state is stored in files. And then, you know, there’s the markdown format for the, you know, for, for the files themselves. And then, and then there’s basically what in Unix is called Aron job. There’s a loop and then there’s a heartbeat for the, there’s heartbeat and, and the thing basically Wake Wakes up.Wakes up. So it’s basically LLM plus shell, plus file system, plus markdown, plus kron. And it turns out that’s an agent. And, and, and every part of that, other than the model is something that we already completely know and understand. And in fact, it turns out that like the latent power of the Unix shell is like extraordinary because basically like all, like, there’s just like an, there’s just enormous latent power in the shell.There’s enormous numbers of Unix commands, there’s enormous number of command line interfaces into all kinds of things already in the, you know, your entire, I mean your entire, just to start with, your computer runs on a shell. If you’re running a Mac or a, or, or a phone, your computer, your computer’s running on a shell, uh, already.And so like the full power of your computer is available at the command line level. Um, and then it turns out it’s really easy to expose other functions as a command line interface. And so like this whole idea where we need like MCP and these like product mm-hmm. Fancy protocols, whatever, it’s like, no, we don’t, we just need like a command, command line thing.So that’s the architecture. And then it turns out what is your agent? Your agent has a bunch of files starting a file system. And then there’s the thing that just like completely blew my mind when I write my head around it as a result of this, which is like, okay. This means your agent is now actually independent of the model that it’s running on.Because you can actually swap out a different LLM underneath your agent and your, your agent will change personality somewhat. ‘cause the model is different, but all of the state stored in the files will be retained.swyx: Yeah. Different instruction set, but you just compiledit.Marc: Right, exactly. And it’s all right.It’s like right. Swapping out a ship and recompiling, but it’s, it’s still, it’s still your agent with all of its memories. Um, and with all of its capabilities. And then by the way, you can also swap out the shell, uh, so you can move it to a different execution environment that is also, is also a b shell, by the way, you can also switch out the file system, right.Uh, and you can, and you can, and you can swap out the, the, the heartbeat for the, the crown framework, the, the loop that the agent framework itself. And so your agent basically is ba basically at the end of the day, it’s just. It’s just, its files. Um, and then, and then there’s of course it a openswyx: call.Marc: Yeah, it’s, it’s basically, it’s, it’s just the files.Um, and then by the way, as a consequence of that, the agent and then the agent itself, it turns out a couple important things. So one is it, it’s, it, it can migrate itself, right? And so you’re, you can instruct your agent, migrate yourself to a different, uh, runtime environment, migrate yourself to a different file system, migrate yourself to a different, you know, swap out the language model.Your agent will do all that stuff for you. And then there’s the final thing, which is just amazing, which is the agent is the agent actually has full introspection. It actually, it actually knows about its own files and it could rewrite its own files. Right. Which by the way, is basically no widely deployed software system in history where the, the, the thing that you’re using actually has full introspective knowledge of how it itself works and is able to modify itself.Like that, that, I mean, there have been toy systems that have had that, but there, there’s never been a widely deployed system that has that capability and then that leads you to the capability. That just like completely blew my mind when I wrap my head around it, which is you can tell the agent to add new functions and features to itself and it can do that.Extend yourself. Yeah. Right? Extend, extend yourself. Like extend yourself. Give yourself a new capability. Right? And so, and so literally it’s just like you run into somebody at a party and they’re like, oh, I have my open claw, do whatever, connect to my eat, sleep bed, and it gives me better advice and sleep.And you go home at night and you tell your claw, or if they’re at the party, by the way, you tell your claw, oh, add this capability to yourself. And your claw will say, oh, okay, no problem. And it’ll go out on the internet and it’ll figure out whatever it needs and then it’ll go out to claw code or whatever.It’ll write whatever it needs. And then the next thing you know, it has this new capability. And so you don’t even have to, like, you can have it upgrade itself without even having to, without having to do anything other than tell it that you want it to do that. And so anyway, so the, the combination of all this is just, I mean, this is just like a massive, incredible, I mean, it’s just incredible.Like if I, if I were, if I were 18, like this is a hundred, this is what I would be spending all of my time on. This is like such an incredible conceptual breakthrough. Yeah. And again, pe people are gonna look at it and they already get this response. People are gonna look at it and they’re gonna say, oh, well, where’s the breakthrough?‘cause these, the, all of these components were already known before. Mm-hmm. But, but this is the key, the key to the breakthrough was by using all these components that were known before, you get all of the underlying capability of that’s buried in there. And so all, and so for example, computer use all of a sudden just kind of falls, trivi, trivial.Of course it’s gonna be able to use your computer. It has full access to the shell. Right. And then, and then you just, you, you give it access to a browser, and then you’ve got the computer and the browser and, and often away it goes. And, and then you’ve got all the abilities of the browser also. Um, yeah.And so, and so the capability unlock here is profound. My friends who are, you know, deepest into this, are having their claw do like a, like, literally like a thousand things in their lives. They have new ideas every day. They’re just like constantly throwing new challenges at the thing. And by the way, it’s early and, you know, these are, you know, these are prototypes and there are, you know, as you guys know, there’s security issues.Yeah. And, and so, you know, there’s a bunch of stuff to be ironed out, but the, the unlock of capability is just incredible.swyx: Yeah.Marc: And I, I have absolutely no doubt that everybody in the world is gonna, is gonna have at least, you know, an agent like this, if not an entire family of agents. And we’re gonna be living in a world where I think it’s almost inevitable now that this is the way people are gonna use computers.swyx: I was gonna say for someone who is deeply familiar with social networks, the next step is your claw talking to my Claw. Mm-hmm.Marc: Postingswyx: on Claw Facebook, uh, posting their jobs on cloud LinkedIn and close posting their tweets on claw XAI or what, whatever, you know. Um, I do think that that is how, uh, you know, we, we get into some danger there in, in terms of like alignment and whether or not we want these things to, to, to run.Marc: You guys know where Rent a, rent a human.com.swyx: Yeah. Rent a,Marc: yeah. Yeah.swyx: I mean, it’s Fiverr, it’s TaskRabbit.Marc: Sure, of course.swyx: MechanicalAlessio: Turk.Marc: Yeah. But flipped, right. The agent hiring the people.Alessio: Yeah.Marc: Which of course is gonna happen, right? It’s obviously gonna happen.Alessio: I’m curious if you have any thoughts on the engineering side.So when you build the browser, the internet, you know, just a bunch of mostly plain text file plus some images, and today the, every website and app is like, so complex. Somehow, you know, the browser kept evolving to fit that in. Mm-hmm. Are there any design choices that were made like early in the browser and kinda like the internet and the protocols that you’re seeing agents similar to this?Like, Hey, this thing is just not gonna work for like this type of new compute and we should just. Rip it out right now.Marc: There were a whole bunch, but I’ll give you a couple. So one is, um, and we didn’t, you know, to be clear like this, this was not, you know, this is totally different. We didn’t have the capabilities we have today, but because Wet have, we didn’t have the language models underneath this, but, um, we did have this idea that human readability actually mattered a great deal.Um, and, and, and so, and specifically in those days, it was, it was not so much English language, but it was there, there was a design decision to be made between binary protocols and text protocols. And basically every, every, every basically old school systems architect that had grown up between like the 1960s and the 1990s basically said, you know, the internet, it’s, what do you know about the internet?It’s star for bandwidth. You, you just, you have these very narrow straws. Uh, you know, look, people, when we did the work on Mosaic, like pe, people who had the internet at home had a 14 kilobit modem, right? So you’re, you’re trying to like hyper optimize every bit of data mm-hmm. That, that travels over the network.And so obviously if you’re gonna design a protocol like HGTP, you’re gonna want it to be binary, you know, highly compressed, binary protocol for maximum efficiency. And you’re gonna wanna have it be like a single connection that persists. And you’re, you’re, the last thing you’re gonna wanna do is like, bring up and tear down new connections.And you definitely, you’re not gonna, not gonna want a text protocol. And so of course we said no. We actually want to go completely the other direction. It’s obviously, we only want text protocols. Uh, by the way, same thing in H TM L itself. We want html to be relatively verbose. You know, we want the tags to actually be like human readable.Um, we wanna useswyx: the most inefficient things possible.Marc: Yeah, we wanna do the, we wanna do the in, we wanna do the inefficient things.swyx: You’re the original token Mixer.Marc: Yeah, exactly. Yeah, yeah, yeah. Basically it’s just like better lessonAlessio: filled.Marc: Well, yeah. Well actually this was, this was actually the, the conscious thing, which basically says just like assume, assume a future of infinite, infinite bandwidth built for that, right?And then basically what it was, is it was a bet that it, it was a bet that if the system, if the, if the latent capabilities of the system were powerful enough, and that was obvious enough to people that would create the demand for the bandwidth that would cause the supply of bandwidth to get built that would actually make the whole thing work.And then specifically what we wanted was we wanted everything to be human readable because we, at the engineering level, we wanted people to be able to read the protocol coming over the wire and be able to understand it with their, with their bare eyes without having to like disassemble it or whatever.Right. Have it converted outta binary. Right. And so the, the, the, all the pro, you know, HTTP and everything else were, were, it was always, uh, text protocols. Uh, and the same thing with HTML and in, in many ways, some people say that the key breakthrough in the browser was the view source option, um, which is every webpage you go to, you could view source, which means you could see how it worked, which means you could teach yourself how to build right new, uh, to, to build new webpages.There was that. So human readability. Um, and, and again, human readability in those days still meant technical, you know, specs. You know, now it means English language, but there’s an incredible latent power in giving everybody who uses the system the option to be able to drop down and actually understand and see how it’s working.And that worked really well for the web and I think it’s working really well for ai. That was one. Um, what was the other, um. A big part of the idea of web servers was to actually surface the underlying latent capability of the operating system and to be able to surface the, uh, also the underlying latent capability of the database because basically what was a web server?What, what, what, what is a web server? Fundamentally? Architecturally, it’s, it’s, it’s the operating system. So it’s, it’s the operating system’s ability to, you know, it’s running on top of an os. So it’s the OSS ability to manage. The file system and do everything else that you wanna do, process everything. Um, and then of course, a lot of early, you know, a lot, a lot of websites are, are front ends to databases.Um, and so you wanted to, you wanted to unleash the underlying latent power of whether it was an Oracle database or some other, you know, some other Postgres or whatever, whatever it was. Um, and so a lot of the function of the web server was to just bridge from that internet connection coming in to be able to unlock the underlying power of the OS and the database.Uh, and again, people looked at it at the time and they were like, well, is this really, does this really matter? Like, is this important Because we’ve had databases forever and we’ve always had, you know, user interfaces for databases and this is just another user interface for a database. And it’s like, okay, yeah, fair enough.But on the other side of that is just like, this is now a much better interface to databases and one that 8 billion people are going to use and is going to be like, far easier to use and far more flexible. And, and, and, and you’re not just gonna have old databases. Now you have a system where people can actually understand why they want to build, you know, a million times more database apps than they have in the past.And then the number of databases in the world exploded. And so again, this goes to this thing of like building, building in layers. Some of the smartest people in the industry look at any new challenge and they’re like, okay, I’m, I’m, I need to build a new kind of application. So the first thing I need to do is build a new programming language, right?And then the next thing I need to do is build a new operating system, right? And then the next thing I need to do is I need to build a new chip. Right? And they, they kind of wanna reinvent everything. And I’ve, I’ve always had, maybe it’s just, I don’t know, pg pragmatic mentality or something, or maybe an engineering over science mentality, but it’s more like, no, you have just like all of this latent power, uh, in the existing systems and you, you don’t want to be held back by their constraints, but what you wanna do is you wanna kinda liberate that power and open it up.Yeah. And so I, I think, I think, and I think the web did that for those reasons. And I think it’s the same thing now that’s happening. It’s a greatswyx: perspective on the web.Alessio: Programming language just is not a good thing. We have Brett Taylor on the podcasts and we were talking about rust. And you know, rust is memory safe by the phone.So why are we teaching the model to not write memory, unsafe code, just use rust, and then you get it for free. How much do you think there’s like. Time to be spent like recreating some of these things instead of taking them for granted. I’ll be like, oh, okay. Python is kind of slow Pythonswyx: type scripts,Alessio: you know?It’s like, yeah.swyx: As, as imperfect as they are, they are the lingua franca.Marc: I mean, I think this is gonna change a lot. ‘cause I don’t think the models care what language they program in. Mm-hmm. And I think they’re gonna be good at programming in every language, and I think they’re gonna be good at translating from any language to any other language.Like, okay, so this gets into the coding side of things. I, I think we’re going through a really fundamental change. And then, look, I, I grew up hand, you know, I grew up hand code, you know? Yeah, yeah, yeah. I grew up hand coding. Everything I did was actually everything I did actually was written in CI wasn’t evenAlessio: back in the days,Marc: I wasn’t even using c plus plus, so I, or like Java or any of this stuff.Right. Uh, and so, um, I, everything, everything I ever did, I was like managing my own memory at, at, at the level of c and then I, you know, I, I’m still from the generation that, you know, I, I knew assembly language and, you know, I, I, you know, um, so I, I could drop down and do things, uh, right on the ship. And so we, we’ve just, we’ve all, all of us, we’ve always lived in a world in which software is like this precious thing that like, you have to think about very carefully.And it’s like really hard to generate good software. And there’s only a small number of people who can do it. And like, you have to be very, like, jealous in terms of thinking about like, how do you allocate, like what are your engineers working on and how many good engineers do you actually have? And how much software can they write?And how can, how much software can human beings, you know, kind of maintain? And I think like all those assumptions are being shot right out the window right now. Like, I think they’re, I, I think those days are just over. And I think the new world is like, actually high quality software is just like infinitely available.Mm-hmm.Marc: And if you need new software to do X, Y, Z, like, you’re just gonna wave your hand and you’re gonna get it. And then if it’s, if you don’t like the languages written in, you just tell the thing, all right, I want the, now I want the rush version. Um, or, you know, se secure, you know, secure. We’re about to, by the way, we’re about to go through computer security is about to go through the most dramatic change ever, which is number one, like every single latent security bug is about to be exposed,swyx: right?Marc: So we’re gonna have like, the in, we’re, we’re, we’re set up here for like the computer security apocalypse for a while. But, but, but on the other side of it, now we have a coding agents that can go in and actually fix all the security bugs. And so how, how are you gonna secure a software in the future?You’re gonna tell the, tell the bot to secure it, and it’s gonna go through and, and fix it all. And so, so this thing that was this incredibly scarce resource of high quality software is just going to become a completely fungible thing that you’re just gonna have as much as you want, right? Uh, and, and that has like, you know, that has like tons and tons of consequences in some sense.The answer to the question that you posed, I, I think it’s just somewhat, I don’t know, simple or something, or straightforward, which is just, if you want all your software and rust, you just, all the bot, you want all your software and rust, like, things that used to be like hard or even like, seem like an insurmountable mountain to get to get through all of a sudden, I think, become very easy.swyx: I, I think Brett had a theory that there would be a more optimal language for lms. And so the contention is, uh, there isn’t like, just don’t bother, just whatever humans already use LMS are perfectly capable, porting.Marc: I think we’re pretty close to being, I don’t know if this would work today. I think we’re pretty close to being able to ask the AI what would its opt optimal language be and let Right, and let it design it.True. Okay, here’s a question. Are you gonna even gonna have programming languages in the future? Um, or the ai, are the AI just gonna be emitting binaries? Let’s assume for a moment that humans aren’t coding anymore. Let’s assume it’s all bots. The bot. What levels of intermediate abstraction do the bots even need?swyx: Yeah.Marc: Or are they just coding binary directly? Did you see there’s actually an experi, somebody just did this thing where they have a, they have a, a language model now that actually emits model weights for a new language model. Right. And so will the bots be justAlessio: predict the weightsMarc: Will, yeah. Will the bots literally be emitting not just coding binaries, but will they, will, will they actually be admitting weights for, for new models?Yeah. Direct directly and. Conceptually, there’s no reason why they can’t do both of those things. Uh, like architecturally. Both of those things seem completely possible. It’sswyx: very inefficient. You’re basically veryMarc: inefficient.swyx: A simulation of a simulation in a simulation inside of the weights. Correct?Marc: Yeah, yeah. Very inefficient. But like, look, LMS are already like incredibly inefficient. Ask an uh, in favor thing, ask Claude, add two plus two equals four. Right? It’s just like, you know, it’s like, you know, it’s, it’s, it’s like whatever, billions and billions of times more inefficient than using your pocket calculator.swyx: Yeah.Marc: But, but, but yet the, the, the payoff is so great of the general capability. And so anyway, like I, I kind of think in 10 years, like, I’m not sure. Yeah. Like, I’m not sure there will even be a salient concept of a programming language, um, in the way that we understand it today. And in fact, what we may be doing more and more is a form of interpretability, which is we’re trying to understand why the bots have decided to, uh, structure, uh, code in the way that they have.swyx: I mean, if you play it through, you don’t need browsers, then like, that’s the depth of the browser.Marc: Well, so I, I would take it a step further, which is you may not need to use your interfaces. So who is gonna use software in the future?swyx: Other bots.Marc: Other bots. Yeah. Yeah. Andswyx: so you still need to, I don’t know, pipe information in,Marc: do we?swyx: And outMarc: reallyswyx: well, what are you gonna do then?Marc: Are you sureswyx: you’re just gonna log off and touch grass?Marc: Whatever you want. Exactly. Isn’t that better?swyx: I want software to do stuff for me.Marc: Isn’t that? But isn’t that better? I mean, look, I, you know, I don’t know. Look like, you know, you know, you, all the arguments here, you know, it was not that long ago that 99% of humanity was behind a plow.swyx: Right.Marc: Right. And what are people gonna do if they’re not plowing fields all day to, to, to grow food? Right. And it just turns out there’s like much better ways for people to spend time than plowing fields. Yeah.swyx: Dooms growing.Marc: Uh, yeah, exactly. Exactly. Or, you know, talking to their friends and look, and I’m not an absolutist and I’m not a utopian.And I, and to be clear, like I’ve, I have an 11-year-old and he’s learning how to code and like I’m, you know, I, I think it’s still a really good idea to learn how to code and so forth, but I just, if you project forward, you just have to think forward to a world in which it’s just like, okay, I’m just gonna tell the thing what I need and it’s gonna do it, and then, and then it’s gonna do it in whatever way is most optimal for it to do it.Mm-hmm. Yeah. Unless I tell it to do it non optimally. Like if I tell it to do it in Java or in Rust or whatever, it’ll do it, I’m sure. But like, if I’m just gonna tell it to do, it’s, gonna do it in whatever way is like the optimal way to do it. Yeah. And then I, and then if I need to understand how it works, I’m gonna ask it to explain to me how it works.Right. And so it’s gonna be doing its own, interpret it, it’s gonna be the engine of interpretability to explain itself. And I, I just am not convinced that, that I’m not, I’m not convinced that in that world you have these historical, the goals of the abstractions will be whatever, the Boston network with the human Right.Alessio: Yeah. Yeah. That, well, I, I’m curious like. If that’s true, then shouldn’t the models providers be building some internal language representation that they can do extreme, kinda like rl uh, and reward modeling around, because it’s like, today they’re kind of like tied to like type script and Python because the users need to write in that language versus they can have their own thing internally and like they don’t need to teach it to anybody.They just need to teach their model. And I think that’s how you get maybe the version between the models, like going back to like the pie open claw thing. It’s like, oh, I built all the software using the open AI model and now switch to the RO model. But the TRO model doesn’t understand the thing. So I I, it feels like there still needs to be some obstruction.But maybe not. Maybe that’s the lockin that the model providers want to have. I don’t,Marc: I’m not even sure that’s lockin though. ‘cause why can’t the second model just learn what the first model has done? Like,swyx: exactly.Marc: Okay. So okay. Give you an example. So as you know, models can now reverse engineer software by, right?Isn’t it the whole thing now where people are reverse engineering, like Nten, Nintendo, gay binaries. Yeah. So you, you have like there’s, I’ve seen a bunch of reports like this where somebody has like a favorite game from the 1980s and the source code is like long dead, but they have like a binary brand to do a chip or something, another reverse engineer to get a version that runs in their Mac.Right. And so if you reverse it, if, this is why I kinda say if you’re reversing like X 86 binaries, then why can’t you reverse engineerAlessio: whatever the degree. Yeah. And because we’re all on a Unix based system, it has to be reversible because it needs to run on the target.Marc: Yeah, yeah, yeah, yeah, yeah. Basically.And so I just, I just think it’s this thing where it’s just like, and by the way, and everything we’re describing is something that human beings in theory could have done before, but just with like, right. Yeah, yeah. But with enormous where, but it was just always like cost and labor prohibitive. Reverse engineer.I learned how to reverse engineer. Human beings can reverse engineer binaries. Yeah. It’s just for any complex binary, you need like a thousand years mm-hmm. To do it. But now with a model, you don’t. And so all of a sudden you get, you get these things. Or, or another way to think about it is so much of human built systems are to compensate for the human limitations.swyx: Mm-hmm.Marc: Yep. Right? Um, and if you don’t have the human limitations anymore, then all of a sudden you have, and, and it’s not that you, you won’t have abstractions, but you’ll have a different kind of abstraction. Yep. Yep.swyx: I have two topics to bring us to a close. And, uh, you could pick whichever ones. Uh, just talking about protocols, was it you or someone else?Uh, I forget my internet history. Who said that? Like the biggest mistake that we didn’t figure out in the early days was payments. Yes. Was that you?Marc: Yes. Itswyx: was a 4Marc: 0 2swyx: 0 2 4Marc: 0 2 payment required.swyx: We have a chance now. Nope. I don’t think we’re gonna figure it out. I don’t know. Like, what’s your take?Marc: Oh, I think, we’ll, yeah, no, now I think it’s gonna happen for sure.swyx: Yeah.Marc: Yeah. And there’s two reasons to example for sure. One is we actually have internet native money now in the form of crypto. Stable coins. Stable coins and crypto. And this is, I, I think this is the grand unification basically of ai, crypto, uh, is what’s about to happen now. Um, I think AI is the crypto killer app, I think is where, where this is really gonna come out.Um, and then the other is it’s just, it, I mean it’s just, I think it’s now obvious. It’s like obviously AI agents are gonna need money and it’s already happening, right? If you’ve got a c if you’ve got a claw and you wanted to buy things for you, you have to give it money in some form.swyx: I would say the adoption’s probably like 0.1% if, if that, but Yeah.Marc: Oh, today? Yeah. Yeah, yeah. But think, think forward, like where is it goingswyx: forward thinkingMarc: The ultimate principle of everything and, and everything that I think I, we, we do is, it’s the William Gibson quote, which is, the future is already here. It just isn’t distributed. Mm-hmm. It isn’t, isn’t distributed yet.My friends who are the most aggressive use users of, of, of, of open claw, just like have given their clause bank accounts and credit cards. Um, and, and, and, and, and not only have they done it. Obvious that they needed to do it because it’s obvious that they needed to be able to spend money on their behalf.swyx: Yeah. Yeah.Marc: It’s just completely obvious. And so, and again, like, so the number of people who have done that today to your point is like, I don’t know, probably 5,000 or something. Yeah. Butswyx: it’ll grow.Marc: That’s how these things startswyx: actually, I mean, since, uh, you keep mentioning,Marc: and by the way, open cloud, by the way, if you don’t give it a bank account, it’s just gonna break into your, your, it’s gonna break high agency, it’s gonna break into your bank account anyway, and, and take your money.So you, you might, as you might as well do it, you might as well do it,swyx: uh,Marc: by the way. I really love, I gotta tell you, I really love the phenomenon. I love the Yolo. Um, I’m not doing it myself to be clear, but, but I love the people that are just like, yeah, what, what is it? Skip, skip, vision,swyx: danger, skip.Marc: Dangerous.swyx: Which by the way, is a Facebook thing.Marc: Okay?swyx: Right. Because, uh, because we, uh, in Facebook, they, they have this culture to name the thing dangerous, so that you are aware when you enable the flag that you are opting into a dangerous thing.Marc: Okay, good.swyx: And they brought it into open ai and of course thatMarc: makes it enticing.swyx: Sam runs Codex, uh, with skip permissions on, on his laptop.Marc: Yes, a hundred percent. And so I, I th I think the way to actually see the future is to find the people who are doing that. There’s a man, you know, and they, you knows,swyx: log everything, you know, just watch it, watch the logs,Marc: but. Let’s actually find out what the thing can do.Yeah. And the way to find out what the thing can do is just like, try everything. Yeah. Let it try everything. Let it unlock everything. By the way, that’s how you’re gonna find all the good stuff it can do. By the way. That’s also how you’re gonna find all the flaws. Yeah. I think the people who turn that on for bots are like, they’re, they’re like martyrs to the progress of human civilization.Like, I feel very bad for their descendants that their bank accounts are gonna get looted by their bots in the first like 20 minutes. But I think the contribution that they’re making to the future of our species is amazing.swyx: It’s like gentleman science, you know?Marc: Yes. It’s, yes, yes. Experi yourself. It’s, uh, Ben Franklin out with the, trying to try, trying to get lightning to strike his, his, uh, his balloon and see, seeing if he gets electrocuted.swyx: Yeah.Marc: It’s, uh, Jonas sk with the polio vaccine, right. Injecting it. Yes. So, yes. I, I, I, I think we should have, like agl, we should have like flags and like we should have like monuments to the people that just let open club run their lives.swyx: More anecdotes of like, what, what are the craziest or interesting things that people listening to this should go, go home and do.Marc: I mean, this is, this is the, this is the, the extreme thing is just like the straight Yolo, like just Yeah. Turn, turn your lifeswyx: on. I mean, that’s a general capability. Yeah. Yeah. Is there like a specific story that was like, wow. And, and everyone in a group chat just lit up.Marc: I mean, like, you know, so there’s tons of, there’s already tons of health, you know, there’s the health dashboard stuff is just, is just absolute personal health.Absolutely amazing. Yeah. The number of stories on, um, I just don’t wanna violate people’s, you know, obviously personal. Yeah. Anonymized. But, um, you know, one of the things open clouds are really good at is hacking into all this stuff in your land. Uh, it’s really good. So, you know, internet of things. AKA internet of s**t.swyx: Yeah.Marc: Likeswyx: super insecure, but great. It’s discoverable.Marc: Yeah, it’s discoverable. O open claw is happy to scan your network, identify all the things. And then my, my, my friends who are most aggressive at this are having open claw take over everything in their house.swyx: Yeah.Marc: Take it takes over their security cameras.It takes over their, their, you know, their whatever their, their access control systems. It takes over their webcams. I have a friend whose claw watches him sleep. Put a webcam in your bedroom. Put the, put the claw, put the claw on a loop. Uh, I have it. Wake up frequently and have it watch, just tell it, watch me sleep.And, and I’ve, I’ve seen the transcripts and it’s literally like Joseph asleep. This is good. This is good that Joe’s asleep. ‘cause you know, I have, I have his health day and I know that he hasn’t been getting enough sleep and so it’s really good that he’s getting sleep. I really hope he gets his full, whatever, you know, five hours of REM sleep.Uh, Joe’s moving. Joe’s moving. Um, uh, Joe might be wake waking up. This is a real pro. If Joe wakes up now, he is gonna ruin his sleep cycle. Oh, okay. It’s okay. Joe just rolled over. Okay. He’s gone back to bed. Okay, good. Alright. Okay. I can relax. This is fine. He’sswyx: monitoring the situationMarc: monitoring, monitoring the situation, and, and being a bot, like, you know, is just like very focused, right?It’s just like, uh, this is like, its reason for existence is to watch Joe sleep. And then, and then I was talking to my friend who did this is like, you know, on the one hand it’s like, all right, this is weird and creepy. Um, and I need to, I need to, maybe this has taken over my life. And then the other thing is like, you know what if I had a heart attack in the middle of the night, this thing literally would like freak out and call 9 1 1.Like, there’s no question. This thing would figure out how to like, alert medical authorities and like, prob probably some in SWAT teams and like, do whatever would be required to save my life. Right? And so it’s like, you know, like, yeah. Like that’s happening. What else? Um, I’ll give, I, um, uh, it’s a company unitary, uh mm-hmm.That makes the robot dogs. Um, and I, I actually have one at home, which is, it’s actually really fun. The Chinese companies, the Chinese companies are so aggressive at adopting, uh, new technology, but they don’t always like, listen, take the time to really.swyx: Package it,Marc: package it, and maybe think it all the way through.And so, so the, at least the industry dog I have, so it, it has a old non LLM just control system, which by the way is not very good in, in markets. Well, but it, in practices, it’s not that good. It has trouble with stairs and so forth. And so it’s not quite what it should be. But then the language model thing comes out in the voice.So they, they add, so they add LLM capability and then they, they add a voice mode to it. Um, but, but that LLM capability is not at all connected to the control system. So, so you’ve got this schizophrenic dog that like, is a complete idiot when it comes to climbing the stairs, but it will happily teach you quantum mechanics.Right. In like a lum English accent. Right. Like, it, it, it is just like absolutely amazing. Jagged intelligence. Yeah. Yeah. Talk about jagged and then, now obviously what’s gonna happen in the future is, is they’re gonna connect together, but they’ll do it. But right now it’s, it’s, and so right now it’s not that useful.And so I, I have a friend who has one of these who had his claw basically hack in and rewrite the code Rew write new firmware. Yeah. Write new firmware for the, for the unit robot. Ooh. And now it’s, now it’s an actual pet dog for his kids.swyx: You could do that before or after like. The motion.Marc: Yeah. It’s, he said it’s completely different.He said it’s a complete transformation. Yeah. And whenever there’s an issue in the thing, now the claw just like reiterates the code. You know, you know, you goes in, it does, does the code and so is it kind of goes to your thing here. So, so like all of a sudden, uh, this is why the way we wanna think about AI code AI coding is not just like writing new apps.It’s also going in and rewriting all the old stuff that should have worked that never worked. And so, like, I, I think, I think basically, I think the internet, the internet of s**t is basically over. Like, I, I think everything, there’s a potential here where like all these devices in your house that have been like basically marginal or you know, basically dumb, you know, like all of a sudden they might all get really smart.Now you have smartswyx: home.Marc: You have to decide if, yes, there are horror movies in which this is just, of which this is the premise. And so you have to decide if you want this. Yeah. But, but, but this is the first time I can say with confidence, I now know how you could actually have a smart home. Yeah. Yeah.With 30 different kinds of things with chips and internet access, where it actually all makes sense and all works together and it’s all coherent in the, in the whole thing. And to have that unlock without a human being having to go do any of that work, like, you know.swyx: You know, I, I’m, I’m waiting for a, sorry, mark.Uh, I can’t let you open that fridge door, you know, likeMarc: Exactly, exactly. Yes, yes.swyx: Because Oh, yeah, yeah. You’re not supposed to eat rightMarc: now. I have all of, yes, I have every shred of health information, you know, and I know you think you’re doing, you know, da da da. I didn’t think you do this, but you know, this is a real, are you really, you know, are you really sure?And you know, you told, you know, you told me last night, you really don’t want me to let you do this, so, you know, I’m sorry, but the fridge door is locked. Um, yes. Openswyx: the fridge doors.Marc: Exactly. And by the way, I know you’re supposed to be studying for a test, so why don’t we, why don’t you go when you can pass the test, um, I will open the fridge door for you.Yeah.swyx: Final protocol and then, and then we can wrap up, uh, proof of humanMarc: Yes.swyx: Uh, right.Marc: Yeah.swyx: That’s the last piece that we gotta figure out.Marc: Yeah. So I would say there’s, there’s two massive, I would say, um, uh, sort of asymmetries in the world right now where we’ve known these asymmetries exist and we, we societally have an unwilling to grapple with them.And I think they’re both tipping right now. And, and they’re, they’re, they’re, they’re the same thing. It’s virtual world version. It’s a physical world version. So the virtual world version is, is the bot problem. We’re just like, you know, the internet, internet is just like a wash and bots, internet’s a wash and fake people.It has been forever. Um, by the way, a lot of that has to do with lack of money, you know? And so this, you know, this is the Yeah, this is this.swyx: My spicy take was these two are the same thing. And corporations of people too, you know? So interesting.Marc: Yeah, yeah, yeah.swyx: Okay. So a bank account is proof of human.Marc: Yeah.Okay. Yeah. Until you, until you give the bots bank accounts. Yeah, exactly. So, okay. Yeah. So there’s that. But yeah, look, look, the bot, I mean, every social media user knows this. The bot, the bot problem is a big problem. You know, the bot, the bot problem has been a big problem forever. It’s, it’s a huge problem.And it’s never really been confronted directly, like at any point, by the way. The physical world version of this is the drone, the drone problem. Um, right. And so we, we’ve known for, you know, we’ve known for 20 years now that the asymmetric threat both in Milit military and actual military conflict, but also in just like security, like, like, you know, security on the home front.The big threat is, is the cheap attack drone. Right? The, the, the cheap, the cheap suicide, you know, drone with the bomb. And we’ve known that forever. And by the way, like, you know, it’s very disconcerting how like every, you know, every office complex in the, in the co you know, in the world is like unprotected from drone attacks.Um, every, every stadium, every school, every prison. Like, like, sure e okay, we’ve known that, we’ve never done anything about what you gonna doswyx: about it. Yeah.Marc: One possibility is just leave, leave them unprotected forever and live in a world of like, asymmetric terrorism forever. Or the other is take the problem seriously and figure out the set of techniques and technologies required to, to be able to deal with that.Whether those are lasers or jammers or early warning systems, or, you know, allswyx: personal force fields,Marc: kinetic, personal for dune, uh, personal, personal force fields. Exactly. And in both cases, the, these are, these are economic asymmetries. These are economic asymmetries, right? ‘cause it’s really cheap to field a bot, but it’s very hard to tell something, a bot.It’s very cheap to field a drone. It’s very hard. It’s very expensive to defend against a drone. But you see what I’m saying is it’s, it’s, it’s the, it’s the virtual version of the problem, and it’s the physical version of the problem. Uh, the virtual version of the problem. What we, what we need quite literally is proof of human.The reason is because you’re, you’re, you’re not gonna have proof of bot. The, the, the, especially now the, the bots are too good. The, the, the bots can pass the Turing test. And if the bots can pass the Turing test, then you can’t, you can’t screen for bot. You can’t have proof of not a bot. But what you can have is you can have proof of human, you can have, you know, cryptographically validated, this is definitely a person, and this is, and then you can have cryptographically validated.This is definitely like something that a person said, yeah, this video is real. Right. Um,swyx: just to double click on, on, uh, do you think Alex Lanya with world? Yeah. Do you think he’s got it or is there an alternative?Marc: Oh, so I mean, there’s gonna be, I think there’ll be, I think many people will try, we’re one of the key, you know, participants in, in, in the World, in the World Project.I dunno that, yeah. So we’re, we’re partisans, but yeah, I, I think so we think world is exactly correct. Okay. And, and the reason is it, it has, it has to be, it, it has to be proof of human. It it has, because you can’t do proof of not bought. You have to do proof of human to do proof of human. You, you need, you need biological validation.You, you needed to start with this was actually a person, right? Because otherwise your bot signing up as fake people. Right? So you, you have to have like something, you have to have a bi. Biometric. And then you have to have cryptographic validation. And then the ability to do, to do, to do the lookup. And then, by the way, the other thing you need, which that you, you also need selective disclosure.Um, so you need to be able to do proof of human without reviewing privacy, all the underlying information. Privacy. Yeah. By the way, another thing you’re need, you’re gonna need proof of age, right? ‘cause there’s all these laws in all these different countries now around you need to be 13 or 16 or 18 or whatever to do different things.And so you’re gonna, you’re gonna need a, you know, sort of validated proof of age, um, you know, to be able to legally operate, right? And so that, that’s coming. And then you’re gonna want like, proof of credit score and, you know, proof of like, you know, a hundred other things.swyx: That’s a tricky one.Marc: It is a tricky one, but you’re gonna, you’re gonna, there, there’s no reason, like if somebody’s checking on your credit, somebody shouldn’t, I’ll give you an example.Somebody shouldn’t need to know your name in order to be able to find out whether you’re credit worthy.swyx: Right? I see. Independently verifiable pieces of information.Marc: Pieces of information, yeah. It’s like selectively disclosed. And this is the answer to the privacy problem wr large, which is, I, I only need to prove, I need to prove at that moment.So like, you’re gonna need that. And I, I think their, their, their architecture makes sense. So that needs to get solved. I think language models have tipped, the bots are now too good. Uh, and, and, and so they’re undetectable. And so as a consequence, you, we now need to go confront that problem directly. And then, and like I said, and then the other problem is we, we need to go actually confront the drone problems.The Ukraine conflict has really unlocked a lot of thinking on that. And now the, um, and now the, the, the, the, the Iran situation is also unlocking that. And so I think there’s gonna be just like this incredible explosion of, of both drone and counter drones.swyx: Our drones are better than their drones to keep it that way.Marc: Yeah. Yeah. And counter drones,Alessio: I think we can sneak in one more question. Go for it. Um, I’m trying to tie together a lot of things that you said over the years. So at the Milken Institute debate with Teal, which is amazing. Um, you talked about the lag between a new technology and kinda like the GDP, um, impact of it.Marc: Yep.Alessio: The other idea you talked about is bourgeois capitalism and how, you know, this kind of managerial class was needed because of this complexity. And I think if you bring AI into the fold, you have like much higher leverage of people. So like if you have, you know, the Musk industries, um, and you give Elon a gi, you can run a lot more things That’s right.At once.Marc: That’s right.Alessio: And then you have the social contract. And I know you reviewed a clip of Sam ing, um, we’re rethinking the whole thing, and you’re like, absolutely not. Yes.Marc: Under,Alessio: and I wa I was in an event with Sam last night, uh, and he actually said in the last couple weeks it felt like now people are taking that seriously.Yeah. So I’m just curious like how you’re seeing the structure of organization changing, especially when you invest in early stage companies and, um, yeah, just like how the impact of. Work structure and, uh, all of that is playing out. Yeah.Marc: So there’s a whole bunch of, there’s a whole bunch of topics. I know, yeah.We, we could spend, and by the way, we’d be happy to spend more time, but we could, we could spend more time on all that. So just for people who haven’t followed this, so the, this, this, this term managerial comes from this thinker in the 20th century, James Burnham, who, um, just one of the great kind of 20th century political thinkers, um, societal thinkers.And he sort of said a as, and he was writing in like the 1940s, 1950s. Um, and he said kind of the, the whole history, capitalism until that point had been in two phases. Number one had been what he called bourgeois capitalism, which was think about as like name on the door, like Ford Motor Company. ‘cause Henry Ford runs the company.Um, and Henry, it’s like a DIC dictatorial model. And Henry Ford just like tells everybody what to do. And he said the problem with bourgeois capitalism is it doesn’t scale. ‘cause Henry Ford can only tell so many people to do so many things. And then he runs at a time in the day. And so, um, he said the second phase of capitalism was what he called managerial capitalism, which was the creation of a professional class of managers, um, that are trained not to be like.Car experts or to be whatever experts in any particular field, but are trained to be experts in management. And then that led to, you know, the importance of like Harvard business, you know, business schools and management consulting firms and all these things. And then you look at every big company today, and like most of the executives at most of the Fortune 500 companies are not domain experts in whatever the company does.And they’re certainly not the founders of those, but they’re professional managers. And in fact, in the course of their careers, they’ll probably manage many different kinds of businesses. They’ll rotate around and they might work in healthcare for a while and then work in financial services and then go work in something else, you know, come work in tech.And what Burnham said is he said that transition is absolutely required because the, the, the, the problem with bourgeois capitalism is, is it doesn’t scale. Henry Ford doesn’t scale. And so if you’re gonna run capitalist enterprises that are gonna have millions to billions of customers, um, you’re gonna need to, you’re, they’re gonna be operating a level of scale and complexity that’s gonna require this professional management class.And he said, look, the, the professional management class has its downsides. Like they’re not necessarily experts at doing the thing. They’re not as inventive, you know, they’re not gonna create the next breakthrough thing. But he is like, whether you think that’s good or bad or whatever is what’s gonna be required.And basically that’s what happened. Right. And so he wrote that book originally in like 1940, you know, over the course of the next 50 years, basically. Managerialism. Well, I mean, today, up till today, managerial managerialism basically took over everything. Mm-hmm. And you know, what I’m describing is basically how all big companies run and how all governments run and how are large scale nonprofits run and kind of everything, you know, everything runs basically what, what, what Venture Capital does is we basically are a rump, uh, sort of protest movement to that.To try to find the next Henry Ford or, or just to say El Elon Musk or, or the next, or the next Elon Musk or the next Steve Jobs, or the next Bill Gays. The next Mark Zuckerberg. And so we, we, we, we start these companies in, in the old model, right? We, we, we start them out as, as, as, as in the Henry Ford model.And so we start them out with a founder or a, or a, or a founder with, with colleagues. But you know, there’s the a founder, CEO, um, and then we basically bet that we basically bet that the startup is going to be able to do things, specifically innovate in ways that the big incumbents in that industry are not gonna be able to do.And so it’s a bet that by, basically by relighting this sort of name on the door, you know, kind of thing. Mm-hmm. This new innovative thing with like a king monarchical, uh, uh, political structure, um, that they’re gonna be able to innovate in a way that the incumbent is not going to be able to because the incumbent is, is being run by managers.Right. And, and, and, and by the way, and of course venture being what it is, sometimes that works, sometimes it doesn’t. But we’re, we’re constantly doing that, but I’ve always viewed it my entire life as like, we’re like raging against the dying of the light. Mm-hmm. Like we’re, we’re, we’re, we’re sort of constantly trying to fight off managerialism, just basically swamping everything and everything.Getting basically boring and gray and dumb and old. Right. And we’re trying to keep some level of energy vitality in the system. AI is the thing that would lead you to think, wow, maybe there’s a third model.Alessio: Mm-hmm.Marc: Right? And, and maybe may and way to think about it would be, maybe it’s a combination of the two, maybe the new Henry Ford or the new Elon or the new Steve Jobs plus ai, right.Is the best of both. Right. Because it’s, it’s, it’s sort of the spark of genius of the name on the door model, the Henry Ford model. But then it’s give that person AI superpowers to do all the managerial stuff and let the boss draw the managerial stuff. That may be the actual secret formula. And we’ve never even known that we wanted this because we never even thought it was a possibility.But I mean, you know, this, what is the thing that these bots are really good, they’re really good at doing paperwork. Like they’re really good at filling out forms, right? Like they’re really good at writing reports, they’re really good at reading, they’re really good at doing all the managerial work. Like they’re amazing at it.And so, yeah, so I, I think, I think the, I a hundred percent, I think the answer, the answer very well might be to get the best, best of both worlds by doing this. And then the challenge is gonna be twofold. The challenge is gonna be for the innovators to really figure out how to leverage AI actually do this.Right? Um, and, and then, and then the, the other challenge is gonna be for the, for the incumbents that are managerial, to figure out like, okay, what does that mean? ‘cause now they’re gonna, they’re, they’re gonna be facing a different kind of insurgent competitor that has a different set of capabilities than they’re used to.And so th the, this really I think is gonna force a lot of big companies to kind of figure out innovation. EE either I say figure out innovation or die trying.Alessio: Do you feel like that structure accelerates the impact on the actual GDPN economy? If you look at Space Act? Yes. The growth is like so fast. Yeah.And like, instead of having these companies kind of like Peter out in growth and impact, they can kind of like keep going if not accelerating.Marc: Yeah, that’s for sure. The hope, um, the, the, the challenge and, and you know, and, and look, the AI utopian view is of course, of course. And, and, and that’s gonna be the future of the economy.And it’s gonna grow 10 x and a hundred x and a thousand x. And we’re entering this regime of like much higher economic growth forever and consumer cornucopia of everything. And it’s, it’s gonna be great. And I, and, and I hope that’s true. I hope that’s, that’s like the u you know, that’s the current kind of utopian vision.I hope that’s true. The problem is, it goes back again. The real world is really messy. Um, and I’ll give you an example of how the real world is really messy. It requires 900 hours of professional certification training to become a hairdresser in the state of California. Um, so it’s like 35% of the economy, something like that.You have to get some sort of professional certification to do the job, which is to say that the, the professions are all cartels, right? Yeah. And so you have to get licensed as a doctor. You have to get licensed as a lawyer, you have to get licensed as a. You have to get into a union. Mm-hmm. Um, by the way, to, to work for the government, you need to be, you, you have both civil service protections and you have public sector unions.You have two layers of insulation, uh, against ever getting fired for anything or anything. Anything ever changing. I’ll give you another example. The the dock work. The dock workers one on strike a couple years ago. Mm-hmm. ‘cause they, you know, robotics, you know, if, if you go look at a modern dock, like in Asia, it’s all robots.If you go to American dock, it’s like all still guys, dragon, dragon stuff, by by hand, the dock works. Goes on a strike. It turns out there are 25,000 dock workers working on, on, on, on Docs in America. It turns out they have incredible political power. Mm-hmm. Because it’s a, it’s, it’s one of these un unified blocks of things.They won their strike and so they got commitments from the dock owners to not implement more automation. We learned a couple things in that. So number one, we learned that even a union as small as 25,000 people still has like tremendous political stroke. We also learned that they, it actually turns out the Dock Workers Union has 50,000 people in it.‘cause there’s 20, they have 25,000 people working in the docks. They have 25,000 people during full paycheck sitting at home from prior union agreements. Oh myswyx: God.Marc: From prior union agreements. I’ll give you another great example. There are government agencies, there are federal government agencies where the employees right of have civil service protections and there are in public sector unions.There are entire federal government agencies that struck new collective bargaining agreements during COVID, where not only are they have their jobs guaranteed in perpetuity, but they only have to report to work in an office one day per month. And so there are entire office buildings in Washington DC that are empty 29 outta 30 days of the year that are still operating and are still, we’re all still paying for it.20 and say, and then what they do, it turns out what the employees do is they’re very, they’re very smart in, in, in this way. And so they figure out, they come in on the last day of a month and the first day of the next month. And so and so, they’re, so, they’re in there, they’re in the office two days per 60 days, which means these buildings are empty for 58 days at a time.And you see what I’m, you see where I’m heading with this? Like this is like locked in, right? This is like locked in in a way that has nothing to do with like, and people say capitalist, it’s like anticapitalistic. It’s like, it’s, it’s basically it’s restrictions on trade, it’s restrictions on the ability to like change the workforce.And so, so much of our economy is, is, you know, the, the, I I’m, I’m describing the entire healthcare system. I’m describing the entire legal profession. I’m describing the entire housing industry. I’m describing the entire education system, right? K through 12 schools in the United States. They’re a literal government monopoly.How are we gonna apply AI and education? The answer is we’re not, because it’s a literal government monopoly, it is never going to change the end. And there is nothing to do, by the way, you can create an entirely new school system. Like that’s the one thing you can do, is you can do what Alpha School’s doing.You can create an entirely new school system. Other than that, you’re not gonna go in and change what’s happening in the American classroom, like K through 12. There’s no chance the teachers are 100% opposed to it. It’s a hundred percent not gonna happen. So, so you see what I’m saying is like there’s this like massive slippage that’s gonna take place.Both the AI utopians and the AI dors are far too optimistic.swyx: Right.Marc: You see what I’m saying? Be because they believe that because the technology makes something possible that 8 billion people all of a sudden are gonna change how they behave. And it’s just like, nope. So much of how the existing economy works.Mm-hmm. It’s just, it. It’s just like wired in. And so we’re gonna be lucky as a society, we’re gonna be lucky if AI adoption happens quickly. Right. Because if it doesn’t, what we’re just gonna have is stagnation.Alessio: Awesome. Mark. I know you gotta run.swyx: Yeah. We all know or still welcome. But, uh, it was such a pleasure talking to you.Uh, we’re truly living in the age of science fiction coming to real life.Marc: Yes. Yes. Could not be more exciting. Yeah. Really. Thank you, mark. You guys awesome.swyx: Thank That’s it.Marc: Good. Thank you. That’s it.

This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit www.latent.space/subscribe

The podcast and artwork embedded on this page are from Latent.Space, which is the property of its owner and not affiliated with or endorsed by Listen Notes, Inc.

Podcast: Latent Space: The AI Engineer Podcast (LS 45 · TOP 1% what is this?)
Episode: Extreme Harness Engineering for Token Billionaires: 1M LOC, 1B toks/day, 0% human code, 0% human review — Ryan Lopopolo, OpenAI Frontier & Symphony
Pub date: 2026-04-07

Get Podcast Transcript →
powered by Listen411 - fast audio-to-text and summarization

We’re proud to release this ahead of Ryan’s keynote at AIE Europe. Hit the bell, get notified when it is live! Attendees: come prepped for Ryan’s AMA with Vibhu after.

Move over, context engineering. Now it’s time for Harness engineering and the age of the token billionaires.

Ryan Lopopolo of OpenAI is leading that charge, recently publishing a lengthy essay on Harness Eng that has become the talk of the town:

In it, Ryan peeled back the curtains on how the recently announced OpenAI Frontier team have become OpenAI’s top Codex users, running a >1m LOC codebase with 0 human written code and, crucially for the Dark Factory fans, no human REVIEWED code before merge. Ryan is admirably evangelical about this, calling it borderline “negligent” if you aren’t using >1B tokens a day (roughly $2-3k/day in token spend based on market rates and caching assumptions):

Over the past five months, they ran an extreme experiment: building and shipping an internal beta product with zero manually written code. Through the experiment, they adopted a different model of engineering work: when the agent failed, instead of prompting it better or to “try harder,” the team would look at “what capability, context, or structure is missing?”

The result was Symphony, “a ghost library” and reference Elixir implementation (by Alex Kotliarskyi) that sets up a massive system of Codex agents all extensively prompted with the specificity of a proper PRD spec, but without full implementation:

The future starts taking shape as one where coding agents stop being copilots and start becoming real teammates anyone can use and Codex is doubling down on that mission with their Superbowl messaging of “you can just build things”.

Across Codex, internal observability stacks, and the multi-agent orchestration system his team calls Symphony, Ryan has been pushing what happens when you optimize an entire codebase, workflow, and organization around agent legibility instead of human habit.

We sat down with Ryan to dig into how OpenAI’s internal teams actually use Codex, why the real bottleneck in AI-native software development is now human attention rather than tokens, how fast build loops, observability, specs, and skills let agents operate autonomously, why software increasingly needs to be written for the model as much as for the engineer, and how Frontier points toward a future where agents can safely do economically valuable work across the enterprise.

We discuss:

* Ryan’s background from Snowflake, Brex, Stripe, and Citadel to OpenAI Frontier Product Exploration, where he works on new product development for deploying agents safely at enterprise scale

* The origin of “harness engineering” and the constraint that kicked off the whole experiment: Ryan deliberately refused to write code himself so the agent had to do the job end to end

* Building an internal product over five months with zero lines of human-written code, more than a million lines in the repo, and thousands of PRs across multiple Codex model generations

* Why early Codex was painfully slow at first, and how the team learned to decompose tasks, build better primitives, and gradually turn the agent into a much faster engineer than any individual human

* The obsession with fast build times: why one minute became the upper bound for the inner loop, and how the team repeatedly retooled the build system to keep agents productive

* Why humans became the bottleneck, and how Ryan’s team shifted from reviewing code directly to building systems, observability, and context that let agents review, fix, and merge work autonomously

* Skills, docs, tests, markdown trackers, and quality scores as ways of encoding engineering taste and non-functional requirements directly into context the agent can use

* The shift from predefined scaffolds to reasoning-model-led workflows, where the harness becomes the box and the model chooses how to proceed

* Symphony, OpenAI’s internal Elixir-based orchestration layer for spinning up, supervising, reworking, and coordinating large numbers of coding agents across tickets and repos

* Why code is increasingly disposable, why worktrees and merge conflicts matter less when agents can resolve them, and what it really means to fully delegate the PR lifecycle

* “Ghost libraries”, spec-driven software, and the idea that a coding agent can reproduce complex systems from a high-fidelity specification rather than shared source code

* The broader future of Frontier: safely deploying observable, governable agents into enterprises, and building the collaboration, security, and control layers needed for real-world agentic work

Ryan Lopopolo

* X: https://x.com/_lopopolo

* Linkedin: https://www.linkedin.com/in/ryanlopopolo/

* Website: https://hyperbo.la/contact/

Timestamps

00:00:00 Introduction: Harness Engineering and OpenAI Frontier00:02:20 Ryan’s background and the “no human-written code” experiment00:08:48 Humans as the bottleneck: systems thinking, observability, and agent workflows00:12:24 Skills, scaffolds, and encoding engineering taste into context00:17:17 What humans still do, what agents already own, and why software must be agent-legible00:24:27 Delegating the PR lifecycle: worktrees, merge conflicts, and non-functional requirements00:31:57 Spec-driven software, “ghost libraries,” and the path to Symphony00:35:20 Symphony: orchestrating large numbers of coding agents00:43:42 Skill distillation, self-improving workflows, and team-wide learning00:50:04 CLI design, policy layers, and building token-efficient tools for agents00:59:43 What current models still struggle with: zero-to-one products and gnarly refactors01:02:05 Frontier’s vision for enterprise AI deployment01:08:15 Culture, humor, and teaching agents how the company works01:12:29 Harness vs. training, Codex model progress, and “you can just do things”01:15:09 Bellevue, hiring, and OpenAI’s expansion beyond San Francisco

Transcript

Ryan Lopopolo: I do think that there is an interesting space to explore here with Codex, the harness, as part of building AI products, right? There’s a ton of momentum around getting the models to be good at coding. We’ve seen big leaps in like the task complexity with each incremental model release where if you can figure out how to collapse a product that you’re trying to.

Build a user journey that you’re trying to solve into code. It’s pretty natural to use the Codex Harness to solve that problem for you. It’s done all the wiring and lets you just communicate in prompts. To let the model cook, you have to step back, right? Like you need to take a systems thinking mindset to things and constantly be asking, where is the Asian making mistakes?

Where am I spending my time? How can I not spend that time going forward? And then build confidence in the automation that I’m putting in place. So I have solved this part of the SDLC.

swyx: [00:01:00] All right.

[00:01:03] Meet Ryan

swyx: We’re in the studio with Ryan from OpenAI. Welcome.

Ryan Lopopolo: Hi,

swyx: Thanks for visiting San Francisco and thanks for spending some time with us.

Ryan Lopopolo: Yeah, thank you. I’m super excited to be here.

swyx: You wrote a blockbuster article on harness engineering. It’s probably going to be the defining piece of this emerging discipline, huh?

Ryan Lopopolo: Thank you. It is it’s been fun to feel like we’ve defined the discourse in some sense.

swyx: Let’s contextualize a little bit, this first podcast you’ve ever done. Yes. And thank you for spending with us. What is, where is this coming from? What team are you in all that jazz?

Ryan Lopopolo: Sure, sure.

Ryan Lopopolo: I work on Frontier Product Exploration, new product development in the space of OpenAI Frontier, which is our enterprise platform for deploying agents safely at scale, with good governance in any business. And. The role of VMI team has been to figure out novel ways to deploy our models into package and products that we can sell as solutions to enterprises.

swyx: And you have a background, I’ll just squeeze it in there. Snowflake, brick, [00:02:00] stripe, citadel.

Ryan Lopopolo: Yes. Yes. Same. Any kind of customer

swyx: entire life. Yes. The exact kind of customer that you want to,

Vibhu: so I’ll say, I was actually, I didn’t expect the background when I looked at your Twitter, I’m seeing the opposite.

Stuff like this. So you’ve got the mindset of like full send AI, coding stuff about slop, like buckling in your laptop on your Waymo’s. Yes. And then I look at your profile, I’m like, oh, you’re just like, you’re in the other end too. Oh, perfect. Makes perfect.

Ryan Lopopolo: I it’s quite fun to be AI maximalist if you’re gonna live that persona.

Open eye is the place to do it. And it’s

swyx: token is what you say.

Ryan Lopopolo: Yeah. Certainly helps that we have no rate limits internally. And I can go, like you said, full send at this stay.

swyx: Yeah. Yeah. So the Frontier, and you’re a special team within O Frontier.

Ryan Lopopolo: We had been given some space to cook, which has been super, super exciting.

[00:02:47] Zero Code Experiment

Ryan Lopopolo: And this is why I started with kind of a out there constraint to not write any of the code myself. I was figuring if we’re trying to make agents that can be deployed into end to enterprises, they should be [00:03:00] able to do all the things that I do. And having worked with these coding models, these coding harnesses over 6, 7, 8 months, I do feel like the models are there enough, the harnesses are there enough where they’re isomorphic to me in capability and the ability to do the job.

So starting with this constraint of I can’t write the code meant that the only way I could do my job was to get the agent to do my job.

Vibhu: And like a, just a bit of background before that. This is basically the article. So what you guys did is five months of working on an internal tool, zero lines of code over a mi, a million lines of code in the total code base.

You say it was cenex, more like it was cenex faster than you would’ve. If you had done it by end. So

Ryan Lopopolo: yeah, that

Vibhu: was the mindset going into this, right?

Ryan Lopopolo: That’s right.

[00:03:46] Model Upgrades Lessons

Ryan Lopopolo: Started with some of the very first versions of Codex CLI, with the Codex Mini model, which was obviously much less capable than the ones we have today.

Which was also a very good constraint, right? Quite a visceral feeling to ask the [00:04:00] model to build you a product feature. And it just not being able to assemble the pieces together.

Which kind of defined one of the mindsets we had for going into this, which is whenever the model just cannot, you always pop open at the task, double click into it, and build smaller building blocks that then you can reassemble into the broader objective.

And it was quite painful to do this. Honestly, the first month and a half was. 10 times slower than I would be. But because we paid that cost, we ended up getting to something much more productive than any one engineer could be because we built the tools, the assembly station for the agent to do the whole thing.

[00:04:43] Model Generations, Build Systems & Background Shells

Ryan Lopopolo: But yeah, so onward to G BT 5, 5, 1, 5, 2, 5, 3, 5 4. To go through all these model generations and see their kind of corks and different working styles also meant we had to adapt the code base to change things up when the model was revved. [00:05:00] One interesting thing here is five two, the Codex harness at the time did not have background shells in it, which means we were able to rely on blocking scripts to perform long horizon work.

But with five, three and background shells, it became less patient, less willing to block. So we had to retool the entire build system to complete in under a minute and. This is not a thing I would expect to be able to do in a code base where people have opinions. But because the only goal was to make the Asian productive over the course of a week, we went from a bespoke make file build to Basil, to turbo to nx and just left it there because builds were fast at that point.

swyx: Interesting. Talk more about Turbo TenX. That’s interesting ‘cause that’s the other direction that other people have been doing.

Ryan Lopopolo: Ultimately I have. Not a lot of experience with actual frontend repo architecture.

swyx: You’re talking that Jessica built the sky. So I’m like, I know the NX team. I know Turbo from Jared [00:06:00] Palmer.

And I’m like, yeah, that’s an interesting comparison.

[00:06:02] One Minute Build Loop

Ryan Lopopolo: The hill we were climbing right, was make it fast.

swyx: Is there a micro front end involved? Is it how how complex react

Ryan Lopopolo: electron base single app sort of thing

swyx: And must be under a minute. That’s an interesting limitation. I’m actually not super familiar with the background shelf stuff.

Probably was talked about in the fight three release.

Ryan Lopopolo: BA basically means that codex is able to spawn commands in the background and then go continue to work while it waits for them to finish. So it can spawn an expensive build and then continue reviewing the code, for example.

swyx: Yeah.

Ryan Lopopolo: And this helps it be more time efficient for the user invoking the harness.

swyx: And I guess and just to really nail this, like what does one minute matter? Like why not five, okay, good. We want no. We

Ryan Lopopolo: want the inner loop to be as fast as possible. Okay. One minute was just a nice round number and we were able to hit it.

swyx: And if it doesn’t complete, it kills it or some something,

Ryan Lopopolo: No.

We just take that as a signal that we need to stop what we’re doing, double click, decompose a build graph a bit to get us to high back under so that we [00:07:00] can able the agent continue to operate.

swyx: It’s almost like you’re, it’s like a ratchet. It’s like you’re forcing build time discipline, because if you don’t, it’ll just grow and grow.

That’s right. And you mentioned that my current, like the software I work on currently is at 12 minutes. It sucks.

Ryan Lopopolo: This has been my experience with platform teams in the past, where you have an envelope of acceptable build times and you let it go up to breach and then you spend two, three weeks to bring it back down to the lower end of the average low bed stop.

But because tokens are so cheap Yeah. And we’re so insanely parallel with the model, we can just constantly be gardening this thing to make sure that we maintain these in variants, which means. There’s way less dispersion in the code and the SDLC, which means we can simplify in a way and rely on a lot more in variance as we write the software.

[00:07:45] Observability, Traces & Local Dev Stack

Vibhu: Lovely.

[00:07:46] Humans Are Bottleneck

Vibhu: You mentioned in your article, like humans became the bottleneck, right? You kicked off as a team of three people. You’re putting out a million line of code, like 1500 prs, basically. What’s the mindset there? So as much as code is disposable, you’re doing a lot of review. A lot [00:08:00] of the article talks about how you wanna rephrase everything is prompting everything, is what the agent can’t see.

It’s kind of garbage, right? You shouldn’t have it in there. So what’s like the high level of how you went about building it, and then how you address okay, humans are just PR review. Like how is human in the loop for this?

Ryan Lopopolo: We’ve moved beyond even the humans reviewing the code as well.

[00:08:19] Human Review, PR Automation & Agent Code Review

Ryan Lopopolo: Most of the human review is post merge at this point.

But post, post merge, that’s not even reviewed. That’s just

swyx: Oh, let’s just make ourselves happy by You

Ryan Lopopolo: haven’t used fundamentally. The model is trivially paralyzable, right? As many GPUs and tokens as I am willing to spend, I can have capacity to work with my hood base.

The only fundamentally scarce thing is the synchronous human attention of my team. There’s only so many hours in the day we have to eat lunch. I would like to sleep, although it’s quite difficult to, stop poking the machine because it makes me want to feed it. You have to step back, right?

Like you need to take a systems thinking mindset to things and [00:09:00] constantly be asking where is the agent making mistakes? Where am I spending my time? How can I not spend that time going forward? And then build confidence in the automation that I’m putting in place. So I have solved this part of the SDLC, and usually what that has looked like is like we started needing to pay very close attention to the code because the agent did not have the right building blocks to produce.

Modular software that decomposed appropriately that was reliable and observable and actually accrued a working front end in these things, right?

[00:09:35] Observability First Setup

Ryan Lopopolo: So in order to not spend all of our time sitting in front of a terminal at most, doing one or two things at a time, invested in giving the model that observability, which is that that graph in the post here.

swyx: Yeah. Let’s walk through this traces and which existed first

Ryan Lopopolo: we started with just the app and the whole rest of it. From vector through to all these login metrics, APIs was, I dunno, half an [00:10:00] afternoon of my time. We have intentionally chosen very high level fast developer tools. There’s a ton of great stuff out there now.

We use me a bunch, which makes it trivial to pull down all these go written Victoria Stack binaries in our local development. Tiny little bit of python glue to spin all these up. And off you go. One neat thing here is we have tried to invert things as much as possible, which is instead of setting up an environment to spawn the coding agent into, instead we spawn the coding agent, like that’s the entry point.

It’s just Codex. And then we give Codex via skills and scripts the ability to boot the stack if it chooses to, and then tell it how to set some end variables. So the app and local Devrel points at this stack that it has chosen to spin up. And this I think is like the fundamental difference between reasoning models and the four ones and four ohs of the past, where these models could not think so you had to put them in [00:11:00] boxes with a predefined set of state transitions.

Whereas here we have the model, the harness be the whole box. And give it a bunch of options for how to proceed with enough context for it to make intelligent choices. So

Vibhu: sales, so like a lot of that is around scaffolding, right? Yes. Previous agents, you would define a scaffold. It would operate in that.

Lube, try again. That’s pivoted off from when we’ve had reasoning models. They’re seeming to perform better when you don’t have a scaffold, right? That’s right.

[00:11:28] Docs Skills Guardrails

Vibhu: And you go into like niches here too, like your SPEC MD and like having a very short agent MG Agent md.

swyx: Yes. Yes.

Vibhu: Yeah. So you even lay out what it is here, but I like

swyx: the table contents.

Vibhu: Yeah.

swyx: Like stuff like this, it really helps guide people because everyone’s trying to do this.

Ryan Lopopolo: This structure also makes it super cheap to put new content into the repository to steer both the humans and the agents.

swyx: You, you reinvented skills, right?

Vibhu: One big agents and

swyx: skills from first princip holds

Ryan Lopopolo: all skills did not exist when we started doing this.

Vibhu: You have a short [00:12:00] one 100 line overall table of contents and then you have little skills, right? Core beliefs, MD tech tracker. Yeah. Yeah. The scale is over

Ryan Lopopolo: The tech jet tracker and the quality score are pretty interesting because this is basically a tiny little scaffold, like a markdown table, which is a hook for Codex to review all the business logic that we have defined in the app, assess how it matches all these documented guardrails and propose follow up work for itself.

Before beads and all these ticketing systems, we were just tracking follow up work as notes in a markdown file, which, we could spa an agent on Aron to burn down. There’s this really neat thing that like the models fundamentally crave text. So a lot of what we have done here is figure out ways to inject text

swyx: into

Ryan Lopopolo: the system right when we get a page, because we’re missing a timeout, for example.

I can just add Codex in Slack on that page and say, I’m gonna fix this by adding a timeout. Please update our reliability documentation. To require that all network calls have [00:13:00] timeouts. So I have not only made a point in time fix, but also like durably encoded this process knowledge around what good looks like.

swyx: Yeah.

Ryan Lopopolo: And we give that to the root coding agent as it goes and does the thing. But you can also use that to distill tests out of, or a code review agent, which is pointed at the same things to narrow the acceptable universe of the code that’s produced.

swyx: I think one of the concerns I have with that kind of stuff is you think you’re making the right call by making, it’s persisted for all time across everything.

Yes. But then you didn’t think about the exceptions that you need to make, right? And that you have to roll it back.

Vibhu: Part of it is

swyx: also sometimes it can follow your s instructions too.

Vibhu: It’s somewhat a skill, right? So it determines when it uses the tools, right? Like it’s not like it’ll run outta every call.

It’ll determine when it wants to check quality score, right?

Ryan Lopopolo: Yeah. And we do in the prompts we give these agents, allow them to push back,

[00:13:51] Agent Code Review Rules

Ryan Lopopolo: When we first started adding code review agents to the pr, it would be Codex, CLI. Locally writes the change, pushes up a PR on [00:14:00] those PR synchronizations of review agent fires.

It posts a comment. We instruct Codex that it has to at least acknowledge and respond to that feedback. And initially the Codex driving the code author was willing to be bullied by the PR reviewer, which meant you could end up in a situation where things were not converging. So yeah, we had to,

swyx: he’s just a thrash.

Ryan Lopopolo: We had to add more optionality to the prompts on both of these things, right? The reviewer agents were instructed to bias toward merging the thing to not surface anything greater than a P two in priority. We didn’t really define P two, but we gave it, you

swyx: did define P two.

Ryan Lopopolo: We gave it a framework within which to score its output

swyx: and then greater than P zero is worse, right?

Yes. P two is very good.

Ryan Lopopolo: P zero is you will mute the code place if

swyx: you merch this

Ryan Lopopolo: thing, right?

swyx: Yeah.

Ryan Lopopolo: But also on the code authoring agent side, we also gave it the flexibility to either defer or push back against review feedback, right? This happens all the time, right? Like I happen to notice something and leave a code review, [00:15:00] which.

Could blow up the scope by a factor of two. I usually don’t mean for that to be addressed Exactly. In the moment. It’s more of an FYI file it to the backlog, pick it up in the next fix it week sort of thing. And without the context that this is permissible, the coding agents are gonna bias toward what they do, which is following instructions.

swyx: Yeah.

[00:15:19] Autonomous Merging Flow

swyx: I do wanted to check in on a couple things, right? Sure. All the coding review agent, it can merge autonomously. I think that’s something that a lot of people aren’t comfortable with. And you have a list here of how much agents do they do Product code and tests, CI configuration and release tooling, internal Devrel tools, documentation eval, harness review, comments, scripts that manage the repository itself, production dashboard definition files, like everything.

Yes. And so they’re just all churning at the same time, is there like a record that, that any human on the team pulls to stop everything

Ryan Lopopolo: Because we are building a native application here. We’re not doing continuous deploy. So there’s still a human in the loop for cutting the release branch.

I see. We require a blessed [00:16:00] human approved smoke test of the app before we promote it to distribution, these sort of things.

swyx: So you’re working on the app, you’re not building like infrastructure where you have like nines of reliability, that kinda stuff?

Ryan Lopopolo: That’s correct. That’s correct. Okay. And also like full recognition here that all of this activity took in a completely greenfield repository.

There’s. Should be no script that this applies generally to

swyx: this is a production thing, you’re gonna ship

Ryan Lopopolo: to

swyx: customers. Of course. Yeah, of course. So this is real

Vibhu: And like one of the things there is, you mentioned you started this as a repo from scratch. The onboarding first month or so was pretty, it was like working backwards, right?

Yeah. And then you had to work with the system and now you’re at that point where you know, you’re very autonomous. I’m curious like, okay, so what, how human in the loop is it? So what are the bottlenecks that you wish you could still automate? And part of that is also like, where do you see the model trajectory improving and offloading more human in the loop?

We just got 5.4. It’s a really good,

Ryan Lopopolo: fantastic model, by the way.

Vibhu: Yeah. Yeah. It’s the first one that’s merged. Top tier coding. So it’s codex level coding and reasoning. So general reasoning both in one model. So

Ryan Lopopolo: and

Vibhu: computer [00:17:00] use vision.

Ryan Lopopolo: Now we now with five four, I can just have Codex write the blog post, whereas for this one I had to balance between chat.

swyx: Oh, I need to, I might be out of a job. Oh my God.

Ryan Lopopolo: Oh,

swyx: I know. You just gave me an idea for a completely AI newsletter that five four could do. Yeah, I get it Now.

Ryan Lopopolo: This sort of thing is just one example of closing the loop, right? Like the dashboard thing you mentioned. We have Codex authoring the Js ON, for the Grafana dashboards and publishing them and also responding to the pages, which means when it gets the page, it knows exactly which dashboards are defined and what alerts.

What alert was triggered by which exact log in the code base. ‘cause all of this stuff is collated together.

swyx: It has to own everything.

Yes. Yeah. Yeah.

Ryan Lopopolo: And it means that if we have an outage that did not result in a page. It has the existing set of dashboards available to it. It has the existing set of metrics and logs and can figure out where the gaps in the dashboard are or [00:18:00] in the underlying metrics and fix them in one go.

In the same way, you would have a full stack engineer be able to drive a feature from the backend all the way to the front end.

Vibhu: So it, it seems like a lot of the work you guys had to do was you as a small team are fully working for a way that the model wants the software to be written. It’s like less human legible for better. Code legibility, agent legibility. How do you think that affects broader teams? So one at OpenAI, do liaison, like this is how software should be written. Like I can imagine, say you join a new team with this methodology, this mindset there’s ways that, teams do code review, teams write code, like teams are structured and a lot of it is for human legibility.

So should we all swap? Like how does this play back one broader into OpenAI and then like broader into the software engineering, right? Is it like teams that pick this up will it’s pretty drastic, right? You have to make a pretty big switch. Should they just full send Yeah.

Ryan Lopopolo: The mindset is very much that I’m removed from the process, right? I can’t really have deep code level opinions about [00:19:00] things. It’s as if I’m. Group tech leading a 500 person organization.

Vibhu: Yeah.

Ryan Lopopolo: Like it’s not appropriate for me to be in the weeds on every pr. This is why that post merge code review thing is like a good analog here, right?

Like I have some representative sample of the code as it is written, and I have to use that to infer what the teams are struggling with, where they could use help, where they’re already moving quickly and I can pivot my focus elsewhere.

Vibhu: Yeah.

Ryan Lopopolo: So I don’t really have too many opinions around the code as it is written.

I do, however, have a command based class, which is used to have repeatable chunks of business logic that comes with tracing and metrics and observability for free. And the thing to focus on is not how that business logic is structured, but that it uses this primitive ‘cause I know that’s gonna give leverage by default.

Vibhu: Yeah.

Ryan Lopopolo: Yeah, back to that sort of systems stinking,

Vibhu: and you have part of that in your blog post, enforcing architecture and ta taste how you set boundaries for what’s used. There’s also a section on redefining [00:20:00] engineering and stuff, but yeah, it’s just, it’s interesting to hear,

Ryan Lopopolo: and as the models have gotten better, they have gotten better at proposing these abstractions to unblock themselves, which again, lets me move higher and higher up the stack to look deeper into the future on what ultimately blocked the team from shipping.

swyx: Yeah. You mentioned so you, this is primarily a, it is like a 1 million line of code base electron app. But it manages its own services as well, so it’s like a backend for front end type thing.

Ryan Lopopolo: We do have a backend in there, but that’s hosted in the cloud.

Yeah. This sort of structure is actually within the separate main and render processes

Within the

swyx: electric.

That’s just how electronic works.

Ryan Lopopolo: Yeah, of course. So have also treated like. MVC style decomposition with the same level of rigor, which has been very fun.

swyx: I have a fun pun. This is a tangent, NVC is model view controller. Any sort of full stack web Devrel knows that.

But my AI native version of this is Model view Claw, the clause the harness.

Ryan Lopopolo: That’s right. That’s right. I do think that there is an interesting space to [00:21:00] explore here with Codex, the harness as part of building AI products, right? There’s a ton of momentum around getting the models to be good at coding.

We’ve seen big leaps in like the task complexity with each incremental model release where if you can figure out how to collapse a product that you’re trying to build, a user journey that you’re trying to solve into code, it’s pretty natural to use the Codex Harness to solve that problem for you. It’s done all the wiring and lets you just communicate and prompts to let the model cook.

Yeah. It’s been very fun. And there’s also a very engineering legible way of increasing capabil. It’s fantastic, right? Yeah. Just give you, just give the model scripts, the same scripts you would already build for yourself.

swyx: Yeah.

Yeah. So for listeners, this is Ryan saying that software engineering or coding against will eat knowledge work like the non-coding parts that you would normally think.

Oh, you have to build a separate agent for it. No, start a coding agent and go out from there. Which open Claw has like it’s pie Underhood.

Ryan Lopopolo: [00:22:00] Yes.

Vibhu: Basically define your task in code. Everything is a coding

swyx: agent by the way. Since I brought it up, it’s probably the only place we bring it up. Is any open claw usage from you?

Any?

Ryan Lopopolo: No. No. Not for me. I don’t have any spare Mac Minis rattling around my house.

swyx: You can afford it? No. I just, I’m curious if it’s changed anything in opening eye yet, but it’s probably early days. And then the other, the other thing I, I wanna pull on here is like you mentioned ticketing systems and you mentioned prs and I’m wondering if both those things have to go away or be reinvented for this kind of coding.

So the git itself and is like very hostile to multi-agent.

Ryan Lopopolo: Yeah. We make very heavy use of work trees.

swyx: But like even then, like I just did a, dropped a podcast yesterday with Cursors saying, and they said they’re getting rid of work trees ‘cause it still has too many merge conflicts.

It’s still un too un unintuitive. But go ahead.

Ryan Lopopolo: The models are really great at resolving merge conflicts. Yeah. And to get to a state where I’m not synchronously in the loop in my terminal, I almost don’t care that there are merge

swyx: with disposable.

[00:23:00] Yeah.

Ryan Lopopolo: We invoke a dollar land skill and that coaches codex to push the PR Wait for human and agent reviewers Wait for CI to be green.

Fix the flakes if there are any merged upstream. If the PR comes into conflict, wait for everything to pass. Put it in the merge queue. Deal with flakes until it’s in Maine. End. This is what it means to delegate fully, right? This is in a, very large model re probably a significant tax on humans to get PRS merged, but the agent is more than capable of doing this and I really don’t have to think about it other than keep my laptop open.

swyx: Yeah. I used to be much more of a control freak, but now I’m like, yeah, actually you could do a better job of this than me. Yeah. With the right context. Yes.

[00:23:47] Encoding Requirements

swyx: Anything else in harness in general? Just this piece, I just wanna make sure we,

Ryan Lopopolo: I think one thing that I maybe didn’t make super clear in the article that I heard on Twitter as an interesting, that’s respond [00:24:00]

swyx: to them.

What’s the chatter and then what’s your response?

Ryan Lopopolo: Ultimately, all the things that we have encoded in docs and tests and review agents and all these things are ways to put all the non-functional requirements of building high scale, high quality, reliable software into a space that prompt injects the agent.

We either write it down as docs, we add links where the error messages tell how to do the right thing. So the whole meta of the thing is to basically tease out of the heads of all the engineers on my team, what they think good looks like, what they would do by default, or what they would coach a new hire on the team to do to get things to merch.

And that’s why we pay attention to all the mistakes, mistakes that the agent makes, right? This is code being written that is misaligned with some as yet not written down, non-functional requirement.

swyx: Sorry, what? Did the online people misunderstand or

Ryan Lopopolo: No,

swyx: what

you

Ryan Lopopolo: responded to? Somebody just literally said that.

I was like, oh yeah,

swyx: okay,

Ryan Lopopolo: This is the [00:25:00] thing. This is what I’ve been doing. Oh, you

swyx: agree? Yeah. I see. Interesting.

Ryan Lopopolo: One other neat thing, which I did totally did not expect is folks were just. Taking the link to the article and giving it to pi or Codex and say, make my repo this,

Vibhu: you achi a whole recursion.

Ryan Lopopolo: And it was wildly effective. Really? It was wildly effective. No

Vibhu: way. It just actually is something I tried with five, four yesterday. I didn’t have time. Last time I was like out speaking of something, and this is one of my things, I was like, okay, I have this article. Can we just scaffold out what it would be like to run this?

And I, I did it first as that and then I was like, okay, let me take another little side repo and say okay, if I was to fully automate this like this because I haven’t written a line of code, it’s

Ryan Lopopolo: like over full, set

Vibhu: it right. The side thing I’m doing of voice. TTS I’m just like, slobbing out, whatever.

It’s nothing production. I’m like, how would I make this like this? And it’s actually like a really good way. It’s like a good way to learn what could be changed, what could be like, it’s just a good analyzing, right? You give it all the codes, you give it all the context, you give it the article and it walks you through it very well.

That’s right. That’s right.

[00:25:57] Inlining Dependencies

[00:25:57] Dependencies Going Away & Brett Taylor’s Response

swyx: I guess one more thing before we go to Symphony is I wanted to cover [00:26:00] Brett Taylor’s response. We had him on the show. He is your chairman, which is wild. Yeah. That he’s reading your articles as well and like getting engaged in it. He says software dependencies are going away.

Basically they can just be like vendored. Yes. Response.

Ryan Lopopolo: A

swyx: hundred percent. A hundred percent agree. You still pro qr, you still pay Datadog. You still pay Temporal. Thank you.

Ryan Lopopolo: Yep. The level of complexity of the dependencies that we can internalize is, I would say low, medium right now. Just based on model capability.

What does the,

swyx: what is medium?

Ryan Lopopolo: I would say like a. A couple thousand line dependency is a thing that we could in-house No problem. Call in an afternoon of time. One neat thing about it is like probably most of that code you don’t even need. Like by in-house and abstraction, you can strip away all the generic parts of it and only focus on what you need to enable the specific thing.

Yes. You’re building,

swyx: I’ve been calling this the end of b******t plugins.

Ryan Lopopolo: Yeah.

swyx: Because there’s so much when I published an open source thing, I want to accept everything, be liberal. I want to accept, this is post’s law, but that means there’s so much bloat. Yes. There’s so much overhead.

Ryan Lopopolo: One other neat thing about [00:27:00] this too is when we deploy Codex Security on the repo, it is able to deeply review and change. The internalized dependencies in a much lower friction way than it would be to like, push patches upstream, wait for them to be released, pull them down, make sure that’s compatible with all the transitive I have in my repo and things like that.

So it’s also much lower friction to internalize some of these things if code is free. ‘cause the tokens are cheap sort of thing.

swyx: Yeah. Yeah. I think like the only argument I have against this is basically scale testing, which obviously the larger pieces of software like Linux, MySQL, he calls up even the Datadog and Temporals and then maybe security testing where Yes.

Classically, I think, is it linis tos, it said security open source is the best disinfectant.

Ryan Lopopolo: Many eyes.

swyx: Many eyes. And if inline your dependencies and code them up, you’re gonna have to relearn mistakes from other people that Yep.

Ryan Lopopolo: Yep. And to internalize that dependency, you’re back to zero and you have to start.

Reassembling all those bits and pieces to Yeah. Have [00:28:00] high confidence in the code as it is written. Yeah.

Vibhu: Even part of the first intro of this, you basically mentioned like everything was written by codex, including internal tooling, right? So internal tooling, like when you’re visualizing what’s going on it’s writing it for itself.

swyx: Yeah. I’m built internal tools way I now, and like I just show them off and they’re like, how long did you spend? And I didn’t spend any time. I just prompted it,

Ryan Lopopolo: very funny story here.

swyx: Yeah, go ahead.

Ryan Lopopolo: We had deployed our app to the first dozen users internally had some performance issues, so we asked them to export a trace for us get a tar ball, gave it to our on-call engineer, and he did a fantastic job of working with Codex to build this beautiful local Devrel tool, next JS app, the drag and drop the tar ball in, and it visualizes the entire trace.

It’s fantastic. Took an afternoon, but none of this was necessary. Because you could just spin up codex and give it the tar ball and ask the same thing and get the response immediately. So in a way, optimizing for human [00:29:00] legibility of that debugging process was wrong. It kept him in the loop unnecessarily when instead he could have just like Codex cooked for five minutes and gotten this same.

swyx: Yeah, you verify your instincts here of this is how we used to do it. Or this is how I would have used to solve it.

Ryan Lopopolo: Yeah. In this local observability stack. Like sure, you can de deploy Yeager to visualize the traces, but I wouldn’t expect to be looking at the traces in the first place because I’m not gonna write the code to fix them.

swyx: Yeah. So basically there needs to be like this kind of house stack and owning the whole loop. I think that is very well established. And it sounds like you might be like sharing more about that in the future, right?

Ryan Lopopolo: Yeah. I think we’re excited to do

[00:29:36] Ghost Libraries Specs

[00:29:36] Ghost Libraries & Distributing Software as Specs

Ryan Lopopolo: We’re gonna talk about Symphony in a little bit, but like the way we distribute it as a spec, which I think folks are calling Ghost Libraries on Twitter.

This is like a such a cool name. It does mean it becomes much cheaper to share software with the world, right? You define a spec, how you could build your own specifying as much as is required for a coding agent to reassemble it [00:30:00] locally. The flow here is very cool. Like we have taken. All the scaffolding that has existed in our proprietary repo spun up a new one.

Ask Codex with our repo as a reference. Write the spec. We tell it. Spin up a team ox spawn a disconnected codex to implement the spec. Wait for it to be done. Spawn another codex and another team ox to review the spec com or review the implementation compared to upstream and update the spec so it diverges less.

And then you just loop over and over Ralph style until you get a spec that is with high fidelity able to reproduce the system as it is. It’s fantastic.

Vibhu: And you’re basically, you’re not really adding any of your human bias in there, right? That’s correct. A lot of times people write a spec and be like, okay, I think it should be done this way, and you’ll riff on something.

And it’s no, the agent could have just handled it like you’re still scaffolding in a sense, right? I want it done this way. It can determine its spec better.

swyx: That’s right. That’s right. Part of me it, I’m, I’ve been working a lot on evals recently, and part of me is wondering if [00:31:00] an agent can produce a spec that it cannot solve.

Is it always capable of things that he can imagine or can you imagine things that it is impossible to do?

Ryan Lopopolo: I think with Symphony, we, there’s like this there’s this axis where you have things that are easier, hard, or established or new, right? And I think things that are hard and new is still something that the models need humans.

Yeah. Drive.

swyx: Yeah. Yeah.

Ryan Lopopolo: But I think those other quadrants are largely salt. Given the right scaffold and the right thing that’s gonna drive the agent to completion,

swyx: it’s crazy that it solved,

Ryan Lopopolo: but it means that the humans, the ones with limited time and attention get to work on the hardest stuff, like the problems where it’s pure white space out in front. Or like the deepest refactorings where you don’t know what the proper shape of the interfaces are. And this is where I wanna spend my time. ‘cause it lets me set up for the next level of scale.

swyx: Yeah. Yeah. Amazing. Let’s introduce Symphony.

I think we’ve been mentioning it every now and then. Elixir. Interesting option.

Ryan Lopopolo: Yeah.

swyx: Yeah. I’m not,

Ryan Lopopolo: again, like the [00:32:00] elixir manifestation here is just a derivative. Is it a model

swyx: chosen? Yeah.

Ryan Lopopolo: Yeah. Yeah. And it chose that because the process supervision and the gen servers are super amenable to the type of process orchestration that we’re doing here.

You are essentially spinning up little Damons for every task that is in execution and driving it to completion, which. Means the mall gets a ton of stuff for free by using Elixir and the Beam.

swyx: I had to go do a crash course in Beam and Elixir, and I think most people are not operating at that scale of concurrency where you need that.

But it is a good mental model for Resum ability and all those things. And these are things I care about. But tell me the story, the origin story of Symphony. What do you use it for? Is this, how did it form maybe any abandoned paths that you didn’t take?

[00:32:46] Terminal Free Orchestration

[00:32:46] Symphony: Removing Humans from the Loop

Ryan Lopopolo: At the end of December we were at about three and a half PRS per engineer per day.

This was before five two came out in the beginning of January. Everyone gets back from holiday with five two and no other work [00:33:00] on the repository. We were up in the five to 10 PRS per day per engineer. And I don’t know about y’all, but like it’s very taxing to constantly be switching like that. Like I was pretty tapped out at the end of the day, again, where are the humans spending their time? They’re spending their time context switching between all these active tmox pains to drive the agent forward.

swyx: Yeah. No way. Yeah.

Ryan Lopopolo: So let’s again, build something to remove ourselves from the loop. And this is what frantic sprinted adapt here to find a way to remove the need for the human to sit in front of their terminal.

So a lot of experimentation with Devrel boxes and, automatically spinning up agents, like it seems like a fantastic end state here, where my life is beach. I open live twice a day and say yes no to these things. Yeah. And this is again, a super, super interesting framing for how the work is done.

Because I become more latency and sensitive. I have [00:34:00] way less attachment to the code as it is written. Like I’ve had close to zero investment in the actual authorship experience. So if it’s garbage. I can just throw it away and not care too much about it. In Symphony, there’s this like rework state where once the PR is proposed and it’s escalated to the human for review, it should be a cheap review.

It is either mergeable or it is not. And if it’s not, you move it to rework. The elixir service will completely trash the entire work tree NPR and start it again from scratch. Okay. And this is that opportunity again to say, why was it trash right? What did the agent do that was

swyx: bad. Yeah.

Ryan Lopopolo: Fix that before moving the ticket to

swyx: end

Ryan Lopopolo: of progress again.

swyx: Yeah. Why is this not in codex app? I guess this, you guys are ahead of Codex app,

Ryan Lopopolo: yeah, so the way the team has been working is basically to be as AI pilled as possible and spread ahead. And a lot of the things we have worked on have fallen out [00:35:00] into a lot of the products that we have.

Like we were in deep consultation with the Codex team to. Have the Codex app be a thing that exists, right? To have skills be a thing that Codex is able to use. So we didn’t have to roll our own to put automations into the product. So all of our automatic refactoring agents didn’t have to be these hand rolled control loops.

It has been really fantastic to be, in a way, un anchored to the product development of Frontier and Codex and just very quickly try to figure out what works and then later find the scalable thing that can be deployed widely. It’s been a very fun way to operate. It’s certainly chaotic. I have lost track very often of what the actual state of the code looks like.

‘cause I’m not in the loop. There was. One point where we had wired playwright directly up to the Electron app. With MCPM CCPs, I’m pretty bearish on because the harness forcibly injects all those tokens in the [00:36:00] context, and I don’t really get a say over it. They mess with auto compaction. The agent can forget how to use the tool.

There’s probably only what three calls in playwright that I actually ever want to use. So I pay the cost for a ton of things. Somebody vibed a local Damon that boots playwright and exposes a tiny little shim CLI to drive it. And I had zero idea that this had occurred because to me, I run Codex and it’s able to, it’s oh, it’s better.

Yeah. Like no knowledge of this at all. Uhhuh.

[00:36:30] Multi Human Chaos

Ryan Lopopolo: So we have had like in human space to spend a lot of time doing synchronous knowledge sharing. We have a daily standup that’s 45 minutes long because we almost have to. Fan out the understanding of the current state.

swyx: Yeah, I was gonna say this is good for a single human multi-agent, but multi human, multi-agent is a whole like po like explosion of stuff.

Ryan Lopopolo: Yeah. And that this is fundamentally why we have such a rigid, like 10,000 [00:37:00] engineer level architecture in the app because we have to find ways to carve up the space so people are not trampling on each other.

swyx: Sorry, I don’t get the 10,000 thing. Did I miss that?

Ryan Lopopolo: The structure of the repository is like 500 NPM packages.

It’s like architecture to the excess for what you would consider, I think normal for a seven person team. But if every person is actually like 10 to 50. Then the like numbers on being super, super deep into decomposition and sharding and like proper interface boundaries make a lot more sense.

swyx: Yeah. To me, that’s why I talked about Microfund ends and I, an anex is from that world, but Cool. It is just coming back to, to, to this I dunno if you have other, thoughts on. Orchestrating so much work coin going through this. Is this enough? Is this like any aha moments?

Vibhu: It’ll be interesting to see like where, okay, so right now you pick linear as your issue tracker, right?

swyx: Or it’s like a is it actually linear? This is actually linear.

[00:37:55] Linear vs Slack Workflow

Vibhu: Oh, that’s linear. It’s linear.

swyx: Oh I never looked at

Vibhu: video. The demo video I had to download to [00:38:00] run.

swyx: So I, because I’m a Slack maxie, but Yeah, linear. Linear is also really good. Yes,

Ryan Lopopolo: we do make a good use of Slack. We we fire off codex to do all these lotion, elasticity, fix ups, the things that like sync that knowledge into the repository.

It’s super cheap. Yeah.

swyx: Yeah.

Ryan Lopopolo: Just do it in Codex.

swyx: My biggest plug is OpenAI needs to build Slack. You need to own Slack. Build yours. Turn this into Slack.

Ryan Lopopolo: I did read about it. You

swyx: did?

Ryan Lopopolo: Yeah.

[00:38:25] Collaboration Tools for Agents

Ryan Lopopolo: I would say that if we think that we want these agents to do economically valuable work, which is like this is the mission, right?

We want AI to be deployed widely, to do economically valuable work, then we need to find ways for them to naturally collaborate with humans, which means collaboration tooling, I think, is an interesting space to explore.

swyx: Yeah, totally. Yeah. GitHub, slack, linear.

Vibhu: Yeah, that was my thing. Okay, where do we see right now Codex has started Codex Model, then CLI, now there’s an app, app can let me shoot off multiple Codex is in parallel, but there’s no great team collaboration for Codex.

And it [00:39:00] seems like your team had some say into what comes out, right? So you talked to ‘em, codex kind of was a thing. From there, if you guys are on the bound, what stuff that like, you might not focus on, but what do you expect other people to be building, right? So people that are like five x 50 Xing.

Should you build stuff that’s like very niche for your workflow, for your team? Should it be more general so other people can adopt? Is there a niche there? ‘Cause part of it is just okay, is everything just internal tooling? Do we have everything our own way? Like the way our team operates has our own ways that we like to communicate or is there a broader way to do it?

Is it something like a issue tracker? Just thoughts if you wanna riff on that.

[00:39:35] Standardizing Skills and Code

Ryan Lopopolo: I think TBD we have not figured this out in a general way. I do think that there is leverage to be had in making the code and the processes as much the same as possible. If you think that code is context, code is prompts, it’s better from the agent behavior perspective to be able to look in a package in directory X, Y, Z, and it not to have to page so [00:40:00] deeply into directory if you C, because they have the same structure, use the same language, they have the same patterns internally.

And that same like leverage comes from aligning on a single set of skills that you’re pouring every engineer’s taste into to make sure that the agent is effective. So like in our code base, we have, I think, six skills. That’s it. And if some part of the software development loop is not being covered, our first attempt is to encode it in one of the existing setup skills, which means that we can change the agent behavior.

Yeah. More cheaply than changing the human driver behavior.

swyx: Yeah.

[00:40:39] Self Improvement via Logs

swyx: Have you ever, have you experimented with agents changing their own behavior?

Ryan Lopopolo: We do.

swyx: Yeah. Or parent agent changing a subagents, behavior or something like that.

Ryan Lopopolo: We have some bits for skill distillation. So for example, there’s one neat thing you can do with Codex, which is just point it at its own session logs to ask it to tell you how you can use [00:41:00] the tool pedal better.

swyx: It’s like introspection

Ryan Lopopolo: or ask it to do things. I use

Vibhu: this session better. What skills should I

swyx: high? I like the modification of, you can do, just do things to you can just ask agent to do things.

Ryan Lopopolo: Yeah. You can just codex things. This is like a, this is like a silly emoji that we have, right? You can just codex things, you can just prompt things.

It’s really glorious future we live in, but okay, you can do that one-on-one. But we’re actually slurping these up for the entire team into blob storage and. Running agent loops over them every day to figure out where as a team can we do better and how do we reflect that back into the repositories?

Yes, though everybody benefits from everybody else’s behavior for free. Same for like PR comments, right? These are all feedback. That means the code as written, deviated from what was good, a PR comment, a failed build. These are all signals that mean at some point the agent was missing context. We gotta figure out how to

swyx: Yeah.

Ryan Lopopolo: Slurp it up and put it back in the reboot.

swyx: By the way, I do this exactly right. I used to, when I use cloud code for [00:42:00] knowledge work, cloud cowork is like a nice product, right? Yes. In I think you would agree. I always have it tell me what do I do better next time? And that’s the meta programming reflection thing.

So I almost think like you have six reflection extraction levels in symphony and almost like the zero of layer. So the six levels are PO policy, configuration, coordination, execution, integration, observability. We’ve talked about a couple of these, but the zero layer is like the, okay, are we working well?

Can we improve how we work? Yes. Can I modify my own workflow without MD or something? I don’t know.

Ryan Lopopolo: Yeah, of course. Yeah, of course you can. Like this thing is also able to cut its own tickets ‘cause we give it full access.

Yeah. Make it a ticket to have it cut. Tickets you can.

Put in the ticket that you expect it to file as on follow up work,

swyx: like Yeah. Self-modifying. Yeah.

Ryan Lopopolo: Yeah.

[00:42:44] Tool Access and CLI First

Ryan Lopopolo: Put, don’t put the agent in a box. Give the agent full accessibility over it. Domain.

swyx: I had a mental reaction when you said don’t put the agent in a box. So I think you should put it in a box. Like it’s just that you’re giving the box everything it needs.

Ryan Lopopolo: Yeah. Context and tools.

swyx: But we’re like, as developers, we’re used to calling [00:43:00] out to different systems, but here you use the open source things like the Prometheus, whatever, and you run it locally so that you can have the full loop. I assume.

Ryan Lopopolo: Yep.

Vibhu: I think like

Ryan Lopopolo: another, you wanna minimize cloud, cloud dependencies.

Vibhu: You also want to make sure that you think about what the agent has access to. What does it see? Does it go back into the loop, like from the most basic sense of you let it see its own like calls, traces it can determine where it went wrong. But are you feeding that back in? So you know, just the most basic level of you wanna see exactly what’s input output, like does the agent have access to.

What is being outputted, right? It can self-improve a lot of these things. It’s all

Ryan Lopopolo: text, right? My job is to figure out ways to funnel text from one agent to the other.

swyx: It’s so strange like way back at the start of this whole AI wave Andre was like, English is the hottest day programming language.

It’s here, it’s just Yeah. The feature as well.

Vibhu: A lot of, okay. Like a lot of software, a lot of stuff. There’s a gui, it’s made for the human. We’re seeing the evolution of CLI for everything, right? All tools have CLIs. Your agents can use [00:44:00] them well, do we get good vision? Do we get good little sandboxes?

Like right now? It’s a really effective way, right? Models love to use tools. They love the best. They love to read through text. So slap a CLI let it go loose. That works for everything.

Ryan Lopopolo: It does. Yeah. Yeah.

[00:44:14] UI Perception and Rasterizing

Ryan Lopopolo: We’ve also been adapting nont, textual things to that shape in order to improve model behavior in some ways, right?

We want the agent to be able to see the UI agents do not perceive visually in the same way that we do. They don’t see a red box, they see red box button, right? They see these things in latent space. So if we want, Hey, yeah, I do. We have

swyx: a ding if that goes off every time. Alien space

Ryan Lopopolo: ding.

Anyway if we wanna actually make it see the layout, it’s almost easier to rasterize that image to ask EOR and feed it in to the agent. Ha. And there’s no reason you can’t do both, right? To like further refine how the model perceives the object it’s [00:45:00] manipulating.

swyx: Cool. Could we, you wanna talk about a couple more of these layers that might bear more introspection or that you have personal passion for?

[00:45:07] Coordination Layer with Elixir

Ryan Lopopolo: I will say that the coordination layer here was a really tricky piece to get right.

swyx: Let’s do it. Yep. I’m all about that. And this is Temporal core.

Ryan Lopopolo: This is where when we turn the spec into Elixir, where like the model takes a shortcut, right? Like it’s oh, I have all these primitives that I can make use of in this lovely runtime that has native process supervision.

Which is I think, a neat way to have taken the spec and made it more choices achievable by making choices that naturally map

swyx: Yeah.

Ryan Lopopolo: To the domain, right? In the same way that like you would prefer to have a TypeScript model repo if you are doing full stack web development, right? Because the ability to share types across the front end and backend reduces a lot of complexity.

And because

swyx: that’s what graph kill used to be.

Ryan Lopopolo: That’s right. And

swyx: I don’t know if it’s still alive, but

Ryan Lopopolo: [00:46:00] no humans in the loop here. So like my own personal ability to write or not write elixir. Doesn’t really have to bias us away from using the right tool for the job. It is just wild.

swyx: Love it. I love it.

Yeah. I wonder if any languages struggle more than others because of this? I feel like everyone has their own abstractions. That would make sense. But maybe it might be slower, it might be more faulty where like you’d have to just kick the server every now and then. I, I don’t know. I think observability layer is really well understood.

Integration layer, CP is dead. I think all these just like a really interesting hierarchy to travel up and down. It’s common language for people working on the system to understand

Ryan Lopopolo: The policy stuff is really cool, right? Yeah. You don’t really have to build a bunch of code to make sure the system wait for the, to pass

swyx: it’s institutional knowledge.

Ryan Lopopolo: Yeah. You just give it the G-H-C-L-I with some text that say CI has to pass. It makes the maintenance of these systems a lot easier.

[00:46:57] Agent Friendly CLI Output

swyx: Do you think that CLI maintainers need to be [00:47:00] do anything special for agents or just as is? It’s good because like I don’t think when people made the G GitHub, CLI, they anticipated this happening.

Ryan Lopopolo: That’s correct. The GH CLI is fantastic. It’s great super industry.

swyx: Everyone go try GH repo create GH pull and then pull request number, right? GH HPR, like 1 53, whatever. And then it like pulls

Ryan Lopopolo: basically my only interaction with the GitHub web UI at this point is GH PR view dash web.

Exactly. Glance

swyx: at the diff

Ryan Lopopolo: and be like Sure thing. Send it. Yeah. But the CLI are nice ‘cause they’re super token efficient and they can be made more token efficient really easily. Like I’m sure you all have seen like I go to build Kite or Jenkins and I could just get this massive wall of build output.

And in order to unblock the humans, your developer productivity team is almost certainly gonna write some code that parses the actual exception out of the build logs and sticks it in a sticky note at the top of the page. And you basically [00:48:00] want CLI to be structured in a similar way, right? You’re gonna want to patch dash silent to prettier because the agent doesn’t care that every file was already formatted.

Just wants to know it’s either formatted or not. So it can then go run a right command. Similarly, like in our PNPM distributed script runner, when we had one, when you do dash recursive, like it produces a absolute mountain of text. But all of that is for passing. Test suites. So we ended up wrapping all of this in another script

swyx: to suppress the,

Ryan Lopopolo: which you can vibe the channel only output the failing parts of the tests.

swyx: You make a pipe errors versus the standard, standard out. I don’t know. Okay. Whatever. Too much thinking have to do that. The CII used to maintain SCLI for my company and yeah, this is like core, very core to my heart. But you’re vibing my job.

Ryan Lopopolo: That’s right.

swyx: Cool. Any other things?

This is a long spec. [00:49:00] I appreciate that. It’s got a lot of strong opinions in here. Any other things that we should highlight? I think obviously you can spend the whole day going through some of these, but I do think that some of these have a lot of care or some of this you might wanna tell people, Hey, take this, but, make it your own.

[00:49:15] Blueprint Spec and Guardrails

Ryan Lopopolo: Fundamentally, software is made more flexible when it’s able to adapt to the environment in which it is deployed, which means that things like linear or GitHub even are specified within the spec, but not required pieces of it. There’s like a more platonic ideal of the thing that you could swap in like Jira or Bitbucket, for example.

But being able to tightly specify things like the ID formats or how the Ralph Loop works for the individual agents. Basically means you can get up and running with a fully specified system quickly that you then evolve later on. I think we never intended for this to be a static spec that you can [00:50:00] never change.

It’s more like a blueprint to get something worth a starting point up and running.

swyx: Yeah.

Ryan Lopopolo: For you then to vibe later to your heart’s content,

swyx: you have like code and scripts in here where it’s oh, I think this is a really good prompt. It’s just a very long prompt.

Ryan Lopopolo: Fundamentally, the agents are good at following instructions, so give them instructions.

And it will, improve the reliability of the result. We, much like the way we use Symphony, we don’t want folks to have to monitor the agent as it is vibing the system into existence. So being very opinionated

Very strict around what these success criteria are means that our deployment success rate goes up. Yeah. It means we don’t have to get tickets on this thing.

Vibhu: Think it all goes back to that like code to disposable, right? Like early on when you had CLI or you’d kick off a Codex run, it would take two hours. You would wanna monitor okay, I’m in the workflow of just using one.

I don’t want it to go down the wrong path. I’ll cut it off and, just shoot off four, like that was my favorite thing of the Codex app, right? Yeah. Just Forex it like, [00:51:00] it’s okay. One of them will probably be right, one of them might be better. Stop overthinking it. Like my first example was probably like deep research.

When you put out deep research and I’d ask it something like, I asked it something about LLM, it thought it was legal something and spent an hour, came back with a report completely off the rails. And I was like, okay, I gotta monitor this thing a bit. No don’t monitor it. Just you want to build it so it’s that it, it goes the right way.

And you don’t wanna, you don’t wanna sit there and babysit, right? You don’t want to babysit your agents

Ryan Lopopolo: with that deep research query that you made. Looking at the bad result, you probably figured out you needed to tweak your prompt Yeah. A bit, right? That’s that guardrail that you fed back into the code base for the task, your prompt to further align the agent’s execution.

Same sort of concept supply there too.

swyx: When you talk, how are the customers feeling

Ryan Lopopolo: for Symphony? I think we have none, right? This is a thing we have put out into the

swyx: world. Symphony’s internal, right? As long as you are happy, you are the customer. That’s right. Just, what’s the external view?

[00:51:53] Trust Building with PR Videos

Ryan Lopopolo: I’d say folks are very excited about this way of distributing software and ideas in [00:52:00] cheap ways. For us as users, it has again, pushed the productivity five x, which means I think there’s something here that’s like a durable pattern around removing the human from the loop and figuring out ways to trust the output.

The video that is shared here

swyx: Yeah.

Ryan Lopopolo: Is the same sort of video we would expect the coding agent to attach to the pr.

swyx: Yeah.

Ryan Lopopolo: That is created. Yeah. That’s part of building trust in this system and that’s, to me, like fundamentally what has been cool about building this is it more closely pushes that persona of the agent working with you to be like a teammate.

I don’t shoulder surf you like for the tickets that you work on during the week. I would never think that I would want to do that.

swyx: Yeah.

Ryan Lopopolo: I wouldn’t want a screen recording of your entire session in Cursor or Claude code. I would expect you to do what you think you need to do to convince me that the code is good and [00:53:00] mergeable

swyx: Yeah.

Ryan Lopopolo: And compress that full trajectory in a way that is legible to me. The reviewer.

swyx: Yeah.

Ryan Lopopolo: It’s Stu. And you can just do that because Codex will absolutely sling some f you can just around. It’s great.

swyx: Oh, F FM P is the og like God, CLI.

Ryan Lopopolo: Yeah.

swyx: Swiss Army Chainsaw. I used to say. There’s a SaaS, micro SaaS that’s called it in every flag in FFM Peg.

Ryan Lopopolo: Oh, for sure.

swyx: You know what I mean? For sure. Just host it as a service, put a UI on it. People who don’t know FM Peg will pay for it.

Ryan Lopopolo: When we were first experimenting with this, it was a wild feeling to be at the computer with just like windows just popping up all over the place and getting captured and files appearing on my desktop, like very much felt like the future to have a thing controlling my computer for like actual productive use.

Like I’m just there

swyx: keeping it. Like awake, jiggling the mouse every once in a while. That’s what some office workers do. So they buy a mouse jiggler. That’s right.

[00:53:59] Spark vs Reasoning Models

Vibhu: One thing I [00:54:00] wanted to ask, so okay, as stuff is so CO is disposable is saying shoot off a budget of agents. One question is okay, are you always like a extra high thinking guy?

And where do you see Spark? So 5.3 Spark, there’s a lot of me wanting to make quick changes. I’m not gonna open up a id, I’m not gonna do anything. But I will say, okay, fix this little thing, change a line, change a color. Spark is great for that, but am I still a bottleneck? Like, why don’t I just let that go back?

I’m like, just riff on that. Is there,

Ryan Lopopolo: spark is such a different model compared to the. The extra high level reasoning that you get in these, five Yeah. To clear for people.

swyx: It is a different model, different architecture, different, like it doesn’t support

Ryan Lopopolo: it, it just, it’s incredibly fast smaller model.

I have not quite figured out how to use it yet. To be honest, I use faster. I was adapting it to the same sorts of tasks I would use X high reasoning for. Yeah. I, and it would blow through three compactions before writing a line of code.

Vibhu: And that’s another big thing with 5.4 right.

Million co context.

Ryan Lopopolo: Yes, it’s

Vibhu: fantastic. Which is huge [00:55:00] ingenix, right? Like you can just run for longer before you have to compact. The more tokens you can spend on a task before compacting, like the better you’ll do.

Ryan Lopopolo: That’s right. That’s right. I’m not sure how to deploy spark. I think your intuition is right, that it’s very great for spiking out prototypes, exploring ideas quickly, doing those documentation updates.

It is fantastic for us in taking that feedback and transforming it into a lint. Where we already have good infrastructure for ES links in the code base these sorts of things it’s great at and it allows us to unblock quickly doing those like anti-fragile healing tasks in the code base.

swyx: Yeah, that makes sense.

[00:55:38] What Models Can’t Do Yet

swyx: So you are push, you guys are pushing models to the freaking limit.

[00:55:41] Current Model Limitations

swyx: What can current models not do well yet?

Ryan Lopopolo: They’re definitely not there on being able to go from new product idea to prototype single

swyx: one shot.

Ryan Lopopolo: This is where I find I spend a lot of time steering is translating end state of a mock for a net new [00:56:00] thing, right?

Think no existing screens into product that is playable with. Similarly, while this has gotten better with each model release, like the gnarliest refactorings are the ones that I spend my most time with, right? The ones where I’m interrupting the most, the ones where I am. Now double clicking to build tooling to help decompose monoliths and things like that.

This is a thing I only expect to get better, right? Over the course of a month, we went from the low complexity tasks to like low complexity and big tasks in both these directions. So this is what it means to not bet against the model, right? You should expect that it is going to push itself out into these higher and higher complexity spaces.

Yeah. So the things we do are robust to that. It just basically means I’ll be able to spend my time elsewhere and figure out what the next bottleneck is.

Vibhu: I do think it’s also a bit of a different type of task, right? Codex is really good at codebase understanding, working with code bases. But companies like Lovable bolt, repli, they solve a very different [00:57:00] problem.

Scaffold of zero to one, right? Idea of a product. And it’s there, there are people working on that and models are also pushing like step function changes there. It’s just different than the software engineering agents today, right?

Ryan Lopopolo: Like I said, the model is isomorphic to myself.

The only thing that’s different is figuring out how to get what’s in here into context for the model and for these white space sort of projects. I, myself, I’m just not good at it. Which means that often over the agent trajectory, I realize the bits that we’re missing, which is why I find I need to have this synchronous interaction.

And I expect with the right harness, with the right scaffold, that’s able to tease that outta me or refine the possible space, right? To be super opinionated around the frameworks that are deployed or to put a template in place, right? These are ways to give the model. All those non-functional requirements, that extra context to acre on and avoid that wide dispersion of possible outcomes.

swyx: Thank [00:58:00] you for that.

[00:58:00] Frontier Enterprise Platform

swyx: I wanted to talk a little bit about Frontier.

Ryan Lopopolo: Yeah, sure.

swyx: Overall you guys announced it maybe like a month ago. And there’s a few charts in here and it’s basic like your enterprise offering is what I view it. Is there one product or is there many,

Ryan Lopopolo: I can’t speak to the full product roadmap here, but what I can say is that Frontier is the platform by which we want to do AI transformation of every enterprise and from big to small.

And the way we want to do that is by making it easy to deploy highly observable, safe, controlled, identifiable agents into the workplace. We want it to work with your company native. I am stack. We want it to plug into the security tooling that you have. Oh, we want it to be able to plug into the workspace tools that you used,

swyx: so you’re just gonna be stripping specs, right?

Ryan Lopopolo: We expect that there will be some harness things there. Agents, SDK is a core [00:59:00] part of this to enable both startup builders as well as enterprise builders to have a works by default harness that is able to use all the best features of our models from the Shell tool down to the Codex Harness with file attachments and containers and all these other things that we know go into building highly reliable, complex agents.

We wanna make that great and we wanna make it easy to compose these things together in ways that are safe, for example, right? Like the G-P-T-O-S-S safeguard model. For example. One thing that’s really cool about it is it ships. The ability to interface with a safety spec. Safety specs are things that are bespoke to enterprises.

We owe it to these folks to figure out ways for them to instrument the agents in their enterprise to avoid exfiltration in the ways they specifically care about, to know about their internal company, code names, these sorts of things. So providing the right hooks to make the [01:00:00] platform customizable, but also, mostly working by default for folks is the space we are trying to explore here.

swyx: Yeah. And this is the snowflakes of the world just need this, right? Yes. Your Brexit of the world stripes. Yeah, it makes sense.

[01:00:11] Dashboards and Data Agents

swyx: I was gonna go back to your, I think the demo videos that you guys had was pretty illustrative. It’s like also to me an example of very large scale agent management.

Yes. Like you give people a control dashboard that if you play, if you like, play any one of these like multiple agent things, you can di dig down to the individual instant and see what’s going on.

Ryan Lopopolo: Yes, of course.

swyx: But who’s the user Is it let’s it like the CEO, the CTO, ccio, something like that.

Ryan Lopopolo: At least with my personal opinion here, the buyer that we’re trying to build product for here is one and employees who are making productive use of these agents, right?

That’s gonna be whatever surfaces they appear in the connectors they have access to, things like that. Something like this dashboard is for it. Your GRC and governments folks, your AI innovation office, your security [01:01:00] team, right? The stakeholders in your company that are responsible for successfully deploying into.

The spaces where your employees work, as well as doing so in a safe way that is consistent with all the regulatory requirements that you have and customer attestations and things like that. So it is a iceberg beneath the actual end. It’s,

swyx: yeah you jump every, I guess layer in the UI is like going down the layer of extraction in terms of the agent, right?

Yep. Yeah. Yeah. I think it’s good.

Ryan Lopopolo: Yeah. The ability to dive deep into the individual agent trajectory level is gonna be super powerful.

Not only for from like a security perspective, but also from like someone who is accountable for developing skills. One thing that was interesting that we also blogged about shipping was an internal data agent, which uses a lot of the frontier technology in order to make our data ontology accessible to the agent and things like that to understand.

What’s actually in the data [01:02:00] warehouse?

swyx: Yeah. Seman layer Yes. Type things. Yes. I was briefly part of the, that, that world is it salt? I don’t know. It’s actually really hard for humans to agree on what revenue is. Yes.

Ryan Lopopolo: Yes.

swyx: What is an active user?

Ryan Lopopolo: There’s what, five data scientists in the company that have defined this Golden.

swyx: They, yeah. And no. And there’s also internal politics. Yes. As to attribution of I’m marketing, I’m responsible for this much, and sales is responsible for this much, and they all add up to more than a hundred. And I’m like you guys have different definitions.

Vibhu: Yeah. And if you’re a startup, everything is a RR,

swyx: So I think that’s cool.

Oh, you guys blog about this. Okay. I didn’t see this. Yeah. Is this the same thing? I don’t know. This is what you’re referring to? Yes. Okay. We’ll send people to read this. This is our data.

Vibhu: Him this one.

swyx: Yeah. I don’t know if you’re you have any highlights? I

Vibhu: No. In general from the playlist.

Yeah. A lot of good things to read.

swyx: Yeah. Yeah. Lot, lots of homework for people. No, but like data as the feedback layer, you need to solve this first in order to have the products feedback loop closed. That’s right. So for the agents to understand and this is not something that humans have not solved.

This like, and

Ryan Lopopolo: this is [01:03:00] how you build artists that do more than coding, right? Yeah.

swyx: Yeah.

Ryan Lopopolo: To actually understand how you operate the business.

swyx: Yeah.

Ryan Lopopolo: You have to understand what revenue is, what your customer segments are. Yeah. What your product lines are.

[01:03:13] Company Context and Memes

Ryan Lopopolo: Like one thing that’s in looping back to the code base that we described here for harnessing, one thing that’s in core beliefs.md is who’s on the team, what product we’re building, who our end customers are.

Who our pilot customers are, what the full vision of what we want to achieve over the next 12 months is these are all bits of context that inform how we would go about building the software. Oh my God. So we have to give it to the agent too.

Vibhu: I’m guessing that stuff is like pretty dynamic and it changes over time too, right?

Like part of it was, it’s not just a big spec. You have it as one of the things and it will iterate.

Ryan Lopopolo: One, one thing that I think is gonna break your mind even more is we have skills for how to properly generate deep fried memes and have Ji culture [01:04:00] and Slack. Because with the Slack Chachi PT app that you’re able to use in Codex, like I can get the agent to s**t post on my behalf.

Just, it’s part of humor.

swyx: Theme humor. Humor is part of EGI. Is it funny? It is pretty good, yeah. Okay. Yeah,

Ryan Lopopolo: it’s pretty good at making

swyx: Deep, it’s a lot of I think humor is like a really hard intelligence test, right? It’s like you have to get a lot of context into like very few words.

This is why make references

Ryan Lopopolo: is why five four is such a big uplift for our it’s the me. Yeah, for sure. Yeah. Yeah.

swyx: It’s very cool.

Vibhu: So five, four can two post. So that’s what we take over here.

Ryan Lopopolo: Yeah. Maybe maybe when y’all are done here today, ask Codex to go over your coding agent sessions and to roast you.

swyx: Love it. I’ll give it a shot. Give a shot. Coming back to the final point I wanted to make is, yeah I think that there, there are multiple other, like you guys are working on this, but this is a pattern that every other company out there should adopt. Yes. Regardless of whether or not they work with you.

To me, this is I saw this, I was like, f**k, [01:05:00] every company needs this. This

is

swyx: multiple billions.

Ryan Lopopolo: This is what it takes to get

swyx: Yeah.

Ryan Lopopolo: People to Yes. Yeah. Actually realize the benefits. Yes. And distribute.

swyx: And it’s, it, I think it sounds boring to people like, oh, it’s for safeguards and whatever, but I think you to handle agents at scale like you are envisioning here I don’t know if it’s like a real screenshot, like a demo, but this is what you need.

This is, or my original sort of view of what Temporal was supposed to be that you, you built this dashboard and you basically have every long running process in the company Yes. In one dashboard and that’s it. That’s right.

Vibhu: Yeah. I think it’s pretty customized towards every enterprise, right?

Like you care about different things.

swyx: There’s a lot of customization, but there’ll be multiple unicorns just doing this as a service. I don’t know. I’m like very frontier field, if you can tell. Amazing. But it, it only clicked ‘cause obviously this came out first, then Harness eng, then symphony and only clicked for me that like, this is actually the thing you shipped to do that.

Ryan Lopopolo: Yeah. Yeah. There’s a set of building blocks here that we assembled into these agents [01:06:00] and the building blocks themselves are part of the product, right? Yeah. The ability to steer revoke authorization if a model becomes misaligned, like all of this is accessible through Frontier. And there’s gonna be a bunch of stakeholders in the company that have the things they need to see in the platform Yeah.

To get to. Yes. So we’ll build all of those in the frontier so that we can actually do the widespread the planet. Yeah. That’s the fun part.

swyx: Yeah. I’m also calling back to there’s this like levels of EGI I don’t know if Opening Eye is still talking about this, but they used to talk about five levels of EGI and one of it was like, oh, it’s like an intern coding software patient.

At some point it was AI organization and this is it. That’s right. This is level four or five. I can’t remember which, which level, but it’s somewhere along that path. Was this.

Ryan Lopopolo: You know how I mentioned that my team is having fun sprinting ahead here. And we do this thing where we’re collecting all the agent trajectories from Codex to slurp them up and distill them.

This is what it means to build our team [01:07:00] level knowledge base, happen to reflect it back into the code base. But it doesn’t have to be that way. And it doesn’t have to be bound to just codex. I want Chacha BT to also learn our meaning culture and also the product we are building and how so that when I go ask it, it also has the full context of the way I do my work and I’m super excited for Frontier to enable this.

swyx: Yeah. Amazing.

[01:07:21] Harness vs Training Tension

swyx: What are the model people say when they see you do this? Like you have a lot of feedback, obviously you have a lot of usage, you have a lot of trajectories and don’t, I don’t imagine a lot of it’s useful to them, but some of it is,

Vibhu: you have this too, you deploy a billion tokens of intelligence a day and this was, this was at the beginning of 2096.

You’re Yeah. Cooking.

Ryan Lopopolo: Yeah, there’s this fundamental tension, which I think you have talked about between whether or not we invest deeper into the harness or we invest deeper into the training process to get the model to do more of this by default. Yeah, and I think success for the way we are [01:08:00] operating here means the model gets better taste because we can point the way there and none of the things we have built actively degrade Asian performance.

‘cause really all they’re doing is running tests and like running tests is a good part of what it means to write reliable software. If we were building an entire separate rust scaffold around Codex to restrict its output, that I think would be like additional harness that would be prone to being scrapped.

But yeah. Yeah. If instead we can build all the guardrails in a way that’s just native to the output that Codex is already producing, which is code, I think. No friction with how the model continues to advance, but also like just good engineering and that’s the whole point.

swyx: Yeah. So I’ve had similar discussions with research scientists where the RL equivalent is on policy versus off policy.

Yeah. And you’re basically saying that you should build an on policy harness, which is already within distribution and you [01:09:00] modify from there. But if you build it off policy, it’s not that useful.

Ryan Lopopolo: That’s right.

swyx: Super cool. Any, anybody thoughts, any things that we haven’t covered that we should get it, get out there?

[01:09:08] Closing Thoughts & OpenAI Hiring

Ryan Lopopolo: Just I’ve been super excited to benefit from all the cooking that the Codex team has been doing. Yes. They absolutely ship relentlessly. This is one of our core engineering values, ship relentlessly, and they, the team there embodies it. To extreme degree, yeah, I have five three and then Spark and five four come out within what feels like a month is just a phenomenally fast.

swyx: It’s exactly a month ago it’s five three and yesterday was five four. Yeah. I mean it’s, do we have every month now is five five next? Exactly.

Ryan Lopopolo: I can’t say that the poll markets would be very upset.

swyx: I think it’s interesting that it’s also correlated with the growth. They announced that it’s 2 million users, but like almost don’t care about Codex anymore.

This is it, this is the gay man. It’s like coding cool, soft like knowledge work.

Ryan Lopopolo: That’s right. That’s right. This is the thing to chase after. Yeah. And this is one of things that my team is excited to support,

swyx: get the whole like [01:10:00] self-hosted harness thing working, which you have done and like the rest of us are trying to figure out how to catch up, but then do things.

You That’s right. With you

Vibhu: do things.

swyx: That’s right. You can just do things. That’s the line for the episode.

Vibhu: That’s it. Any other call to actions. You’re based in Seattle, your team, I’m guessing. New Bellevue office.

Ryan Lopopolo: New Bellevue office. We just had the grand opening yesterday as of the recording date which was fantastic.

Beautiful buildings. Super excitedly part of the Bellevue Community building the future in Washington. And I would say that there is lots of work to be done in order to successfully serve enterprise customers here in Frontier. We are certainly hiring and if you haven’t tried the Codex app yet, please give it a download.

We just passed 2 million weekly active users growing at a phenomenally fast rate, 25% week over week. Come join us.

swyx: Yes. And I think that’s an interesting no. My, my final observation opening is a very San Francisco centric company. I know people who have been. [01:11:00] Who turned down the job or didn’t get the job ‘cause they didn’t want to move to sf and now they just don’t have a choice.

You have to open the London, you have to open the Seattle. And I wonder if that’s gonna be a shift in the culture, obviously you can’t say, but

Ryan Lopopolo: I was one of the first engineering hires out of our Seattle office, so Yeah.

swyx: See I was very natural.

Ryan Lopopolo: Its success has been part of what I have been building toward and it is, it has grown quite well, right?

Yeah. We have durable products in the lines of business that are built outta there a ton of zero to one work happening as well, which is the core essence of the way we do applied AI work at the company to sprint after it new to figure out where we can actually successfully deploy the model.

Yeah. Yes. A hundred percent. We also have a New York office too that has a ton of engineering presence.

swyx: Yeah. Exact. Exactly. That’s these are my road roadmaps for a e wherever people hiring engineers, I will go. That’s right. Ra it’s

Vibhu: a cool office to New York is a old REI building, I believe the REI office.

swyx: It’s just No, you’ll never be as big. New York is you can’t get [01:12:00] the size of office that they need.

Ryan Lopopolo: The New York office, Seattle user has a very office Mad Men vibe. It’s beautiful. The Bellevue one is very green, gold fixtures, very Pacific Northwest is very cool place to the vibe.

Be local

Vibhu: little, yeah. A lot of people are like there for people like New York. They wanna be in New York, right?

Ryan Lopopolo: Yeah. Yeah. We have a fantastic workplace team that has been building out these offices. It really is a privilege to work here. Yeah. Excellent. Okay. Thank you for your time. You’ve been very

swyx: generous and you’re, you’ve been cooking, so I’m gonna let you get back to cooking.

It’s been amazing to be with you folks. Happy Friday. Happy Friday.

This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit www.latent.space/subscribe

The podcast and artwork embedded on this page are from Latent.Space, which is the property of its owner and not affiliated with or endorsed by Listen Notes, Inc.

Podcast: Level-up Engineering (LS 33 · TOP 5% what is this?)
Episode: Building a new engineering team by turning another one around - Tips from Tinder
Pub date: 2024-10-16

Get Podcast Transcript →
powered by Listen411 - fast audio-to-text and summarization

Transforming teams doesn’t go without its challenges.

Let’s look at Tinder’s example. 

In this episode, Chris O'Brien, Director of Engineering at Tinder, shares his insights on building and leading engineering teams, particularly focusing on turning around existing teams. He discusses transforming teams, transitioning into a leadership role, Tinder’s culture and hiring process and a lot more.

Sign up to the Level-up Engineering newsletter!

In this interview we're covering:

• Building a new team by turning another one around
• Transitioning into a leadership role
• Tinder’s culture
• Keeping business, customer and team needs aligned
• Tinder’s hiring process

Excerpt from the interview:

“Change isn't easy for anyone, especially in the workplace where stability and predictability matter. Switching teams suddenly can be unsettling, and it takes time for people to adapt and build trust with their new colleagues. That's why I've always believed in prioritizing relationship-building. It's something my mentor taught me early on, and it's proven to be invaluable. When there's already a foundation of trust and camaraderie, transitions become smoother, and teams become stronger.”

The podcast and artwork embedded on this page are from Apex Lab, which is the property of its owner and not affiliated with or endorsed by Listen Notes, Inc.