TwoSetAI: Decoding AI Builders

From venture capital investor to co-founder of a company that makes physical materials: Jorge Colindres was investing in software and machine learning companies when he grew frustrated with what they were building. He read, by his count, north of 200 research papers to find out whether the technology was the problem. At the time of this conversation he is a co-founder of Radical AI, which uses machine learning to design new materials and a robotic lab to make and test them. By his account the company has gone from two people in a living room to about 37, and has raised about $65 million.

In this episode, Jorge joins TwoSetAI to argue that science needs a new operating model. Radical AI, he says, is not about automating science or predicting materials a little better. It is about joining design, experiment and analysis in one continuous feedback loop that compounds, so that a timeline of 10 to 20 years can be compressed into one to two. He starts with materials because, as he puts it, so many hard problems are bottlenecked by the fact that materials take too long to develop.

You can learn more about Radical AI here.

We go deep on how the system works. An agent draws on hundreds of thousands of scientific publications, quantum mechanical methods, simulations it writes code to run and the lab's own earlier results, and crafts a hypothesis for an experiment. A robotic lab with about nine tools executes it, automated models analyze what came out, human scientists add context and build benchmarks, and all of it feeds the next design. Jorge gives high entropy alloys as the example: by his account academia has validated about 3,500 compositions in about 40 years, and Radical validated a few hundred in the last couple of months. He also explains why the company wants to be fully verticalized, from discovery to manufacturing, and why the business is simple to state: "We sell materials."

We also get into the harder parts. Asked whether he ever feels vulnerable as a founder, he says "every day," and talks about who he leans on. He explains where he thinks the moat is, why the failed iterations that never make it into papers are among the most important things a scientist can capture, where hallucinations show up, and why a long-term vision with no intermediate progress is a fantasy. On the team: most of it does not come from a materials background, everyone works in person five days a week, and a third party torch test at 3,000 degrees, which he calls an early test, gave them a sign that the approach works. He closes with three books, for himself at 20, today, and in 20 years.

We discuss:

Radical AI in the simplest terms: machine learning imagines materials, a robotic lab creates and tests them

Why he calls it a new version of science, with AI as a tool in the process

Compressing 10 to 20 years of materials development into a compounding cycle

Fusion, lithium ion batteries and airliners: problems he says are bottlenecked by materials

"We sell materials," and why he thinks the big materials companies are not incentivized to take risks

Hypersonic flight at Mach 5, and what a material has to survive

From investor to founder: calendar optimization, email copy and north of 200 research papers

"Why are we not solving cancer with machine learning?"

Finding a third co-founder after talking to 30 to 40 academics

Feeling vulnerable "every day," and who he talks to

Leaning into risk: quicker decisions with less information

High entropy alloys, and the pace of validating new compositions

Why Radical wants to be fully verticalized and sell to end customers

Inside the robotic lab: about nine tools, with data captured all along the way

What the models predict, and what goes into them

How an agent crafts a hypothesis and the lab sends the results back up

Where humans are still in the loop

Hallucinations, in-house models and models from the frontier labs

The moat: data from the lab, and later from processing, scale and manufacturing

Why failure data matters: the 90 to 95% of iterations that papers leave out

Two clocks: long-term vision and short-term execution

Where he sees AI's biggest opportunity

"You are not here for a job. You are here for a mission."

The torch test at 3,000 degrees

Three books: Essentialism, The Nvidia Way and The Age of Reason

Jorge Colindres

LinkedIn: https://www.linkedin.com/in/jorgecolindres/

Radical AI: https://radical-ai.com

Mentioned in the episode: Essentialism; The Nvidia Way; and The Age of Reason by Thomas Paine

Timestamps

00:00:00 Why is 95% failure rate actually the goal?

00:02:27 Are you an AI company or a materials company?

00:04:48 What does Radical AI actually sell?

00:07:15 How did three cofounders from different worlds find each other?

00:12:57 What does founder vulnerability actually feel like?

00:17:13 How did you validate hundreds of alloys while academia did 3,500 in 40 years?

00:21:12 How does the robotic lab actually work?

00:25:49 How does an AI agent design a scientific hypothesis?

00:28:11 Where do humans still matter in the loop?

00:32:02 Where does the moat actually live?

00:36:21 Why is failure data more valuable than success data?

00:38:05 How do you think long-term when everything moves fast?

00:42:08 Could this approach work for cancer research?

00:44:04 What kind of person thrives at Radical?

00:45:52 What happened when you torch-tested at 3,000 degrees?

00:51:05 Outro

From a document site that drew cease and desist letters within days to a company built to get authors paid by AI: Trip Adler is the co-founder and CEO of Created by Humans, which he describes as the company that is sticking up for human creativity. Its aim, in his words, is to solve copyright in the AI era, with a business model where rights holders get paid for their data and AI companies get the data they need with legal confidence. So far the focus is books. Before this he started Scribd in college and was there 17 years, and by his account that is where he learned copyright, by first getting it wrong.

In this episode, Trip joins TwoSetAI to argue that copyright and AI can coexist. The conversation has become polarized, he says, between ending copyright law and following it to the letter, and his answer is a new set of rights. Created by Humans has defined three categories of what it calls AI rights, and he walks through what each one covers, which two are getting the most traction and which one is held back while the lawsuits play out. His comparison is music: Napster was a legal gray area too, and the answer to it was iTunes and Spotify.

You can learn more about Created by Humans here.

We go deep on how the deals work. Trip describes one that had not been announced when this was recorded, with a well-known AI company he does not name: books are licensed for use through RAG, an answer found on a page of a book links back to where the book can be purchased, and incremental AI revenue is shared with the author and publisher. Payment so far is mostly by usage, he says, though he can imagine a revenue share or upfront payments. He explains the multi-rights holder problem, where clearing one trade book can take sign-off from the author, a co-author, more than one publisher and a literary agent, and the platform his team built so an author can sign up, verify their identity, select the rights to license and sign an agreement with one click. By his account Created by Humans has done deals with big AI companies and with a government agency, has interest from startups and universities, and finds that most of the bigger AI companies already have data acquisition teams whose job is to license data.

We also get into the harder parts. Opinions on both sides are all over the place, he says: some authors do not want the conversation at all, and some AI companies want the lawsuits to play out before they do deals. Many of those companies, by his account, have already trained on the books using pirated versions, which is where it gets trickiest. He will not name a fair price ("we're not in the business of deciding what's fair"), and when Angelina sketches how she would build a rival marketplace, he agrees that getting AI companies to want this is one of the challenges. He also covers the 90% human-generated line the product uses, why he left Scribd and calls starting from scratch a joy, why he thinks most AI companies will eventually license willingly, and what he tells founders who ask how to keep going: persistence, and blind faith.

We discuss:

* Why the company is called Created by Humans, and what it is trying to solve

* The big bet: that copyright and AI can coexist

* His first emotion at the lawsuit headlines: confusion, and a little sadness

* Angelina's own conflict as a creator and as an engineer who trains models

* The three categories of AI rights, and which ones are getting traction

* Why buying and scanning every book is not an efficient model for AI companies

* Speed and cost: what AI companies want from a licensing deal

* How the lawsuits are changing what AI products will output

* Authors who will not have the conversation, and concerns beyond money: creative control and misrepresentation

* His prediction for the courts: some sort of middle ground

* Who is licensing: big AI companies, startups, a government agency, universities

* A deal not yet announced with a well-known AI company, and paying by usage

* Advice for indie founders: the earlier you embrace copyright, the better

* How Scribd taught him copyright: pirated uploads, cease and desist letters, a DMCA takedown process

* Napster, then iTunes and Spotify

* Why transformative rights are the most complicated, and the ones he is most excited about

* Angelina's sketch of a rival marketplace, and "Are we going to be competitors?"

* Getting AI companies to the table: brand, relationships and data acquisition teams

* "We're not in the business of deciding what's fair": finding a market price one deal at a time

* Microsoft, Amazon and the multi-rights holder problem

* Scribd's steady growth, by his account an average of 30% a year, and long-term bets

* The checkbox for human-written books, and the 90% line

* Rapid fire: one-person billion-dollar startups, Bitcoin and Claude Code

* Persistence and "blind faith"

* Three books: one for the present, one for his 20-year-old self, one for the future

Trip Adler

* LinkedIn: https://www.linkedin.com/in/tripadler

* Created by Humans: https://www.createdbyhumans.ai

* Mentioned in the episode: Good to Great by Jim Collins; Siddhartha by Hermann Hesse; and Barbarian Days by William Finnegan

Timestamps

00:00:00 Did I say your company name right?

00:02:00 What's your big bet on solving copyright in AI?

00:03:18 When Britannica sued and fair use won

00:05:10 Three new categories of AI rights

00:09:50 Why AI companies come to you instead of building this

00:13:20 What kind of AI companies are willing to pay?

00:18:20 The RAG deal that's working

00:22:30 The Napster-to-iTunes analogy for AI

00:26:30 Naive implementation test

00:27:30 The multi-rights holder problem

00:35:00 Scribd's 17-year growth vs starting over

00:36:30 Why did you leave Scribd after 16 years?

00:42:30 What AI tools do you use daily?

00:44:30 "It's just blind faith"

00:46:00 Book recommendations

00:48:00 Outro

A former rapper, a university professor, a TEDx speaker and a LinkedIn coach: Jasmin Alić, who goes by Jay, was running the Link Up community at the time of this conversation and coaching founders on brand and business. By his own count he has coached more than a thousand people, around 300 of them one-on-one and 800-plus in his community, and trained more than 60,000 through live sessions, workshops and webinars. He says he never wanted to put himself in the LinkedIn bubble, that LinkedIn is just the starting point for the work, and that what he wants to be known for is helping people do what they thought they couldn't.

In this episode, Jay joins TwoSetAI to argue that writing is thinking, and that founders who hand all of it to AI are sending someone else to the gym to lift the weights. By his account, scientists and neuroscientists were tracking the human brain after ChatGPT launched, and for the first time in history it has slowed down. He puts the cause plainly: we have outsourced our thinking. His fear for his clients and students is a business that outgrows its founder. He says he is not against AI in any sense; his point is about writing, and he does not use AI at all for his own.

You can learn more about Jay's work here.

We go deep on LinkedIn as a go-to-market channel for technical founders. Angelina lays out her own plan (optimize the profile, batch write 30 posts for the next 60 days, hire a virtual assistant to comment) and Jay starts somewhere else: the funnel. Do not send people to the homepage, he says; build a shorter LinkedIn landing page that is linked only from the profile, with pricing near the top, because the posts have already done the educating. For copy, his shortcut is to start at the outcome and show how you save people stress, time or money, which he walks through with Samsung's privacy screen. He explains his BND framework (broad, niche, deep), why the first three lines of a post should carry no niche language, and why the audience that matters is the network of whoever engages: target networks, not individuals. A virtual assistant can take over comments and DMs, he says, but automation should not, because by his account LinkedIn is cracking down on it.

We also get into the harder parts: why going viral is usually bad for a business, the creators with 800,000 followers who came to him making zero dollars from LinkedIn, and why he tells people to say no to engagement pods. Jay explains how accepting every connection request builds what he calls a dead network, so that engagement drops as the account grows. He is blunt about reciprocity: comments do not come back one for one, and he says his own reach rests on leaving thousands of comments a week for the last 5 years. Angelina adds a fact-checking story of her own, a number in a ChatGPT research report that matched its cited source, where the source itself was wrong, and calls it a compounding error. Jay closes on the one thing he says AI cannot take away from us.

We discuss:

* How many people Jay has coached and trained, and what he wants to be known for

* Why he never wanted to be put in the "LinkedIn bubble"

* Why tech founders build visibility "the harder way": Product Hunt, conferences, PR and SEO

* Angelina's LinkedIn go-to-market plan, and what Jay would look at first

* Why founders should not send LinkedIn traffic to their homepage

* The LinkedIn landing page: shorter, pricing sooner, linked only from the profile

* The blue link under your name, and the post as a lead magnet

* The copy shortcut: start at the outcome, then name the feature

* Stress, time and money, the three things he says people most worry about wasting

* Samsung's privacy screen as an outcome-first example

* A prompt for three positioning statements under 20 words

* BND (broad, niche, deep) and the first three lines of a post

* Target networks, not individuals

* Why the algorithm will not hand a new account its audience: "Coach, put me in"

* Hiring a virtual assistant for comments and DMs, and why he says not to automate

* His claim that the human brain has slowed down, and the fear that a business outgrows its founder

* Managing AI versus letting AI manage you

* Angelina's "compounding error" story from fact-checking a research report

* "Writing is thinking," the gym analogy, and the you versus AI competition

* "Start empty": why he says an outline is not enough

* Feedback in the Link Up community: scores out of 10, and a website written live within 2 hours

* Lightning round: likes, profile views, and creators with 800,000 followers making zero dollars

* Engagement pods, reciprocity, and leaving comments every week for 5 years

* Connections versus followers, and how a dead network lowers reach

* His closing message for an AI world

Jasmin Alić

* LinkedIn: https://www.linkedin.com/in/alicjasmin/

* Hey Jay (consulting): https://hey-jay.com

* Link Up community: https://whop.com/linkupbyjasmin/linkupbyjasmin?a=angelinamagr

* Mentioned in the episode: the Link Up community and its in-house writing coach, Victoria Michaels; and BND (broad, niche, deep), his framework for writing posts

Timestamps

00:00:00 Cold open

00:01:08 Meet Jasmin Alić: who he coaches and what he wants to be known for

00:06:15 Why don't more tech founders post on LinkedIn?

00:09:17 What's wrong with my LinkedIn go-to-market plan?

00:10:28 Why shouldn't founders send LinkedIn traffic to their homepage?

00:15:50 What's the shortcut for writing copy that actually sells?

00:18:34 How does the Samsung privacy screen prove outcome-first copywriting?

00:22:59 Why does talking to your target audience keep you invisible?

00:24:09 What is the BND (Broad Niche Deep) framework?

00:26:59 How do you target networks instead of individuals?

00:31:41 The human brain has slowed down, and AI caused it

00:33:13 Will your business outgrow your own brain?

00:38:16 "Writing is thinking": why he refuses to use AI

00:38:53 The competition exercise: you vs. AI, same topic

00:41:59 Writing feedback, scores out of 10 and a website written live

00:45:52 800K followers making $0: what went wrong?

00:47:01 Why virality is usually bad for your actual business

00:49:18 Should you use LinkedIn engagement pods?

00:50:55 The compound interest of genuine engagement

00:54:49 How do LinkedIn connections kill your reach over time?

01:00:04 What's the one thing AI can't take away from us?

01:01:32 Outro

From a physics professorship to software for humanoids: Jan Liphardt started out as a physics professor at UC Berkeley, moved into healthcare and computing on sensitive data, and at the time of this conversation is a professor of bioengineering at Stanford and the founder and CEO of OpenMind, the company building what it calls the Android for robots. About two years before the recording he got curious whether large language models could do more than help his students with their homework and could also control physical hardware. His picture of it: give a ChatGPT arms, legs and a head, and let it explore the park and talk to people.

In this episode, Jan joins TwoSetAI to make the case that the robots are already here, in far larger numbers than most people know, and mostly in places most people never go. He names one company that has deployed 1.1 million of them, and describes hospitals where the robots that move the laundry work in the middle of the night, so visitors never see them. A humanoid in a typical home or workplace is still extremely rare, he says, and his own house is one of the exceptions.

OM1, OpenMind's open-source software stack, is on GitHub here.

We go deep on the hardware and the software. On hardware: a humanoid that Jan expects to cost about $20,000 this year, a supply chain that he says sits mostly in Asia, and his suspicion that the real suppliers of humanoids at scale will be car companies, with their decades of experience making millions of reliable things (he names Tesla, BYD, Xiaomi and Hyundai). He explains why he thinks a kid-sized form factor wins for many uses: the Booster is about a meter in height and weighs 24 kilograms, so it can be picked up with one hand. On software: OpenMind's app store for humanoids, with education, home safety and health companionship as the first categories, and OM1, which handles decision-making and data fusion one layer above most robotics models. In OM1 everything becomes text. Some 10 to 15 specialized models talk to one another in natural language, which makes the system easy to eavesdrop on and to put guardrails on, at the cost of making big decisions only every second or two. That rules out drone targeting, ballerina humanoids and onion chopping.

We also get into the harder parts. A robot at home is a camera running around the house and, in his words, a big battery on legs, and when Angelina asks how far away safe is, he compares it to asking when the internet will be perfectly safe: never. OpenMind started from Asimov's three laws of robotics and has stayed with them for the last year and a half; Jan says the company lacks the legal or philosophical acumen to do better, and asks lawyers, historians and philosophers for a better take. On hallucination, he has graded undergraduate homework for 15 years and says it is not a computer thing. OpenMind's own answer is a second model, the "mother LLM", that watches the robot and writes a critique every 30 seconds, plus humans in the loop. He also explains what he meant by a time of maximum uncertainty and least preparation, why he tells his kids that learning one skill in college for life is "completely dead", and why OpenMind is a .org: the .com was 2 million, he says, and he does not want a future where a robot's software is "secret and magic".

We discuss:

* How a physics professor got curious about giving a large language model arms, legs and a head

* Physical AI, embodied AI, world models: why he says the terms are not settled yet

* The humanoid he expects to cost about $20,000 this year

* Hardware supply chains in Asia, and where he thinks the US is behind and ahead

* The boxes of masks that arrived from research collaborators in China during COVID

* Whether a humanoid becomes a household essential

* Where robots are already deployed today, and why most people never see them

* Who he expects to build humanoids at scale: the car industry

* Tall and strong versus kid-sized: the Booster, about a meter in height and 24 kilograms

* Why a humanoid shape at all, when homes, cars and stairs are built for people

* An app store for humanoids, and its first three categories

* The developer who wants a humanoid alarm clock with no snooze button

* The robot dog that started doing his son's math homework, and the prompt change that followed

* Safety at home: cameras, data, batteries and pinch hazards

* "When is the internet going to be perfectly safe?" and trading privacy for convenience

* Asimov's three laws as a starting point, and the school bus problem

* OM1: a decision-making layer where models talk to each other in natural language

* Hallucination, from someone who has graded undergraduate homework for 15 years

* The "mother LLM" that writes a critique every 30 seconds, and keeping humans in the loop

* FABRIC and robots that share skills: "incredible opportunity, but also incredible peril"

* "Maximum uncertainty and least preparation": airplanes, Waymo and the speed of change

* Why he calls one skill from college for life "completely dead", and which universities he thinks last

* Learning by building things, and why he is not writing a book yet

* Why OpenMind is a .org

Jan Liphardt

* LinkedIn: https://www.linkedin.com/in/jan-liphardt/

* OpenMind: https://openmind.org

* OM1 on GitHub: https://github.com/OpenMind/OM1

* Mentioned in the episode: Asimov's three laws of robotics, which OpenMind started from; and The Matrix, Black Mirror and Space Odyssey, science fiction he says his team learns from

Timestamps

00:00:00 Why you've never heard of embodied AI or world models until right now

00:04:12 How much does a humanoid cost in 2026, and what can it actually do?

00:11:38 Amazon already deployed how many warehouse robots?

00:15:39 Who's really building humanoid robots, and why it's not who you think?

00:17:26 The Booster: 1 meter tall, 24 kilograms, pickupable by hand

00:23:19 What can you actually install on a humanoid app store?

00:27:50 His son asked a robot dog for homework help

00:38:09 Everything the robot sees becomes text: here's the architecture

00:46:15 I've been grading essays for 15 years: here's what hallucination really is

00:48:49 The "mother LLM" that watches your robot's every move

00:51:03 One robot learns a skill, all robots learn it instantly

00:54:01 Why this is the time of maximum uncertainty and least preparation

00:57:23 Is the traditional career path dead?

01:03:04 Outro

From six years trying to be a rock star to a newsletter that HubSpot acquired: Adam Biddlecombe started Mindstream, a daily AI newsletter, with his childhood friend Matt Village in June 2023. By his account they began with a blank sheet of paper in an office at the end of his garden, two meters by two meters, where he locked the door and said it would open when they had started a business. They left their jobs three months later, and he says it was 17 months from start to acquisition. At the time of this conversation he is a year and a half into his job at HubSpot, where he describes his job as building the Mindstream brand.

In this episode, Adam joins TwoSetAI to talk about what a media brand is for. Angelina asks how a founder interview show should think about goals, sponsors, paid tiers and platforms, and Adam answers from what he did at Mindstream, including his first order of business with the paid subscription that came with a newsletter he acquired. His view of where creators go wrong when they build a brand runs through the whole conversation.

You can read Mindstream here.

We go deep on how Mindstream grew. Adam describes it as one deal growing to a bigger deal: an investment from someone he knew that came with a bookkeeper, SEO support, brand support and mentorship, which he says was probably more valuable than the cash; then the December 2023 acquisition of another newsletter, AutoGPT, which by his account took Mindstream from about 15,000 subscribers to over 100,000; then HubSpot. He walks through who did what on a small team, why he had the editorial team study The Hustle every day, and why the newsletter is written by people while he uses AI to write all of his own LinkedIn content. He says 99% of Mindstream's sponsorships came inbound, and that probably 30% to 40% of that inbound came from how Mindstream showed up for one Google search. He also explains why he thinks building a community is 10 to 100 times harder than building an audience, and the difference between repurposing content and re-sharing it.

We also get into the harder parts. In February 2024, three months after the AutoGPT acquisition, a spreadsheet the founders had spent hours building told them the business was eating itself. Adam recounts the night that followed and what they found the next day; two months after that, he says, they started talking to HubSpot. He is open that only three years before this conversation he was in a very bad place, unhappy with his work and with himself, and he explains the ideas he leans on now: that action cures fear and anxiety, that any state of mind you sit in is one you are practicing, and that everything in life moves like a stock ticker, either up or down. He also describes his own form of procrastination, checking dashboards seven times a day, and the question he asks himself to cut through it.

We discuss:

* Why Adam's first question is what the end goal is

* Six years trying to be a rock star, and the phone call to Matt that started it

* The garden office, the locked door and a blank sheet of paper

* Why AI was the obvious thing to write about in June 2023

* Leaving their jobs three months in, and going after revenue and funding at the same time

* An investment that came with services, and why he rates those above the cash

* The AutoGPT acquisition: from about 15,000 subscribers to over 100,000

* "One deal growing to a bigger deal": how the raise, the acquisition and HubSpot connect

* Whether building to sell gives you the right motivation to keep going

* Content to commerce, and why he says more tech companies are acquiring media companies

* A personal brand as an opportunity magnet

* Starter Story, built over 8 years, next to Mindstream's 17 months

* Sponsors for an interview show: the top five guests and top five sponsors exercise

* Is the content the product, or the funnel for something else?

* What he did with AutoGPT's paid subscription, and what The Hustle did differently with Trends

* Why community is 10 to 100 times harder than audience, and why churn shows

* One platform or many: LinkedIn, X and the 80/20 rule

* Repurposing versus re-sharing

* A newsletter signed off as written by humans, and where he does use AI

* Who did what on a small team, and who sold the ads

* The top AI newsletters search behind 30% to 40% of inbound

* AEO, and asking Claude or ChatGPT for the next audiobook

* Inside HubSpot: being brutal with time, meetings and dashboards

* The spreadsheet that said the business was destined to die

* Action cures fear and anxiety, practicing anxiety, and the stock ticker

Adam Biddlecombe

* LinkedIn: https://www.linkedin.com/in/adam-biddlecombe/

* Mindstream: https://mindstream.news

* Mentioned in the episode: My First Million, his favorite podcast when he started; The Hustle and its Trends product; Starter Story; and AutoGPT, the newsletter Mindstream acquired

Timestamps

00:00:00 What's the real goal when building a media company?

00:04:11 Should you build to sell or build to keep?

00:06:10 How do you start a newsletter from nothing?

00:08:40 When is the right time to raise funding?

00:10:12 Why would you give away 30% at 10K subscribers?

00:12:29 Does building to sell give you the right motivation?

00:14:07 What does content-to-commerce look like inside HubSpot?

00:17:46 What does the Starter Story acquisition signal for media?

00:21:05 How should interviewers think about sponsorships?

00:24:50 Is the content the product or the funnel?

00:26:36 Why would you kill paid subscribers on day one?

00:29:10 Why is community 10-100x harder than audience?

00:32:47 Should you expand beyond one platform?

00:35:33 What's the difference between repurposing and resharing?

00:40:13 How did a 2-person team run a 200K newsletter?

00:47:05 How did one blog post drive 40% of sponsorship revenue?

00:49:22 How does AI search change content strategy?

00:50:18 What surprised him about working inside HubSpot?

00:52:11 What's the most important thing you could be working on?

00:57:26 What almost killed Mindstream?

01:00:55 How does "action cures fear and anxiety" work in practice?

01:01:36 Are you practicing anxiety without knowing it?

01:04:15 What is the stock ticker philosophy for life?

01:08:04 What does personal transformation actually look like?

01:13:00 Outro

From pens and paint to the canvas other companies build on: Steve Ruiz has a fine art background and, at the time of this conversation, is the founder of tldraw, an engine for building canvas apps such as whiteboards and design tools. tldraw grew out of an open source project he meant to finish in the couple of months before starting a job at Adobe. It got popular, then kept getting popular, and by his account it was about a 4-year-old startup with a team of 15 people, most of them in London.

In this episode, Steve joins TwoSetAI to explain why he did not turn a viral demo into a company of its own. In 2024, when GPT-4 with vision had just come out, his team made Make Real: draw an interface on the canvas, click a button, and a model sends back a working prototype. He calls it one of the first pieces of software that used AI to let people without programming skills do things, and says it broke the internet when it came out. Asked why he did not become a Lovable or a Bolt, he tells Angelina she sounds like his investors, then gives his reasons. To him it felt like a pattern, not a product, "a thousand products and a thousand applications rather than just one," and he wants tldraw to be the software those applications build on.

You can learn more about tldraw here.

We go deep on what makes this canvas different. Most canvas apps, he says, show an image that is repainted the way a video game would do it. His bet was that the normal building blocks of a website, HTML and CSS, could be made just as fast, so that anything that runs on the web can sit on the canvas. He walks through who builds on it: Observable, the data visualization tool he names as a favorite customer, AI design tools such as UX Pilot, Variant and Magic Path, and internal collaboration tools, with Shopify as his example. He explains the pricing, which is sold to teams and not to individuals: one price under 10 people, a negotiated price above that. He also demos fairies, AI agents that live on the canvas, elect a leader, split up a task and work in parallel, and tldraw Computer, where blocks connected by arrows become what he calls a little machine.

We also get into the harder parts: why he could not have sold access to Make Real even if he had wanted to (people had to bring their own API keys, and by his account only about 20% of visitors knew enough to do that), and whether charging $25 for fairies worked out. His answer is that it was still not profitable, but pretty close, and that most of these projects are marketing by building. He argues that group chats, with or without an AI, are one-dimensional, while on a canvas everyone can work at the same time. On the future of work he is, in his words, slightly booster: he thinks we have had a shallow engagement with the knowledge work that could be done, and that there will still be more to do. He closes with the AI tools he uses day to day and the science fiction series he recommends to new grads.

We discuss:

* The open source project he meant to finish before taking a job at Adobe

* The explain-it-like-I'm-five version of a canvas, and the technical one

* Why tldraw is "just a normal website" built out of HTML and CSS, and why that is unusual for a canvas

* "Appropriately overengineered": what changes when the product is an SDK

* What people do on tldraw that he never planned for: Bible study groups, technical interviews, language tutoring

* Make Real in 2024: from a drawing to a working prototype

* Debugging a generated page by drawing on top of it

* "You sound like my investors": why he did not become a Lovable or a Bolt

* API keys, 30 responses per day, and a phone call with OpenAI

* A product or a pattern: "a thousand products and a thousand applications"

* A liquid simulation someone else built on tldraw's geometry system

* Why a fine artist chose the canvas: software where the details matter

* Observable: SQL queries, data and visualizations on one canvas

* Who buys a canvas SDK, and how tldraw prices it

* Starter kits, and an agent that draws back on the canvas

* Canvas or chat as the interface for AI, and why he calls group chats one-dimensional

* Fairies: agents with a location that elect a leader and work in parallel

* Whether charging $25 for fairies was profitable, and "marketing by building"

* tldraw Computer, and "the world's most expensive counter"

* PMs, marketers and multi-stage prompting flows: who builds without being an engineer

* Why he thinks "a grounded familiarity" with AI will matter

* "God in a box": what programmers do once they have these tools

* Claude Code, Cursor and Gemini Flash: what he uses day to day

* Red Mars, and fearlessness about learning something new

Steve Ruiz

* LinkedIn: https://www.linkedin.com/in/steve-ruiz-61a150239/

* tldraw SDK and docs: https://tldraw.dev

* tldraw web app: https://tldraw.com

* tldraw on GitHub: https://github.com/tldraw/tldraw

* Make Real: https://makereal.tldraw.com

* tldraw Computer: https://computer.tldraw.com

* Mentioned in the episode: Red Mars by Kim Stanley Robinson, the first book in the series he recommends

Timestamps

00:00:00 Why didn't he become Lovable after inventing vibe coding?

00:01:29 Who is Steve Ruiz and what is tldraw?

00:02:33 How do you explain a canvas to a seven year old vs an engineer?

00:08:08 How did the first AI vibe coding tool get built before Lovable existed?

00:12:35 Why didn't you just become Lovable or Bolt?

00:14:54 What happens when a stranger builds liquid physics on your canvas?

00:16:55 Why did a fine artist choose to build canvas infrastructure?

00:20:47 How did Observable turn a canvas into data visualization notebooks?

00:23:00 Who buys canvas infrastructure and how does tldraw price it?

00:25:25 What can you build with tldraw inside a chat app?

00:29:29 Is canvas replacing chat as the default UI for AI?

00:31:13 What are AI fairies and how do agents coordinate on a canvas?

00:35:25 Was charging $25 for fairies actually profitable?

00:36:55 What is tldraw Computer and how do executable workflows work?

00:42:32 Are non-engineers building real products with canvas tools now?

00:47:39 How are you raising your daughter in the age of AI?

00:49:08 What happens when engineers get handed God in a box?

00:51:46 What book would you recommend to new grads starting their career?

00:53:21 What strikes me most about this conversation

Dara Ladjevardian had no mentors when he started his first company. So he built a digital version of his grandfather — one of the most successful entrepreneurs in Iran. That experiment became Delphi, now powering digital minds for Arnold Schwarzenegger, Jay Shetty, and thousands of high-inbound professionals. In this conversation, Dara shares why he believes AI is not the main character, why he killed video and bet on voice, and the day OpenAI and Anthropic banned him — nearly ending the company.⏱️ Chapters:00:00 - He built an AI of his grandfather01:36 - The muscle every AI founder needs to build07:05 - Will AI affect education? 08:47 - "AI is not the main character"10:09 - How is Delphi different from LinkedIn?15:04 - How is Delphi different from Character AI?17:29 - How do you build a digital mind?22:39 - How does Delphi handle hallucination?24:35 - Why he killed video and bet on voice27:37 - The idea he'll never build30:26 - How do people communicate with Delphi vs. ChatGPT?34:27 - The day OpenAI and Anthropic banned him36:28 - His post-AGI prediction39:26 - How will our lives change with AI?42:02 - Is AI learning net positive or net negative?45:42 - outro🔗 Guest Links:Talk to Dara's Delphi: https://delphi.ai/daraDelphi: https://delphi.aiDara on X: https://x.com/daraladjeQuestions answered in this episode:What is Delphi and how does it work? → Delphi lets you create a digital version of your mind that can answer questions 24/7 in your voice. You upload content — YouTube videos, podcasts, Notion, Google Drive — and Delphi builds a temporal knowledge graph that captures not just what you know, but how you reason.How is Delphi different from LinkedIn? → LinkedIn shows credentials — where you went to school, where you worked. Delphi shows what you know and how you think. The best relationships aren't based on someone's resume; they're based on values, experiences, and perspective — things a LinkedIn profile can't show.Who should use Delphi? → Anyone with high inbound. Celebrities like Arnold Schwarzenegger use it to scale mentorship. But most users are founders, investors, consultants, and content creators who answer the same questions repeatedly. "Never repeat yourself again" is the core promise.How does Delphi prevent hallucination? → Delphi has two modes. "Strict" only answers questions it knows directly — otherwise it says "I don't know" and redirects. "Adaptive" infers what you might say based on related data. Hallucinating someone's opinion is higher stakes than hallucinating a fact.Why did Delphi stop doing video? → Video AI is a "shiny object" — people try it once but don't come back. Once the novelty wears off, you notice the quirks and it becomes creepy. Voice and text leave room for imagination, which is essential for users to suspend disbelief.Why won't Delphi build a legacy/grief product? → "You don't want to be a company associated with monetizing grief — if you stop paying us, you lose your grandfather." It's the most-requested feature, but Dara believes it's a horrible strategy that would make people hate the company.What is Dara's post-AGI prediction? → When AI is smart, fast, and nearly free, humans become the scarce resource. We'll crave things with lived experience. You watch human chess players even though AI beats them. The bet: humans matter more in a post-AGI world.What does "AI is not the main character" mean? → Other companies try to automate humans. Delphi says AI should augment humans. You own your data, your identity. AI scales you — it doesn't replace you. A Delphi is an artifact of your mind, like a book or painting, that people can experience through conversation.#Delphi #DigitalMind #AIStartup #FounderInterview #AIClone #PostAGI #TwoSetAI

From visuals for nightclubs and fashion shows to real-time 3D worlds on festival stages: Joris Corthout is the CEO of Prismax, which he says he founded exactly 20 years before this conversation as a graphic studio for the entertainment industry. Angelina introduces Prismax as the team behind the visuals for Tomorrowland, and Joris calls the festival the biggest of his clients. When COVID stopped the events industry, the company was eight people and he was asking when it would go bankrupt. He and his business partner, both gamers, decided to rebuild their work on Unreal Engine, the game engine they had been playing in for a month.

In this episode, Joris joins TwoSetAI to show what his team builds and where generative AI fits into it today, which by his account is in the concept phase and almost never in production. He takes Angelina through a demo reel that includes Alicia Keys at a piano that was recreated in 3D, and he explains why, with all the AI tools his team tries, a show is still run by people.

You can see Prismax's work here.

We go deep on real-time rendering: why a film frame could take a day to render while a game engine calculates everything on the spot, and how data from the fireworks, the lights and the lasers sets off the virtual effects on screen. Joris describes the first virtual Tomorrowland, planned as one or two stages and delivered as eight stages with a hundred DJs over two days; a stage that is 150 meters wide and 40 meters high; and the biggest real-time show his team has done, at 16,000 by 16,000 resolution. Against that he sets what AI video gives him now: HD, and sometimes not even 4K, without the control his 3D artists have. He says an AI startup texted him two days earlier about VJ software whose effects are generated in real time by AI, and he names the two things he would need to see before using a tool like that in a show. He lists what the team does use, from Midjourney for storyboards and mood boards to HeyGen, which made a CEO's hologram speak eight languages, and he puts AI's share of actual production at 0.05% of the time.

We also get into the harder parts. Joris thinks generating video with AI uses an amount of energy that is not sustainable, and that the big companies will at some point have to charge what it costs; Angelina cites a professor she had interviewed on the water used to train one large model. He says people who make commercials have a real problem, that everybody is afraid, including in his own company, and that creative people who embrace AI as a tool will be steps ahead of their competition. He describes what he would ask an AI studio to build for Prismax, what he thinks a machine would get wrong if it built a show, and what he would rather hear than a robot playing the violin. His next product, he says, puts a recorded orchestra of 120 musicians inside worlds built with game technology.

We discuss:

* How a graphic studio for nightclubs and fashion shows ended up on festival stages

* COVID, eight people, and the question of when they would go bankrupt

* Two gamers and the decision to rebuild on Unreal Engine

* The first virtual Tomorrowland: eight stages, a hundred DJs, two days

* Real time versus rendered: why nobody waits 48 hours for a change anymore

* Virtual flames that go off with the real ones

* The "virtual window": a room of LED walls showing a 3D world with real weather data

* Who wants one, and whether it will be accessible to everybody in another 10 years

* Tomorrowland next to Glastonbury: 400,000 people in one weekend

* The demo reel: Formula One, a virtual festival and Alicia Keys at a piano that was not real

* A show at 16,000 by 16,000 resolution, synced to the lights and the lasers

* Midjourney for storyboards and mood boards: one or two days where a proposal took one or two weeks

* Will Smith eating spaghetti as a reference for how fast AI video moves

* What AI video still lacks for his work: resolution, control and real time

* The AI startup that texted him, and Doom simulated in AI

* Why he says it takes something like COVID to make the next step

* Veo 3, HeyGen, Runway and Grok: what the team is trying

* A CEO as a hologram speaking eight languages

* Where AI output gives itself away: hair, lamps and millions of leaves

* Energy, water and the price of RAM

* ChatGPT on his contracts, Gemini on his emails

* What he would ask an AI studio to build for Prismax

* "A show should be like a sinus"

* A real violinist or a robot

* An orchestra of 120 musicians in a game technology world, and who he is hiring

Joris Corthout

* LinkedIn: https://www.linkedin.com/in/joris-corthout-92740a4/

* Prismax: https://www.prismax.com/

* Mentioned in the episode: Unreal Engine, Midjourney, Runway, HeyGen, Grok, Google's Veo and OpenAI's Sora

Timestamps

00:00:00 What does 16K resolution running in real-time actually look like?

00:01:46 Who is Joris Corthout and what does he do?

00:03:00 How did COVID force a pivot of the business?

00:04:56 What makes real-time 3D different from traditional rendering?

00:08:08 What is the virtual window experience billionaires are buying?

00:11:25 How big is Tomorrowland compared to Glastonbury?

00:14:00 Was Alicia Keys' piano real?

00:16:09 How does Prismax use AI in production?

00:18:46 What would an AI startup need to show before he'd use it in a show?

00:23:04 Which AI tools is the team actually experimenting with?

00:30:28 Why does AI generation consume so much energy?

00:39:06 What do humans understand about shows that AI can't replicate?

00:41:38 A human violinist vs. a robot?

From the first non-engineer at a developer tool company to chief operating officer: Alexander Berger joined in 2020 as employee number seven, when the company was called StackBlitz and sold a code editing environment that ran in the browser. He came in as its first enterprise sales rep, closed a first sale of $13,500, then worked through most of the other departments. Today he is COO of Bolt.new, which Angelina introduces as the AI platform that went from nearly shut down to 40 million in ARR in just five months. All seven of those early employees, he says, are still at the company.

In this episode, Alexander joins TwoSetAI to tell the story of that last experiment. By the final stretch of 2024 the company was four years in, under 1 million in annual recurring revenue, with a team of about 15, and was thinking about shutting down and returning the remaining capital to investors. It decided to run one more big experiment first: hook AI into the product and put it out there. Bolt launched on October 3rd, 2024 with nothing but a tweet thread, and he says he has never experienced, and probably never will experience, another product launch like it.

You can try Bolt.new here.

We go deep on how that launch worked and what the product does. Alexander builds a calculator app with a race car and confetti effects on screen while he explains Bolt for an elementary schooler and a college student. He breaks down the tweet thread format the team still uses for product launches: he and Eric, the CEO, spent at least 12 hours on the thread, most of it on the first 240 characters, because by his account the first tweet gets 90% of the views, and no link goes in it. He says the two accounts behind the launch had roughly 40,000 followers, mostly frontend developers, yet over 70% of the people signing up had no previous coding experience. He also covers Web Container, the technology that runs the whole development environment in a browser tab, two features he says a lot of users do not know about (visual inspector and plan mode), and why he prefers "software composing" to "vibe coding".

We also get into the harder parts: his warning to founders who are not 100% sure they have product market fit to be very careful with cash, why he does not believe a company is a family, and what is still unsolved. Getting AI-generated code to match an enterprise's design system and code requirements is, in his words, the hardest problem at the cutting edge of what Bolt is working on. On how product, design and engineering teams will be organized, he says the book has not been written. He argues that anybody can build software now, that doing it at a professional level is a skill, and that the skill is not software engineering. He also says where he would put his money if he had to start a company with Bolt, how he positions Bolt against Cursor and the other vibe coding tools, and which books he would hand to a 20-year-old, a 40-year-old and a 60-year-old.

We discuss:

* Employee number seven, and the first non-developer at a developer tool company

* The first enterprise sale: $13,500 against a $54,000 target, then a six-figure deal

* Bolt explained for an elementary schooler and a college student, with a live demo

* Why he says these tools were not possible until Anthropic came out with Sonnet 3.5

* "All jobs are really the same thing": an objective, tools and constraints

* What a COO does at a startup: turning the CEO's vision into tactical execution

* StackBlitz, the design system teams that loved it, and why it was too niche

* Four years in, under 1 million in ARR, and the decision to run one more big experiment

* Web Container, and the business model of "renting out their own computers to them"

* Being very careful with cash until you are sure of product market fit

* "We're not a family... we are a team": keeping the original squad together

* The launch: a tweet thread, roughly 40,000 followers, and signups from people who had never coded

* 12 hours on one thread: the first 240 characters, no links in the first tweet, a graphic for each point

* Visual inspector and plan mode, two features he says a lot of users do not know about

* "Vibe coding" versus "software composing"

* Who can compose software, and the skill that is not software engineering

* Learning the language of digital products: hero section, pancake menu, padding, margin, APIs

* Where we are on the adoption curve: "ahead of 99% of the population"

* Making money with Bolt: why distribution matters more now, and the case for agencies

* Landing a client with a prototype built in five minutes: "show instead of tell"

* How he positions Bolt against Cursor and the other vibe coding tools

* Hallucinations and enterprise code requirements: engineers deployed into customers' companies

* How product, design and engineering teams are changing, and his book recommendations

Alexander Berger

* LinkedIn: https://www.linkedin.com/in/alexanderbergerlinkedin/

* Bolt.new: https://bolt.new

* StackBlitz: https://stackblitz.com

* Mentioned in the episode: Rich Dad Poor Dad by Robert Kiyosaki; The Intuitive Parent, the book by his father-in-law; and Ender's Game by Orson Scott Card

Timestamps

00:00:00 Intro

00:01:04 Who is Alexander Berger, COO of Bolt.new?

00:06:20 What is Bolt.new and how does it work?

00:12:07 Can Bolt.new build a full production app or just prototypes?

00:18:38 How did Bolt.new grow from $1M to $40M ARR in 5 months?

00:26:15 How did Bolt.new go viral on Twitter?

00:29:36 How to write a viral product launch tweet

00:37:00 What is vibe coding and why does Alexander Berger disagree with the term?

00:38:48 Can I build an app without knowing how to code?

00:40:18 Is Bolt.new good for beginners with no coding experience?

00:46:22 Can I make money building apps with AI tools?

00:47:42 How to start a software agency using AI coding tools

00:49:11 How to get clients for AI app development

00:49:38 Bolt.new vs Cursor vs Lovable: Which AI coding tool is best?

00:55:28 Book recommendations for founders and builders

About 20 years in AI, from computer vision research to generative models: Devi Parikh was faculty at Virginia Tech and Georgia Tech, then spent about eight years at Meta, the first five at FAIR and then a couple of years in its GenAI organization, where she was a senior director leading the multimodal research teams. In early 2024 she and her co-founders left to start Yutori, where at the time of this conversation she is co-CEO with Abhishek Das, alongside co-founder and chief scientist Dhruv Batra. As she describes it, Yutori builds web agents that do the clicking for you.

In this episode, Devi joins TwoSetAI to take apart a naive design for her own product. Angelina describes how she would build Scouts, Yutori's product for monitoring the web, from scratch: a search API, query rewriting, similarity matching and an email digest. Devi explains what that misses. Much of what people want from the web, she says, sits in databases behind forms and filters, a heavy tail that a search engine does not list: whether a local tennis court is free on Monday at 7:00 a.m., an apartment that fits your filters, the price of a ticket. Getting to it takes an agent that clicks.

You can try Scouts here.

We go deep on how the agents work. Devi says Yutori trains its own in-house browser use model, and that it reads screenshots of the page, not the HTML: her team started with the HTML and the DOM, found it too noisy and too different from one website to the next, and wrote a blog post on why. Around that model sit APIs and MCP servers, and because models start to fall apart at around 100 tools, she says, the system is a hierarchy: sub-agents that optimize for coverage and recall, and an orchestrator that optimizes for precision and keeps track of what it has already told you. She also offers a mental model for cost. An LLM is a for loop over token generation, an agent is a for loop around an LLM, and a Scout is a for loop over a whole team of agents, which is why, by her account, having the model in-house is what makes the product feasible to open up.

We also get into the harder parts: what hallucination means once an agent can act, and why Scouts was read-only when this was recorded, with evals through the entire stack, citations, and an inspect work button on each report. Devi describes logging in and buying on a user's behalf as roadmap items that need higher user trust and guardrails, such as checking with you on anything over $50. She compares Scouts with ChatGPT Pulse, says the scaling bottleneck would be infrastructure and vendors, and says the right way to deliver reports (email, a text, a digest) is still something her team brainstorms about. She also explains why she left a job Angelina calls arguably the best research job in AI, what the name Yutori means, and recommends two books: Beginning of Infinity by David Deutsch and Scout Mindset by Julia Galef.

We discuss:

* Devi's path: computer vision, vision and language, generative models, academia, then FAIR and GenAI at Meta

* What Yutori does, explained for an elementary schooler

* The "AI chief of staff" idea, and why it is hard to picture today

* Why she left Meta: missing being in the weeds, and three founders who had talked about it for years

* What motivates her after 20 years in AI

* Angelina's naive design for Scouts: search API, query fan-out, cosine similarity, email digest

* The tennis court problem: what Google search is not going to find

* The heavy tail of the web: information behind forms, filters and sliders

* Why Yutori trains its own browser use model

* Screenshots versus HTML and the DOM, and when a text-based approach still makes sense

* A founder's Scout for negative reviews of a competitor

* APIs and MCP servers beyond web search, and why 100 tools is too many for one agent

* Sub-agents for recall, an orchestrator for precision, and reports that do not repeat themselves

* How the orchestrator decides what to report and which sources to trust

* Replying to a Scout's email as feedback, and memory per Scout

* LinkedIn, public information, and why Scouts does not log in for you

* The roadmap: buying on your behalf, and the user who found a flight under $500

* How often a Scout checks

* The for loop mental model: LLM, agent, Scout

* Getting ready for GA, and where the scaling bottleneck sits

* ChatGPT Pulse compared with Scouts

* Onboarding: 70 example Scouts and four suggested ones

* Hallucination, evals, citations and the inspect work button

* Guardrails for purchases, and email versus a text versus a digest

* What "Yutori" means, who the product is for, and two book recommendations

Devi Parikh

* LinkedIn: https://www.linkedin.com/in/deviparikh/

* Yutori: https://yutori.com

* Scouts: https://yutori.com/scouts

* Yutori blog: https://yutori.com/blog

* Mentioned in the episode: Beginning of Infinity by David Deutsch; Scout Mindset by Julia Galef; and Yutori's blog posts on screenshots and on Navigator, its N1 model

Timestamps

00:00:00 Cold Open

00:01:02 Introduction + 20-Year AI Journey

00:02:57 What Yutori Does

00:05:24 Why She Left Meta

00:08:50 My Naive Approach to Building Scouts

00:12:45 Why Screenshots Beat HTML

00:19:32 Multi-Agent Architecture Deep Dive

00:25:02 MCP Servers + Coverage Strategy

00:28:54 The Roadmap: Authentication & Transactions

00:31:49 The "For Loop" Mental Model

00:34:50 ChatGPT Pulse Comparison

00:38:09 Hallucination + Evals

00:43:37 What "Yutori" Means

00:46:45 Book Recommendations

00:51:29 Closing