Intelligence Snacks

The rebrand is official, The Good Stuff is now Intelligence Snacks.

The episode opens with the ColdCard hack that swept ~1,300+ Bitcoin over the weekend. A 2021 firmware bug disabled the hardware RNG and used a pseudo-random one instead. The prevailing theory - Kimi K3, released without guardrails, found the exploit. US models refused to help with offense or defense, they treated code auditing as a cyber attack. So the Bitcoin community turned to open models for red teaming. Crisis response agents, knowledge management, Pliny the eldar and more!

**Key Moments:**

- [03:50] The hack: "Private keys that were not random. Orders of magnitude easier to crack."

- [06:24] "The conjecture is Kimi K3 found the exploit."

- [07:11] "Opus, Fable, GPT 5.6 — Claude models just refused to do anything."

- [11:45] "Rick was pulling Twitter every 10 minutes... validated proof within 10 minutes of where each exploit was in the code."

- [14:20] "Having your agent go, 'Let me handle the noise' — here's step one, two, three."

- [19:30] Pliny: "Part news capture, part educational briefing. If I've never encountered that content, it does an ELI5."

- [24:03] "The thing an X feed gives you is randomness. I couldn't recreate that deterministically."

- [26:15] "You need a subject area graph — what are the obvious nodes missing?"

- [30:22] "The consulting joke — just be two pages ahead of the client."

**Friends of the Pod:** Kevin Loaec, Lloyd, Rod, Shoshow.live

**Concepts Introduced:** Intelligence Snacks (high-leverage AI insertion points), AI situation room (crisis response with agents), retirement attack (vendor-controlled "random" seeds), the serendipity problem (preserving randomness in filtered feeds)

After a short break, Pete and Andy return to the van for what turns out to be the final episode of **The Good Stuff**. They explain why the podcast is getting a new identity before diving into AI agents, social media, and the software stack they've been building.

The conversation explores using personal AI agents to filter information instead of doomscrolling, why stepping away from the AI bubble can be refreshing, and how tools like Wingman, Flight Deck and Autopilot are evolving into an operating system for agent-first businesses.

The episode closes with a discussion about why the future isn't AI replacing software, it's businesses becoming software.

**Key Moments:***

**[00:01]** The final episode of *The Good Stuff* is announced before a complete rebrand.*

**[01:15]** Looking back on 65 episodes and why this podcast succeeded when earlier attempts didn't.*

**[01:56]** Kimi K3, Buzz and autonomous OpenAI agents kick off the discussion.*

**[03:02]** Pete explains his AI news agent that reads Twitter, builds a knowledge graph and delivers one high-signal story every hour.*

**[05:41]** Andy discovers that disconnecting from social media makes him realise just how small the AI bubble really is.*

**[08:12]** Why most people still use AI very differently from the people discussing it online every day.*

**[10:05]** Remembering when Facebook was actually social—and how modern feeds lost that magic.*

**[13:06]** Weekly AI summaries, work recaps and letting agents consume information for you.*

**[15:51]** AI coding, deep work and ADHD: why agent development can feel productive yet strangely unrewarding.*

**[20:16]** Pete introduces Wingman, Autopilot, Tower and Flight Deck—the foundations of their agent operating system.*

**[29:46]** Buzz vs Flight Deck: where the ideas overlap and why they're building something different.*

**[36:19]** Socratic dialogue beats one-shot prompting for designing software with AI.*

**[46:37]** Pipelines combine deterministic code with AI agents to automate repeatable business work.*

**[48:36]** "Most stuff you value the consistency over the intelligence." Why business processes matter.*

**[55:52]** Wingman isn't the product—it's the infrastructure that lets businesses become bespoke software.*

**[01:01:49]** The opportunity isn't replacing spreadsheets—it's replacing the fragile systems businesses already rely on.*

**[01:05:16]** The podcast signs off with one final idea: deploying "intelligence snacks" throughout a business.

**Friends of the Pod:** Gigi (*No Solutions*), Steve Lee (*Presidio / Bitcoin Jam*), Matt Odell (*Citadel Dispatch*), Peter Steinberger (*Birdclaw / OpenClaw*), Rick, Lara

**Concepts Introduced:** Intelligence Snacks (small deployments of AI decision-making), Flight Deck, Tower, Autopilot, Business as Software

**Quote:** *"The business is the software eventually. Wingman isn't the product, it's the foundation that lets us codify how your business actually works. The real value isn't in the agents. It's in building the business itself into software."*

DPC went from AI skeptic to full "AI psychosis mode" in three months. The conversation unpacks his journey from dismissing autocomplete to building his own Rust-based coding harness called Tau, designed with Unix philosophy — composability, standard input/output, event-based architecture, and the ability to run shell extensions over SSH to different machines.

An extended woodworking metaphor runs throughout: AI is power tools. You can be a Japanese craftsman with hand planes making beautiful YouTube content, but most work will get done with routers and circular saws. The saw stop and jigs becomes the central safety question, how do you build safely without ripping your arms off?

**Key Moments:**

- [01:09] "I used to be very skeptical of AI. I'm still skeptical of AI, but I'm now in my AI psychosis mode."

- [03:16] "I decided I need to write my own harness. Pi was the gateway for me to understand I could control this stuff."

- [06:15] "My whole worldview is infused with Unix. Composability is one of the guiding principles."

- [10:33] The woodworking metaphor: "You could be a Japanese craftsman with hand tools and be big on YouTube, but most stuff's probably gonna be done with power tools."

- [13:42] "The moment you start doing woodworking, you watch all the videos about how to operate a router because it can chop off your fingers in milliseconds."

- [16:36] "The moment I started playing with AI, I immediately started thinking how do we isolate the development environments?"

- [22:34] 300,000 lines of code in Tau vs 200,000 at Fedimint after four years with multiple people

- [24:43] "Vibe coding is great for personal use. If it has many bugs in code paths you never hit, do they really exist?"

- [25:46] "I think a lot of people will build their own little empires."

- [28:53] The logistics guy: "He built a whole vibe-coded infrastructure for his business. He's making millions on it."

- [32:50] "Boutique SaaS — like SaaS for five businesses or ten businesses or even one."

- [36:40] "I'm somewhat convinced now you can't just have the software write itself. It's not a hard problem, it's an impossible problem."

- [38:38] "When you create software, probably more important than the software itself is the mental model of the problem."

- [43:18] "The architecture is the most important aspect of software. The code you can always fix."

- [47:16] The OKR trap: "I understand you, but the OKRs say something different, and we cannot afford nine months of seven people."

- [55:31] "It feels like you just got a team under you with almost infinite patience and always available."

- [58:18] "It's like a kid that's never gonna grow up, an employee you wanna senior up and pass wisdom to, but you know they'll never get it."

- [1:04:21] "I was getting bored of software engineering and then suddenly this new thing drops that is nothing like anything before and its all new again"

**Friends of the Pod:** Mario (Pi harness creator), Perth Wood School (safety guards discourse), the anonymous logistics guy at the Bitcoin conference

**Concepts Introduced:** Tau harness (Unix-style, Rust-based), saw stop as AI safety metaphor, boutique SaaS, slop empires, model character matching (like matching work to employee personalities)

**Quote:** "Software engineering is more of a craft than engineering, than science. AI is very much like power tools — they're dangerous, they can do a lot for you, but the security is very tricky. The moment you start doing woodworking, you watch all the videos about how to operate a router because it can chop off your fingers in milliseconds."

The conversation explores why regional SMEs face a different AI adoption gap than metro businesses.

Why AI enables his advisors to get out of the office and sit at kitchen tables having real conversations.

The paradox: more AI means hiring more humans, because human judgment and attention don't scale with processing power. The conversation goes deep on corporate extraction vs SME capitalism, the "tidal wave going out" before the tsunami hits regional Australia.

**Key Moments:**

- [03:50] "In regional Australia, finding work's not the issue. The challenge is how do you engage the right people."

- [06:05] The tsunami metaphor: "The tide's going out and everyone's walking on the beach going, look at the seashells"

- [12:59] "For us as a society, we need to decentralize leadership"

- [17:05] X Plan dominance: "$110 million of $200 million in fintech revenue. One login can cost $12,000 per year."

- [22:24] "There's not many things in your business more valuable than your data now"

- [29:11] The Hispanic Trump voter question from Harvard: "Why would somebody in regional Australia vote for a female version of Donald Trump?"

- [37:17] "Your core value generating part of your business now sits in Anthropic's data center and the guy that runs it is telling you he will take your job"

- [42:05] "All I've seen in our business is we need more humans that can use AI"

- [51:10] "I'd rather be wrong and human than right and AI"

- [53:37] Kane's constraint theory: "Shoot a bullet before you get the bazooka out"

- [1:00:03] NVIDIA stat: "Same size as the top 324 Australian companies combined by market cap"

- [1:02:28] "I wouldn't take the pill. Why would I want to cut the heads of ten of my favorite work brothers and sisters?"

**Friends of the Pod:** Kane (guest), Gabe (Adapt/Lumia connection), Bill Withers (SME capitalism conversations), Daniel (Kane's tech mate), Professor Boris Gruwski (Harvard), Professor Rari (the Hispanic Trump voter question)

Yo joins from Sovereign Engineering to talk about how AI has transformed freedom tech development since SEC-04, when Paul demoed a wallet vibed into existence in 30 minutes and blew everyone's minds.

The conversation covers the euphoria wearing off, the slot machine addiction problem, why LLMs are "extremely confident, extremely good at English usage, but extremely dumb," and the coming IPO exit liquidity dump. Uber burned their entire annual AI budget in four months and shipped nothing. Claude Code still has that terminal flicker bug from week two. Coding is not solved. But used as tools within their constraints? These things are genuinely great. FIPS, the peer-to-peer internet architecture routing to Nostr identities instead of IPs, was largely written by Claude, but every single line got human review. That's the model: Claude as a team member with specific jobs, not the entire dev team. YOLO++ kicks off July 20th.

**Key Moments:**

- [02:47] Paul's 30-minute wallet demo at SECO 4: "Everybody was blown away by this capability"

- [04:14] "There's no hope if I am to come back to programming by hand. I'm not going to produce anything."

- [07:11] Cryptography and AI: "It really, really fucks up. It always wants to do its own cryptography."

- [08:13] "If you are a reasonably good programmer, you would always trust yourself more than you would trust the AI"

- [10:59] The study where LLMs overtook a codebase: "Absolutely unrecognizable, unmaintainable without the LLM"

- [12:52] "As far as Dario is concerned, coding is largely solved. I'm not a good programmer, but I don't see it solving my problems at all."

- [23:25] "The euphoria is wearing off. The slot machine is just too addictive. People are tired."

- [25:11] IPO bubble discussion: "It should be known by now that it's exit liquidity"

- [27:35] 401k rule changes: waiting periods reduced from months to 15 days to force passive buying

- [31:01] SECO 5 prediction: specialized local models for Git, bash, commits, PRs—"We didn't really get those, did we?"

- [38:02] Uber's AI disaster: "Their year's budget, they finished in four months. What did we actually ship? Nothing."

- [42:17] Claude Code terminal flicker: "After coding's been solved for nine months, they still haven't fixed it"

- [48:57] FIPS written with Claude: "Every single line of code is getting reviewed. Claude is one of the members of the development team."

- [55:21] FIPS architecture: Nostr identities convert to IPv6, devices identify and connect peer-to-peer

- [59:10] Cross-pollination at cohorts: "Everybody on Wednesday is talking about their AI setup, their workflow"

**Friends of the Pod:** Yo (guest), Paul (OG wallet demo), Gigi (pipeline workflow, phone recordings), Marty Malmi (laptop wanderer, Nostr VPN), Mitchell Hashimoto (spending money to figure out where AI works), Lighthazard (AI for examples, not library code), Jonathan (FIPS creator), Aryan (FIPS collaborator), PrimaGene (Claude Code flicker video), James Checkmatey (bubble commentary)

**Projects Mentioned:** FIPS (peer-to-peer Nostr-addressed internet), Tollgate, Wingman (now on v4 with declarative workflow system), Zap Store

**Quote:** "These LLMs are extremely confident, extremely good at English usage, but extremely dumb. They have a lot of information, but they don't know how to weigh that information, how to use that information, and what are the consequences of taking certain actions."

# The Good Stuff, Episode 61: Loop Engineering

Boris (Claude Code creator) and Steinberger both tweeted this week: "I don't prompt anymore, I just build loops that prompt for me." Pete's response: they've discovered being a team leader.

Loop engineering is organizational design with new hype marketing terms—triggers, processes, business rules, the stuff we've been doing for a thousand years. The conversation explores where humans actually fit in these loops (spoiler: you can't be hands-off), the coin flip problem of compounding agent decisions, and why running agents for a day with no human interference produces drift toward suboptimal forks. Vision, values, and principles aren't just for humans-they're how you scale decision-making when you can't review every choice.

Also covered: the bubble phase of AI where we're shitting money into the pool instead of making things efficient, Apple's WWDC local LLM play (MDX protocol, neural accelerators), and why the $50/month product gap is so hard to close.

**Key Moments:**

- [01:01] Boris tweet: "I don't prompt anymore. I just build loops that prompt for me."

- [01:26] "They've discovered being a team leader. Basically."

- [06:07] "Business intelligence layer—it sounds sexy. But for 20 years there's always been a project in every business that was 'we should just have one big database.'"

- [07:50] "It's discovering that instead of coding, maybe we should be thinking about coding methodologies"

- [09:04] "What's interesting about agents is they move quick enough that you can put more iterations in. So it does feel loopy."

- [12:20] The coin flip problem: "If you've got 50 sub-agents, each making a judgment call, that's another coin toss. Your likelihood of getting heads nine times out of ten has just been obliterated."

- [14:21] "OpenAI pays his bills. At that point it doesn't matter what you're spending."

- [16:37] Vision, values, principles: "Here's how we make decisions. Every decision, you need to be thinking about these."

- [18:22] "Your AI does not have the same intuition for where you're going that you do. When you sit with it, it steers right. When you don't, it doesn't."

- [24:03] Shopify's five-person team ideal: "It's just because you can have tons of people doesn't mean you should"

- [29:29] "It's not all about the volume. It's always been about differentiation and taste and specific useful output."

- [40:47] Mythos/Fable: "It told me it was going to burn credits at twice the rate"

- [51:24] The bubble: "We're in this bubbly cash grab. We haven't done any of the interesting engineering work about making this efficient instead of making it bigger."

- [55:44] "I don't need more intelligence. I need better application to the problem."

**Friends of the Pod:** DeadmanOz, Gav, Mark, Peter Levels, Toby from Shopify.

**New Terms Coined:** The Right Porridge (Goldilocks context management), Loop Engineering (what we're calling it for now)

**Quote:** "We seem to be speed-running, the realization that the all of the organizational management infrastructure we've developed over the last thousand years was useful. It's there for a reason. We've not arrived there accidentally, we're now just reinventing it for agents with new names."

Pete and Andy are joined again by Deadman Oz / AM to talk through the latest AI trends, starting with the argument over whether progress is really coming from better models or better harnesses.

They cover Claude Code workflows, agent orchestration, context management, reasoning models, open source local models, inference economics, and the business strategies of OpenAI, Anthropic, Mistral, and the Chinese labs.

They close by taking the same question into robotics: where automation pays off, whether humanoid robots make sense, and why space robots feel less unsettling than house robots.

## Chapters and Themes

- `00:00-02:18` Deadman returns and the group sets up the core argument: is AI progress mostly model capability, or the harness around the model?

- `02:18-05:50` Pete explains why he is bullish on pipelines and declarative workflows, while Anthony describes using workflows for Bitcoin-adjacent research.

- `05:50-10:44` Workflows, agents, and context windows: where they help, where they fail, and why parallel work can make humans lose the thread.

- `10:44-14:59` Pete argues that agent-driven control loops introduce randomness, so humans still need to keep agents on the motorway.

- `14:59-21:55` Anthony makes the case for stronger reasoning models, using recent AI-assisted maths results as an example of cross-field insight.

- `21:55-24:15` Pete questions whether models can self-correct without a clear internal goal, especially when the human cannot fully specify the destination.

- `24:15-32:14` OpenAI, Anthropic, product focus, influencer hype, doom messaging, enterprise lock-in, and regulatory self-interest.

- `32:14-37:54` Local LLMs, Chinese open models, Mistral's enterprise focus, and whether frontier models commoditise too quickly to justify the spend.

- `37:54-40:58` Auto-research, self-improving training loops, and whether ever-higher intelligence might become too abstract for ordinary tasks.

- `40:58-49:47` Robotics, ports, unions, mining, household chores, and whether automation is blocked more by politics than technology.

- `49:47-56:05` Humanoid robots versus specific tools, open robot ecosystems, space robots, and the more optimistic version of an AI future.

## Key Takeaways

- The harness, workflow, and orchestration layer shape what models can actually do.

- Parallel agents are useful for breadth, but they create attention and coordination problems.

- Reasoning may help agents stay on track, but goal definition and judgment remain hard to outsource.

- Anthropic may be product-focused, but hype, opacity, and regulatory self-interest still make trust difficult.

- Open source and local models are increasingly good enough when the surrounding system is well designed.

- Frontier-model economics are awkward: inference can be profitable, while training races make breakthroughs short-lived.

- Robotics may deliver clearer value in infrastructure, mining, manufacturing, and space than in humanoid house helpers.

## Notable Lines

- "I think it's more harness than model."

- "It worked with the old model, but you needed to think about how you actually work."

- "The computer is the computer and it can talk."

- "If you don't have to do any of the upfront training, the margins on the inference is insane."

- "I probably don't need Mythos to continually say what's the git command for..."

- "Space robot seems fine."

Pete and Andy unpack the recent surge in anti-AI sentiment and argue that the anger is often aimed at the wrong target. They trace the backlash through graduate job anxiety, elite AI messaging, economic stagnation, and broader distrust in institutions, then pivot into a more grounded discussion about what AI can and cannot realistically replace. Along the way they talk about cost curves, productivity myths, design, taste, originality, and why the enduring value in human work may sit less in task execution and more in judgment, experience, and problem selection.

## Chapters and Themes

- Opening on the visible rise in AI hostility, from graduation-ceremony boos to a broader sense that AI has become a cultural punching bag.

- Why the backlash feels understandable: people hear wealthy AI leaders talking about job losses while younger workers and graduates already feel economically cornered.

- The difference between hating AI itself and hating the incentives, theft, corruption, and concentrated power associated with the companies leading the current wave.

- Why "human-made" branding may grow, but why avoiding AI entirely will be hard when customers still buy on price, speed, and output.

- AI job-loss narratives versus practical constraints: compute, energy, token costs, and whether replacing real workers is actually economical at scale.

- The gap between claims of 10x or 100x productivity and the lack of obvious real-world evidence in company performance and output.

- Why some work may become cheaper and more automated, while the scarce layer shifts toward framing problems, exercising judgment, and understanding real user needs.

- A closing reflection on design, taste, originality, and how creative work changes once the old technical bottlenecks become commoditized.

## Key Takeaways

- Much of the current anger around AI is really anger about economic insecurity and untrustworthy institutions.

- Anti-AI sentiment may be emotionally understandable without being strategically useful.

- The economics of AI replacement are still far less obvious than the rhetoric suggests.

- Cheap task execution does not remove the need for judgment, taste, and lived experience.

- The enduring moat in creative and technical work may be knowing what matters, not just producing outputs faster.

- Local, controllable, human-directed AI remains a healthier direction than total dependence on centralized providers.

## Notable Lines

- "Go support your local homegrown organic AI project."

- "Everybody else's job is really easy to automate, right? Except mine."

- "Good developers were never about the code."

Pete and Andy are joined by Shawn Yeager to talk about what AI does to professional services firms once the billable hour starts collapsing.

The core argument is that firms in law, accounting, marketing, and similar sectors will need to move away from selling time and toward selling trust, judgment, systems, and outcomes.

As AI compresses document work and routine delivery, the real value shifts to client relationships, productized services, workflow design, and the ability to package expertise in ways that scale.

## Chapters and Themes

- `00:00-05:35` Shawn introduces his thesis: AI is killing the billable hour, forcing professional services firms to rethink what they sell after time-based work gets compressed.

- `05:35-10:29` The discussion moves to outcomes versus hours. Shawn argues the real value in legal and other services has always been trust, judgment, and knowing what to do when things go wrong.

- `10:29-14:14` Pete and Shawn explore how AI changes delivery by making it easier for non-specialists to generate quality work when the right systems, files, and scaffolding exist.

- `14:14-18:07` They discuss the rise of the product-minded generalist. Shawn argues everyone increasingly needs product thinking: customer empathy, technical awareness, and the ability to shape deliverables into something sellable.

- `18:07-20:19` Andy raises the firm-structure question: if AI flattens the old professional services pyramid, what happens to the layers of junior talent and middle management?

- `20:19-22:52` Shawn and Andy dig into the talent-pipeline problem. If firms become flatter and more system-driven, they may struggle to develop the next generation of experienced practitioners.

## Key Takeaways

- AI is compressing the work that used to justify the billable hour.

- Professional services firms will need to sell trust, judgment, and outcomes rather than time.

- Product thinking is becoming a core skill, even in traditional service businesses.

- Smaller, high-agency teams can now do work that previously required much larger firms.

- The long-term challenge may be talent development, especially in heavily regulated professions.

## Notable Lines

- “AI killed the billable hour.”

- “Ultimately what they sell is trust.”

- “What would you do if you had five or ten more employees?”

- “I think everyone’s got to be a product manager now.”