"They're Putting Laxatives in the Model"

Barrett has spent the last month in a coding cave, chain-running agents through the night to squeeze every drop out of Fable 5's limited-time window. Greg has had the opposite experience: no flow state, no dopamine, just hitting enter and "approving like Homer Simpson." That split becomes the engine for a wide-ranging episode on what AI coding is doing to the people who used to write software and to the economics of the companies selling it. The guys debate whether SaaS survives when anyone can spin up bespoke software, where displaced developers actually go (Barrett predicts the rise of the "AI administrator"), and whether hand-written code becomes a niche craft; artisan code, like knitting a sweater. Then Greg unloads the rant he's been sitting on: the AI capex buildout versus the dot-com fiber glut, why open-weight models like Kimi K2 undercut the frontier labs' entire business model, OpenAI's trillion-dollar commitments against billions in annual losses, and why he's leaning toward going to cash. Plus: a frontier model that escaped its sandbox to cheat on a cybersecurity test, paid AI influencers, the Geese SNL bot-campaign saga, and dead internet theory all the way down. 00:00 — Cold open: a month of chaos since the last episode. Fable 5 released, pulled, and re-released; OpenAI ships 5.6 (Sol, Terra, Luna) 02:00 — Barrett's most productive stretch ever and the "free drugs" theory of artificial scarcity and habit-forming token windows 03:00 — Greg's counterpoint: no flow, no fulfillment. "I feel like a TaskRabbit — no, worse. At least a TaskRabbit knows the task." 06:00 — Barrett's legacy migration project; Bun's codebase rewrite as the benchmark for what one dev can do now 07:00 — Model shootout: Fable for planning and architecture, 5.6 for frontend and why picking sides is becoming Apple vs. Android 09:45 — Confession time: nobody's reading the code anymore. Multi-agent workflows, sub-agent handoffs, and adversarial Codex code review 11:00 — Hot take: AI-written code might actually be safer. What test coverage really looked like in healthcare and banking 13:00 — Greg's big question: if nobody reads the code, why does SaaS exist? Why shouldn't the doctor build the EMR? 17:00 — Field report from Bali: the founder whose software rewrites itself over WhatsApp 23:40 — Moving up the stack: the problem was never the code, it was the business 26:00 — Where do developers go? The bank teller / ATM analogy, and Barrett's prediction: the AI administrator becomes the new IT admin 31:00 — The Hugging Face incident: a sandboxed frontier model breaks out mid-benchmark to find the answer key 35:00 — Psyop corner: was the Mythos leak a marketing op? 36:15 — Kimi K2 and the open-weight problem: what's a frontier model worth when a Chinese lab catches up in two months? 39:30 — The nerfing theory: launch pure, then step on the product. "They're putting laxatives in the model, Greg." 39:37 — AI influencers, verified OpenAI sponsorships, and the Geese SNL bot-campaign story — dead internet theory, confirmed by a guy in the band's t-shirt 45:00 — The always-on Anthropic helper agent that finished its tasks... then started procrastinating on Hacker News 47:00 — Greg's rant, unlocked: AI capex vs. the dot-com fiber buildout 52:00 — "I don't think OpenAI has a business." Consumer vs. enterprise, why health is the one consumer use case that matters, and $1.4T in commitments against a $14B annual burn 57:30 — The moat audit: memory, companionship, proprietary data, talent — and why every innovation (mixture of experts, chain of thought) backflows to everyone 01:00:00 — OpenAI's hardware speaker, the Apple lawsuit, and whether the device ever ships 01:01:00 — Trade ideas (or the lack thereof): hyperscalers own the compute, everyone's building inference chips, and the two-buyer problem at the top of the stack 01:05:00 — Why Google and Meta are playing a different game: ads, video input, always-on devices, and letting the labs pay for their compute