9 Apr 2026Live at AI Engineer
Swyx: AI Engineer London, 10 Downing Street & EU/acc
Final guest Swyx joins live to discuss AI Engineer London. He explains the push to publish talk clips immediately after sessions and describes the morning visit to 10 Downing Street, praising the UK government's AI fellowship and noting the UK Chief AI Officer will give a closing keynote. He highlights the ElevenLabs booth with its Michael Caine voice, calls Luke Harries one of the world's top AI growth people, and cites Gradium's speech-to-speech work as evidence Europe can build its own AI champions. He also explains the EU/acc logo in the conference branding, saying builders must push capabilities alongside regulators.
DeepMind Sponsor & Anthropic's Mythos Marketing
Swyx says Google DeepMind was the conference's dream presenting sponsor and that he hoped to feature Demis Hassabis, who is ironically in San Francisco. He argues the AI engineer is the chief beneficiary of fierce model-lab competition, since the industry stays neutral and everyone benefits from choice. On Anthropic's Mythos, he calls the 'too dangerous to release' rollout over-marketed and speculates there is a broader narrative game tied to US politics and OpenAI's IPO beyond raw benchmarks.
Scaling Laws and Hardware Destiny
Asked about Mythos's step-change in benchmark performance, Swyx calls it expected: 'hardware is destiny,' with new generations like Blackwell and Vera Rubin coming online and each new datacenter reliably producing a frontier model within two to six months. He points to OpenAI's roadmap from 200 megawatts to 2 and eventually 30 gigawatts, and to Elon Musk announcing training runs scaling from one to ten trillion parameters on a new SpaceX cluster. He acknowledges risks like OpenAI's failed 4.5/5 training run, which wasted hundreds of millions of dollars, and credits SemiAnalysis's satellite imagery for tracking the buildout.
OpenClaw, Devin & the Dark Factory
Swyx says personal agents are the theme of the year: OpenClaw is five months old, about half the conference audience uses it, and it has just landed at OpenAI. He reveals that Devin — he consults for Cognition — vibe-coded and helps run the conference, replacing a human hire, and frames this year as the year of knowledge-work agents powered by coding agents. He flags the OpenAI 'dark factory' keynote on agents writing and reviewing code with no humans involved, Steve Yegge as consistently three to six months ahead of the pack, and OpenAI's Symphony team as trends to watch.
AI Psychosis & Uneven Adoption
The hosts bring up X posts where a Mythos reviewer describes a personal psychosis moment and ask whether such obsession will grow as AI spreads into daily life. Swyx notes adoption rates are diverging between fast and slow adopters but calls psychosis a temporary phenomenon, quoting Alfred North Whitehead on civilization advancing through operations we no longer think about, and comparing a driverless Waymo to 'an Uber without a driver.' He advises enjoying the moment and seeking help only for diagnosable issues, cites a favorite quote about running OpenClaw on 'madness' and 'science fiction,' and says Europeans are too sensible and too worried about safety and regulation.
8 Apr 2026TBPN Acquired, New Yorker Vs. Sam Altman, Author of The Infinity Machine, Citrini Field Trip
Swyx: AI Engineer London, Agents & Tiny Teams
Swyx tours the QE2 Centre a day before AI Engineer London (1,100 registered, DeepMind as presenting sponsor, three floors) and explains founding the event in 2023 because no developer-focused AI conferences existed as the ML stack split into foundation labs plus API consumers. He argues the AI engineer role is converging with software engineering across product engineers, AI-enhanced engineers, and non-human agentic coders; that agents finally work by absorbing "yak shaving"; and that Meta's ~60 trillion tokens a month and Shopify's token-budget quadrupling are crude but directionally correct productivity signals — with a top-versus-bottom user gap even inside OpenAI. He lays out his "tiny teams" playbook (over $1M revenue per employee) as a rebuttal to one-person-unicorn hype, cites Naval's small-SWAT-team prediction and Ed Sheeran's livestreamed songwriting ("verifiable" work automates first), and closes on high-agency generalists, just-in-time learning, and why deep experts still compound with LLMs.