TL; DR: AI’s 1997 Internet Moment — Food for Agile Thought #547
Welcome to the 547th edition of the Food for Agile Thought newsletter, shared with 35,532 peers. This week, Benedict Evans tells Lenny Rachitsky that today marks AI’s 1997 Internet Moment and asks whether automation kills tasks or jobs, which Casey Newton’s guest, Kathryn Anne Edwards, treats as real but manageable, faulting unemployment insurance rather than fearing an idle underclass. Itamar Gilad warns that cheap AI coding tempts teams to build before validating, while Malcolm Spittler and Dylan Patel name the value GDP misses ‘Dark Output,’ and Joost Minnaar prefers autonomous-team networks over a single chain of command. Also, Marina Favaro and Jack Clark sketch the implications of recursive self-improvement of AI.
Next, Rich Mironov warns that AI ships 100x more code, yet attention and budgets don’t scale, so products that skip discovery rarely stick. METR, with Anthropic, Google, Meta, and OpenAI, finds that AI agents could plausibly go rogue but not robustly, while Teresa Torres notes that Cowork’s VM hosts the Mini Shai-Hulud worm without blocking it. Mike Fisher likens siloed teams to hand-copying the Diamond Sutra, and Jeff Gothelf redefines “done” as acceptable variance.
Lastly, Marc Abraham borrows venture capital’s ‘terminal value’ to help product managers judge whether a product merits more investment. Addy Osmani calls the opposite reflex ‘cognitive surrender,’ in which you stop thinking and accept the AI’s answer without checking it. Mark Graban shows confession works only when fixes follow, citing Burger King and Domino’s, while Tobi Lütke runs Shopify’s agent River in public Slack so everyone learns by watching. Finally, Anthropic open-sourced 11 role-specific Claude Cowork plugins for knowledge workers.
