AI Strategy & Transformation
Anthropic publishes its recursive self-improvement thesis: "When AI builds itself" (June 5, 2026)
Anthropic Institute released a paper this morning outlining three trajectories: stall, humans-in-charge, and full recursive self-improvement (RSI) in which "human labor stops being competitive." The supporting Anthropic data: more than 80% of code merged inside the company in May was Claude-authored; engineers ship roughly 8x more code per day than in 2024; Claude's success rate on open-ended engineering jumped 50 points to 76% in six months; Claude Mythos sped up internal training code about 52x. Co-author Jack Clark frames the question as governance, not capability, and says Anthropic would pause if its peers did the same.
So what
This is the first time a frontier lab has put its own RSI math on the record while still selling the underlying model. The paper gives growth-company boards usable language to challenge AI capex assumptions and to ask vendors what their own self-improvement posture looks like. Put a written governance memo in front of your audit committee this quarter that names the three Anthropic scenarios and your company's position on each. Pair it with the question every CFO should now be asking: how much of next year's engineering plan depends on the velocity numbers Anthropic just published.
⚠ Hype check: Anthropic published this warning the same week it advanced an IPO process and operated off a fresh major capital round. The company is asking peers to slow down while accelerating its own pace, and most of the productivity numbers come from Anthropic's own systems without independent reproduction. Treat the trajectory framework as useful; treat the velocity figures as vendor-reported until external benchmarks confirm.
Confidence: Informed Perspective — Anthropic's analysis of its own data, not independently reproduced.
Go deeper Anthropic Institute report · Curated: The Deep View
Organizational Design
AI velocity dies inside bureaucracy, not in the tooling (June 5, 2026)
A guest post on Lenny's Newsletter today argues the velocity gap between AI-native and legacy companies has nothing to do with model access. Tools amplify high-agency teams; in low-trust, low-context organizations they amplify nothing. The author's line: "Agents don't have agency. High-agency employees do." Two practical moves: stand up small cross-functional "wedge pods" with end-to-end authority, and assign every internal agent a human "parent" responsible for accuracy and currency.
So what
For growth-company CEOs, this reframes the question from "are we using the right models" to "do our best people have the access and decision rights to act on what the models tell them." Audit where AI output is gated by approval queues that did not exist for an analogous human action. Name agent owners explicitly; remove the bureaucratic excuse that "no one has the context." The phrase to use in your next staff meeting: more agency, not more agents.
Confidence: Informed Perspective — opinion piece with credible operator examples; the prescriptions are reproducible inside most growth companies.
Go deeper Lenny's Newsletter post · Curated: Lenny's Newsletter
