Get The RSI Decode
The companion field guide to the video decoding Anthropic's own paper "When AI Builds Itself" (Marina Favaro & Jack Clark, 2026) — the honest, plain-English read on recursive self-improvement: AI systems that fully autonomously design, develop and train better versions of themselves. Inside: THE TRAJECTORY in four of Anthropic's own numbers (80%+ of code merged into Anthropic's codebase is now written by Claude as of May 2026; the typical engineer merges 8× as much code per day vs 2024; the length of task an AI can finish on its own is doubling roughly every 4 months — Opus 3 did 4-minute tasks in Mar 2024, Sonnet 3.7 ~1.5-hour tasks in 2025, Opus 4.6 12-hour tasks in 2026; benchmarks saturating in ~2 years). THE THREE FUTURES, plain English and which Anthropic believes (the trend stalls — least likely; compounding gains continue where a 100-person company operates like 100,000 — most likely; full recursive self-improvement where AI designs its own successors — alignment "deeply uncertain"). THE RECEIPTS with their source (AI agents recovered 97% of the weak→strong model gap on their own in 800 hours / ~$18K compute; 800 fixes cut a class of API errors 1,000× — est. 4 years of human work; an internal model "Mythos Preview" hit ~52× research speedup; a poll of 130 researchers reported a median 4× boost; Project Glasswing surfaced 10,000+ high/critical vulnerabilities). THE ASTERISKS THEY BURIED (nearly every figure is self-reported by the company with the most to gain; Claude's code only reached parity recently; Amdahl's Law — speed up one stage and the bottleneck just moves, so the human edge is judgment, not typing speed) plus Anthropic's unusual ask for a verifiable, coordinated slowdown (not a unilateral pause). And THE HUMAN-EDGE CHECKLIST drawn from the paper's own bottleneck analysis — choose the problem, kill dead ends fast, set direction, orchestrate agents, review critically, develop taste — because the winners won't be the fastest coders; they'll have the best judgment about where to point all of this.
Free. No spam. Unsubscribe anytime.