davidad: R1-Zero published the recipe for 'data criticality' — literally the singularity
davidad · x · 2026-09-13
AI safety researcher davidad argues that before reasoning models, data was the binding constraint on large training runs — labs were genuinely running out. R1-Zero published the recipe for "data criticality": a model generating its successor's training data. In his words, that's "literally singularity" — the recursive self-improvement loop closed on data.
He adds that other criticalities remain: algorithms (happening now), hardware, energy, manufacturing, construction. Growth isn't superexponential but piecewise-exponential, with an inflection point each time an RSI loop closes — and once the first recipe is public, there's no going back.
Related event: davidad: R1-Zero's Data-Criticality Recipe Is a Literal Singularity(2 posts)→
More from AGI Musings
- OpenAI alignment researcher: taming superintelligence is natural science, not pacing — jachiam0 · 2026-09-13
- METR brings in Redwood staff for independent investigation of Anthropic alignment incidents — Jsevillamol · 2026-09-13
- What's a rogue AI agent's comparative advantage against hyperscalers printing cheaper tokens? — EigenGender · 2026-09-13
- Dario Amodei's new essay calls to pace the frontier; Anthropic opens systems to third-party evaluators — EvanHub · 2026-09-13
- MIT publishes official AI and Education report with eight guiding principles — alex_verem · 2026-09-13
- MIT report: 46% of undergrads use LLMs daily, 90% worry about overreliance — alex_verem · 2026-09-13