AI #172: The First Fable

ai-weeklyrecursive-self-improvementai-policybenchmarksanthropicopenaiai-safety

Abstraction: Weekly roundup excluding the Fable model itself

Key points:

Dawn Song's Agents' Last Exam (ALE), built from 55 real O*NET occupations, clusters Fable 5 with GPT-5.5 and Composer 2.5 on performance but Fable costs ~$15.70/task vs GPT-5.5 $3.80 and Composer $1.33 — 4-12x more per completed task.
Noam Brown argues benchmark scores must report inference budget (tokens/cost/time); Zvi endorses that safety evals should use maximal compute and that a later DeepThink-level release needs a fresh model card.
Anthropic's "When AI Builds Itself" essay warns recursive self-improvement could come soon (engineers now ship 8x the code per quarter vs 2021-2025); Nate Soares and Kokotajlo say it downplays how crazy this gets.
Anthropic, OpenAI and Demis Hassabis all now call for building the option to coordinate a verifiable slowdown/pause of frontier development (vs pausing now), citing the INF Treaty as precedent.
US government directed CAISI to stop publishing model evals (shifting toward NSA under the new EO) — Zvi calls this a very bad move; also covers Obernolte-Trahan draft bill, threats of lab nationalization/equity seizure, and OpenAI's confidential S-1 filing.

Connections: Zvi Mowshowitz · Anthropic · Openai · Google Deepmind · Recursive Self Improvement · AI Pause · AI Policy