AI #173: AI Pauses

Abstraction: Weekly AI roundup during the Fable takedown

Key points:

• Framed around the White House shutting down Fable 5 / Mythos 5 via export controls; the "jailbreak" was literally "fix this code," which Opus 4.8 and GPT-5.5 also do; the demand to "fix" it is impossible without removing coding ability.
• Dario Amodei's essay "Policy On The AI Exponential" proposes FAA-style mandatory third-party testing above a compute threshold for cyber/bio/loss-of-control/automated-R&D risks; Nate Soares criticizes it as soft-pedaling with loss-of-control buried in one line.
• Anthropic released two policy frameworks (economic + advanced AI); developer threshold 10^25 FLOP AND ($500M revenue or $1B R&D); Peter Wildeford notes they ignore internal development/deployment, a main failure mode.
• DeepMind paper (Legg, Hutter, Dafoe, Gabriel, Leibo) lays out four AGI-to-superintelligence paths: scaling, paradigm shifts, recursive self-improvement, multi-agent collectives; assumes alignment away.
• Scott Alexander's AI opinions: 25% AGI by 2027, 50% by 2034; 20% chance first past-point-of-no-return AI wants to eliminate humanity given current safety effort; 66% simulation-related (Zvi's is much lower). Frontier LLMs beat specialized clinical AI (OpenEvidence/UpToDate) — the bitter lesson.

Connections: Zvi Mowshowitz · Anthropic · Dario Amodei · Deepmind · AI Policy · Superintelligence · Recursive Self Improvement

Source: https://thezvi.substack.com/p/ai-173-ai-pauses