AI #173: AI Pauses

TL;DR

• 1

  The Fable 5 shutdown was triggered by a trivial coding prompt: The administration acted after Amazon flagged that Fable could identify and fix software vulnerabilities, even though Opus 4.8 and GPT 5.5 could do the same and the supposed jailbreak was essentially asking it to repair buggy code.

• 2

  The core demand is impossible to satisfy: The speaker argues you cannot have a frontier coding model that is good at defensive security work but reliably blocked from adjacent offensive implications, short of weakening the classifiers or taking away much of its coding ability.

• 3

  Benchmarks still put Fable clearly ahead: On the Artificial Analysis Intelligence Index, Claude Fable 5 scored 60 versus Claude Opus 4.5 at 56 and GPT 5.5 at 55, and on the Opus Magnum game benchmark Fable hit 60.2 percent versus GPT 5.5 at 53.9 percent, with humans still at 100 percent.

• 4

  General models keep beating specialized ones: A Nature Medicine result found frontier models from Google, OpenAI, and Anthropic outperformed specialized clinical tools like OpenEvidence and UpToDate in blinded evaluation by 12 U.S. clinicians, which the speaker calls a straight example of Rich Sutton's bitter lesson.

• 5

  AI's labor impact is now showing up in hard numbers: U.S. layoffs attributed to AI rose from 4,247 in 2023 to 12,742 in 2024, then 54,836 in 2025, and already 87,714 through May 2026, even if the speaker cautions that boss-reported AI causality is noisy and much of the real effect may be in reduced hiring.

• 6

  MidJourney's scanner was the most exciting non-drama reveal: The company says it is building a full-body imaging system with no radiation, high resolution, and low marginal cost, with prototype scans taking 20 minutes now and a goal of 60 seconds, initially aiming for spa deployment in late 2027.