AI #173: AI Pauses
Short summary
The White House slapped export controls on Anthropic’s Claude Fable 5 and Mythos 5 after a code-based jailbreak showed they could identify and fix security flaws. Regulators demand a “fix” that experts say can’t distinguish defensive from offensive coding skills. The post also surveys recent AI news—from new full-body medical scanners to benchmark upgrades and policy proposals.
- This is satire/fiction dressed as a news roundup—the models, companies (Claude Fable 5, Mythos 5, MidJourney Medical), and events (export controls, the pause) appear to be invented rather than real.
- The core absurdist point: regulators demanded a "fix" for an AI's ability to find and patch security flaws, even though experts say defensive and offensive coding skill can't be separated—so the "fix" is technically incoherent.
- Markets are already treating the situation as a betting market, pricing 50-50 odds the export-control pause gets resolved by July 1.
- The piece frames this fictional crisis against a backdrop of real-seeming AI industry noise (new benchmarks, funding rounds like DeepSeek's $7.5B raise at a $50B valuation, competing models like GLM-5.2 and Grok 4.3) to satirize how policy panic outpaces actual technical understanding.
Full summary
At 5:23 pm last Friday the White House slapped export controls on Anthropic’s flagship Claude Fable 5 and Mythos 5 models. The move follows a so-called “jailbreak” demo—Fable was asked to “fix this code,” identified security flaws and suggested patches, which in turn exposes those same flaws for exploitation. Rather than disabling the feature, Anthropic’s co-founder Dario Amodei argued it posed no real risk. That response didn’t fly with regulators. Since then Anthropic has been in damage-control mode, flying executives to Washington, negotiating a path to “fix” the jailbreak—something experts say is technically impossible without crippling Fable’s coding skill. As of day seven, the pause remains in effect, and markets are pricing a 50-50 chance of resolution by July 1.
Beyond the pause, the AI world is buzzing with new benchmarks, hardware and policy moves. MidJourney Medical unveiled a full-body scanner that uses no radiation, promises high resolution and low marginal cost, slated for next-year trials. Anthropic floated several policy frameworks on developer obligations, resilience measures and economic redistribution—ideas that feel quaint against an export-control backdrop. On the tech front, OpenAI’s LifeSciBench rolls out 750 expert-authored biotech tasks, while the EvalEval Coalition aims to catalog and verify evaluation metrics across models. Benchmarks like Opus Magnum challenge models to solve Zachtronics puzzles: Claude Fable 5 topped GPT-5.5 and GLM-5.2, but it’s offline. And across corporate labs, GLM-5.2, Grok 4.3, Gemini 3.5 Flash and others jockey on speed, cost and ethical consistency tests. Meanwhile, companies like DeepSeek raise billions—$7.5 B at a $50 B valuation—planting bets on cheap, fast AI agents.