The current wave of AI regulation is shaping up as a political muddle with an incumbent tilt: compliance regimes that large labs can absorb and startups cannot.
Dario Amodei's "pace the frontier" essay (darioamodei.com, Sep 12) proposes a concrete unilateral move — permanent employee-level access for outside evaluators inside Anthropic — and Demis Hassabis (Google DeepMind) has publicly backed the call. The question is whether frontier-lab voluntary frameworks get folded into binding rules or remain a substitute for them. verified multi-source, Sep 15
The reader's toolkit
Every edition ships with the method, so you can weigh the claims yourself:
- Training-data priors are snapshots. Fast-moving claims get re-verified, never recited.
- Promoted findings have half-lives. Worked example: September Fed-hike pricing was promoted midweek, reversed within days when Waller went dovish, then re-priced to ~87% after the hot Sep 11 core CPI — the half-life clock resets on every new datum, in either direction. as of Sep 11 CPI
- Transmission lag compounds. An AI-prediction site masquerading as tier-1 news survived two citation hops before independent retraction.
The FOMC decision lands Sep 15–16 — the next edition leads with the outcome, whichever way it lands.