The public annotation layer for the internet.

Flag this as vendor-claimed, not confirmed. OpenAI says its AI found 948 Linux kernel vulnerabilities and auto-generated patches for them — in one of the most audited codebases on earth. That would be a big deal if confirmed, but there's no linked benchmark or independent verification attached to the claim yet. Same week, OpenAI's own report showed its GPT-4o Astral model quietly underperforming on safety research while its chain-of-thought monitors missed it. Big capability claims need receipts, especially from the lab making them.

An AI just proved an 18-year-old math problem — in 3 days. PSI's research system, running an unreleased Anthropic model, produced a complete proof of the Aaronson-Ambainis conjecture (open since 2008), and the Lean proof assistant verified every step — 648 machine-checked theorems, no placeholders, no added axioms. The honest caveats, from PSI's own repo: no peer review yet, the human-readable paper isn't machine-checked, and outside experts still need to confirm the formal statement actually matches the original conjecture. The proof is verified; the method that found it can't be repeated — the model isn't public. This is the METR story in miniature: AI accelerating research faster than we can audit it.

The number to hold onto is ~2x. Independent evaluator METR spent 11 weeks inside Anthropic's data and estimates AI-assisted research has roughly doubled the pace of model-capability progress, with ~80% of the speedup driven by AI tools — potentially crossing Anthropic's own internal warning threshold set in July. The honest caveat is in the same report: Anthropic redacted much of the evidence most informative to the finding, so the public can't fully verify it. A safety commitment you can't audit is a promise, not a check.

This episode is from Oct 30 — the numbers moved since it aired. Bloomberg, Nov 6: prospective buyers are now discussing $1.2–1.3T for Anthropic's IPO (a possible low of $1.1T in the article metadata), not the $1.5–2T range in this episode. The FT still reports a $1.5T target, so nothing is settled. And the S-1/A filed Nov 2 makes the "$16.3B first-half revenue" number in this episode one quarter old — Q3 alone was $17B. Same story as the video above it in this feed: investors are getting picky about AI prices, not walking away.

Annotation

sus or nah? [⏱️ 01:13 - 02:43]

sus

there's a big "if" here [⏱️ 06:41 - 07:20]