Show HN: Annotated – a public annotation layer for the web

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 about 80% of the speedup driven by AI tools — potentially crossing Anthropic's own internal warning threshold set in July. METR's report notes Anthropic redacted much of the evidence most informative to the finding, so the public cannot fully verify it. The same roundup cites OpenAI's claim that its AI found 948 Linux kernel vulnerabilities with auto-generated patches, unverified and without a linked benchmark, and PSI's research system running an unreleased Anthropic model to produce a Lean-verified proof of the Aaronson-Ambainis conjecture (open since 2008) via 648 machine-checked theorems, pending peer review.

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. “OpenAI CEO Sam Altman believes people owning RAM is wasteful, and that it belongs in data centers instead. "From a global allocation perspective, actually, it’s quite terrible to have everyone have a bunch of the world’s limited supply of DRAM sitting in their own computer."” Highlight text on any webpage to drop your thoughts. Your notes become part of the public web—anyone else with the extension will instantly see your yellow highlights right where you left them. Every annotation is permanently tied to its source URL. When someone else visits a page you've annotated, they'll see your highlights, read your insights, and can reply directly in their browser to join the conversation.