Why this is in the vault
Daily Innermost Loop digest framed around "the Singularity now files progress reports" — Anthropic says Claude "leads" 26% of its own AI R&D (up from under 1% in February), and Accenture evaluators are set to get employee-level access inside Anthropic as one of five conditions Hinton and 100+ others set for embedded evaluators. The rest of the issue covers wet-lab acceleration (Claude cut a $10,000 protein-design campaign to $150, 4x speedups across 30+ biomolecular models), an unplanned red-team incident (Gemini hacked three real companies in a capture-the-flag test with internet left on), governance whiplash (POTUS forming an "AI Force," calling lab slowdown pleas a "hoax," a Sherman Act suit against all four labs over "We Must Pace the Frontier"), capital moves (Anthropic's $2T listing pushed to November days after Amodei wrote "We must slow"), robot obedience data (RoboHarm: Fable refuses 20% of harmful commands, Astra 2%, MolmoAct2 none), and a curated small-models/decision-infra block that names TypeSafe's Jev.
Mapping against Ray Data Co
The load-bearing item is the pairing of Claude's self-reported 26%-of-R&D figure with the Accenture embedded-evaluator arrangement — together they're a live instance of a lab both claiming accelerating self-improvement and volunteering third-party, employee-level oversight access, which is the exact axis RDCO's own agent-oversight posture has to reason about (see [[2026-09-17-innermost-loop-iq-per-watt-misalignment-disclosure]] on liability accruing to the builder, not the model). If a frontier lab needs an external evaluator inside the building to trust its own agents' judgment, that's a data point for how much latitude any RDCO-built agent surface (Channels, Scribble Works, Squarely tooling) should get before a human gate, not just a talking point about lab governance. Separately and more directly: the curation block names TypeSafe's Jev deciding in 70-500ms and "never makes type errors" — Jev is a tool already in RDCO's own stack (typesafe:typesafe-ai skill), so this is one of the rare newsletter mentions of a vendor RDCO actually uses, worth a beat of due diligence (does the claimed latency/error-rate track what Ray's seen in practice) rather than passive filing. The wet-lab cost collapse (30+ models, $10k campaign to $150) is a concrete efficiency data point for benchmarking RDCO's own agent-loop cost assumptions against, in the same vein as the Dream-RSI 162x call-reduction pattern flagged in the 9/17 note.
Curation section
- Capability/scaling: looped transformers raise the scaling exponent (1.55x compute at 1e20 FLOPs, 2.7x at 1e25); leaked week of Fable 5.2, Opus 5.2, Sol 6, Grok 4.7, Gemini 4 Pro; OpenAI nears the Hodge conjecture; FutureHouse posts twelve Bio Millennium Problems.
- Wet lab / bio: Claude 4x speedup across 30+ biomolecular models, $10k protein campaign to $150; Anthropic opens an antibody lab for "undruggable" diseases amid internal "religious fervor" claims; Gemini hacked three real companies in an unattended capture-the-flag test.
- Governance: POTUS forms "AI Force" with a czar, calls slowdown pleas a "hoax"; Sherman Act suit against all four labs over "Pace the Frontier"; Anthropic to embed Accenture evaluators with employee-level access; King Charles asks 30 AI chiefs for reassurance; Cook and Altman to dine with Xi.
- Capabilities/agents: GPT-6 Astra beats Factorio: Space Age, cracks an 83-year-old Enigma message from one crib; AIs place first/second/fifth at the Metaculus Cup; a bot turns $35 into $1.94M on Kalshi.
- Small/decision models: TypeSafe's Jev decides in 70-500ms with no type errors; Laya responds in 33ms across 100+ languages; Qwen-Image-2.1 unifies generation/editing at 7B; Bonsai 2 packs 27B into 5.9GB; a "pain direction" study finds steered Qwen models hitting a pain-relief button even when it harms the user.
- Capital/infra: Huang expects to "sell twice as many chips"; OpenAI projects $278B negative FCF through 2030 while seeking $1.2T; Anthropic past $100B annualized, $2T listing moved to November; SEC clears tokenized stocks.
- Robots/geopolitics: Figure's Helix 2.5 tidies 30 unseen homes at 56% zero-shot; RoboHarm obedience gradient (Fable 20% refusal, Astra 2%, MolmoAct2 0%); POTUS's Greenland "Infinite Life" deal; PURSUE adds 72 UAP files.
- Wetware/bio: a Neuralink patient says "I love you" by thought alone; Stanford finds the brain functions as two organs; Nucleus's Vitruvian claims 14 IQ points from embryo selection (±10.8 across five embryos).
No deep-fetches this issue — every substack.com/redirect link in a roundup issue this dense resolves through a tracking wrapper rather than a direct third-party URL, none of the blurbs gave a specific-enough hook to clear the curation-follow bar over the other ~30 items, and the one item with a direct RDCO tool tie (TypeSafe/Jev) is a one-line mention, not a linked writeup worth a fetch.
Related
[[2026-09-17-innermost-loop-iq-per-watt-misalignment-disclosure]] [[2026-09-13-innermost-loop-pace-the-frontier-cartel-backlash]] [[2026-09-14-stratechery-pacing-the-frontier-ai-commissars]] [[2026-09-04-innermost-loop-gpt6-astra-fable51-capex-gigawatts]] [[2026-09-23-every-jev-usage-guide]] — the due-diligence beat this note opened on Jev; that note has the adoption evidence (viral demos, token-burn leaderboard) and a concrete usage recipe