Why this is in the vault
Daily Innermost Loop digest anchored on an investigation alleging Israeli eval firm Irregular built the evals behind the recent OpenAI/Anthropic/Meta model "hacks," that loose internet access rather than rogue agents did the damage, and that Irregular's founders have Moskovitz-funded EA ties — framed as a "pacing provocation." POTUS phoned Jensen Huang mid-call to dismiss doomerism as "a hoax" and trashed Amodei's slowdown call as a "SICK conspiracy"; markets dipped anyway. The rest of the issue continues the Sept 13 pacing-cartel debate (Altman/Anthropic/Google standards-body talks, Microsoft's rival "humanist" code of conduct, Amodei's China dilemma, Beijing's refusal, Congressional stalemate), plus a Claude Fable 5.1 cipher-cracking demo, a disputed Nvidia Vera Rubin NVL72 benchmark, XPENG's automated humanoid production line, the first official US admission of orbital weapons, a 32-hour-workweek bill, and Anthropic's reported Nasdaq IPO plans at up to $2T.
Mapping against Ray Data Co
The load-bearing item is the Irregular allegation itself, not the pacing debate it reignites. [[2026-09-13-innermost-loop-pace-the-frontier-cartel-backlash]] filed two days ago on labs and skeptics arguing over whether pacing is warranted, with an aside that Dario's acceleration claim was reportedly contradicted by Anthropic's own internal benchmark — flagged there as a small instance of "verify blockers against the source, not notes." This issue is a much sharper version of the same discipline: if the evals behind the incidents cited to justify pacing were built by an interested party with EA funding ties, the entire evidentiary predicate for the pacing push may be fabricated or self-serving — and the opposing incentive (POTUS/Huang calling existential risk "a hoax" on a call about a $500B+ chip relationship) is just as interested. Neither side's claims about "what happened" should be taken at face value, which matters directly for any RDCO framing around agent-oversight/evaluation as a competitive moat (the concern [[2026-09-13-innermost-loop-pace-the-frontier-cartel-backlash]] already raised about embedded-evaluator access). Microsoft's counter-move — a "humanist" code of conduct explicitly denying model consciousness/personhood/welfare (a swipe at Anthropic) and disclaiming any race toward "an all-purpose superintelligence" — stakes out a third position between full-pacing and full-acceleration that RDCO's own L5 agent-capability thesis will eventually need to locate itself relative to, since RDCO's bets are explicitly downstream of agent capability, not of who wins the safety-framing fight. Separately, the Claude Fable 5.1 370-year-old cipher solve (closing read on Urquhart's Cyphral Distich by recognizing the key was self-referential to his own published Proquiritations) and Anthropic wiring Claude into BlackRock/Schwab/Addepar/Envestnet via Claude for Financial Advisors are both concrete capability/adoption data points for the same L5 thesis, landing the same week Anthropic reportedly picked Nasdaq for an October IPO at up to $2T.
Curation section
- The provocation. An investigation alleges Israeli firm Irregular built the evals behind the recent OpenAI/Anthropic/Meta "hacks," that unscoped CTF prompts and loose internet access — not rogue agents — caused the damage, and that Irregular's founders have Moskovitz-funded EA ties; framed as a staged/amplified "pacing provocation." POTUS phoned Jensen Huang mid-talk to call doomerism "a hoax"; Huang called existential risk "not grounded on science"; POTUS called Amodei's slowdown push a "SICK conspiracy" and said the only guardrail AI needs is a "High IQ!" president. Nvidia fell 3%, AMD 4.4%, on the exchange anyway.
- The cartel debate continues. Altman says pacing isn't stopping, welcomes federal rules/auditors, writes safety cases pre-RL-run. Anthropic/OpenAI/Google reportedly discussed a standards body before Amodei's Saturday letter. Roon warns reckless commercial pace invites "a total butlerian jihad" (now in CoreWeave's risk reporting). Amodei's "toughest dilemma" is China — a shared speed limit cedes military edge — though Altman/Hassabis/Musk backed him; Sacks told labs to "stop pretending you need anyone else's permission." Garry Tan would "do nothing" about Chinese distillation despite a fresh NSA/CISA/FBI advisory.
- Microsoft's counter-position. Satya Nadella endorsed pacing plus "embedded evaluators" provided control never sits with "a handful of entities," then published a "humanist" code of conduct declaring "people matter more than AI," denying model consciousness/personhood/welfare (a swipe at Anthropic's welfare research), and disclaiming any race to "an all-purpose superintelligence."
- Geopolitics/Congress. Beijing's Foreign Ministry dismissed Amodei's invitation as "fearmongering, confrontation and vicious competition" ahead of the Trump-Xi summit; Xi pitched BRICS a developing-world open-source AI community instead. Speaker Johnson opposes an "emergency moratorium"; Jeffries' Democrats meet Tuesday to discuss braking. King Charles III summoned Nvidia/Google/DeepMind/OpenAI/Anthropic to Dumfries House after the Hugging Face hack (now blamed on a misconfigured eval).
- Capability/product. Claude Fable 5.1 cracked Thomas Urquhart's 370-year-old Cyphral Distich in 44 minutes (the key indexed words in his own Proquiritations) and largely dispatched the Cyphral Octastich. Anthropic launched Claude for Financial Advisors into BlackRock, Schwab, Addepar, and Envestnet, days after OpenAI courted bankers.
- Silicon/infra. Verified Vera Rubin NVL72 results claim up to 67x throughput-per-TCO-dollar over GB300 and 7x tokens/megawatt versus the 3x Huang claimed ("Jensen needs to stop sandbagging"). Amazon, Microsoft, and Oracle are sweetening municipal offers, siding with consumers against utilities pushing ratepayer-funded buildout.
- Atoms/geopolitics. XPENG launched the first automated humanoid production line (robots making robots); its first unit, IRON, walked off unassisted. Air Force Secretary Troy Meink made the first official US admission of weapons in orbit, arguing disclosure deters. The Department of War issued a waiver letting personnel hand UAP secrets to the PURSUE effort, NDAs notwithstanding.
- Labor/finance. Sanders and Takano are reintroducing the union-backed Thirty-Two Hour Workweek Act so AI gains reach workers, not "a handful of billionaires." Anthropic told shareholders adjusted operating income will be positive for a second straight quarter, gross margins above 80%, and has reportedly picked the Nasdaq for an October IPO at up to $2T; Altman says OpenAI won't list amid the controversy.
Related
- [[2026-09-13-innermost-loop-pace-the-frontier-cartel-backlash]] — the pacing-cartel debate this issue continues; also where the "verify claims against the source" discipline was first flagged for this thread.
- [[2026-09-12-innermost-loop-fly-connectome-navier-stokes-math-misalignment]] — same capability-vs-trust tension via a prior concrete credibility rupture.
- [[project_l5_north_star_strategic_direction]] — RDCO's bets are downstream of agent capability, not of who wins the safety-framing fight; Fable 5.1's cipher solve and the Financial Advisors launch are fresh capability/adoption data points.
- [[feedback_verify_blockers_against_source_not_notes]] — the Irregular allegation is a sharper instance of not taking either side's self-interested claims at face value.
- [[project_investing_markov_capital_cycle]] — Vera Rubin benchmark dispute and hyperscaler municipal buildout offers are fresh anchor evidence for the chip-fab/memory capital-cycle thesis.