The AI Brief

Three people familiar with the plans told Yahoo Finance on October 21 that Anthropic would publicly file its S-1 as early as Friday, October 23. The SEC's 15-day disclosure rule, which requires a public prospectus at least 15 calendar days before a roadshow, points to October 23 as the last practical EDGAR filing day for a November 9 roadshow start, given that October 25 falls on a Sunday. Anthropic confidentially submitted its draft Form S-1 on June 1, 2026, announcing the confidential filing under Securities Act Rule 135.

Key figures that have circulated from draft documents: Q3 2026 revenue of $16.8 billion (up 45% sequentially from Q2's $11.5 billion), adjusted profitability, and an H1 2026 net loss of $80.7 billion driven by $72 billion in non-cash accounting charges tied to convertible-note fair-value adjustments. Previously reported infrastructure commitments exceed $518 billion. Bloomberg has reported a projected valuation of $1.8 to $2 trillion for the offering, expected to top SpaceX's June IPO in size. No price range, share count, or ticker is set until the public filing is live on EDGAR.

The public filing is the first time audited financials and the complete risk-factor section become available. Prior disclosures came from draft documents seen by Reuters and the Wall Street Journal. Any figure cited before today should be treated as reported rather than confirmed. First covered in this brief in Vol. I, No. 118 (Oct. 2); the public filing is the materially distinct event the prior coverage anticipated.

Disclosure: Claude, which generates this brief, is built by Anthropic.

Meta disclosed the $825 billion figure in its Q3 2026 quarterly filing, breaking it into $355.73 billion in off-balance-sheet operating and finance leases (data centers and energy infrastructure through 2036) and $454.52 billion in non-cancelable contracts for cloud capacity, servers, networking, and energy, plus up to $14.72 billion in contingent cloud-capacity purchases. The company also raised its 2026 capital-expenditure guidance to $135–145 billion, up from a prior floor of $130 billion, after reporting Q3 revenue that exceeded Wall Street expectations but profits that fell short. Meta chief executive Mark Zuckerberg signaled plans to build a "large-scale Muse network" and shift the company's focus toward AI agents.

Markets responded sharply. Meta shares fell roughly 7–8% on the day, dragging Nvidia, AMD, SpaceX, and CoreWeave lower with it. The Nasdaq fell 2.6%, its largest single-session decline since June. The S&P 500 fell 1.2%. Analysts at Investing.com noted that $825 billion in fixed commitments against Meta's $241.61 billion in trailing revenue is a leverage ratio that credit investors will price as meaningful risk, even though the obligations are structured as leases and contracts rather than bonds.

Anthropic and Theorem jointly published research on October 22 showing Claude generated approximately 180 million lines of Rocq proof code covering 239 fuzz-test entry points across 123 libraries drawn from Google's OSS-Fuzz initiative. The team compiled each library to RISC-V, then had Claude write Rocq proofs stating that the resulting binary would always exit cleanly without triggering memory-safety errors or assertion failures for any input under 64 MB, under a model of single-threaded, statically linked binaries on Linux. Theorem provided formal semantics for the RISC-V instruction set and Linux process behavior; the Rocq kernel verified each proof. A typical proof runs to roughly 600,000 lines of Rocq for a binary with a median length of 18,000 RISC-V instructions.

The process surfaced hundreds of previously undetected bugs across the libraries. Anthropic is withholding proofs for entry points where bugs remain unpatched. The public release includes 138 proved statements, a subset of the 239 reported in the research. Outside readers should treat the released set and the full experimental result as separate evidence bases. Anthropic describes the work as the first large-scale application of AI-driven formal verification to production binaries at the assembly level.

Disclosure: Claude, which generates this brief, is built by Anthropic.

Microsoft AI published the Science paper, titled "Sequential diagnosis with language models," on October 22, 2026. MAI-DxO (the Microsoft AI Diagnostic Orchestrator) scored 83.2% on 304 complex NEJM teaching cases and 80.3% on 325 emergency department admissions running on GPT-5, against 36.1% and 46.8% respectively for generalist physicians with access to reference materials. Physicians without reference access scored 22.3% on the NEJM cases. GPT-5 alone, without the orchestration layer, scored 78.9% on the NEJM cases, meaning the multi-agent orchestration contributed roughly four additional percentage points. Estimated diagnostic cost was modeled as lower for the system, though this is a projected figure, not an observed clinical cost.

The paper's comparator group is generalist physicians and general internists, not specialists or clinical teams. The cases are retrospective, meaning all information was already documented; the system did not interact with real patients or navigate incomplete, real-time data. Suleyman put the headline figure on X and framed the work as a step toward "medical superintelligence," while simultaneously stating the system needs clinical validation. Microsoft announced a partnership with Beth Israel Deaconess Medical Center in Boston to evaluate MAI-DxO in live hospital settings as the next phase.

The Wall Street Journal reported that OpenAI suspected Jasmine Wang, Tomek Korbak, and Mikita Balesni, three safety and alignment researchers fired on or around October 1, of sharing internal model-design details with Ryan Greenblatt of Redwood Research through Signal chats set to disappear. OpenAI called the firings a "significant breach of trust" over sensitive-information handling; the three dispute that characterization. Wang was an alignment researcher, Korbak was OpenAI's technical liaison to METR (Model Evaluation and Threat Research) during the investigation of the July 2026 Hugging Face incident, and Balesni worked on monitorability. Korbak says he was told in his exit meeting that the reason was "the way I communicated with METR." Balesni says his exit call pointed to his frequent conversations with third-party safety organizations.

The three published a four-page open letter to OpenAI's safety and oversight bodies on October 8 titled "OpenAI cannot make AI safe on its own," stating that the manner of their dismissals had made remaining employees afraid to speak to outside safety evaluators. Days after the firings, the trio wrote separately to OpenAI's board urging a halt on frontier models whose reasoning humans cannot audit. OpenAI answered on October 9 that its investigation found a breach of trust beyond what the letter described. Wang has tagged California and Delaware attorneys general and asked them to hold OpenAI to its stated commitments on evaluator independence. First covered in Vol. I, No. 121 (Oct. 2).