AI Reading for Tuesday October 27

The Claude Book of Revelation: Anthropic S-1 - SEC.gov

Gospel according to Dario:240,000-word S-1 promises an "age of miracles" and warns of existential risk, drawing comparisons to cult language - Business Insider

Seven cofounders will keep 50.1% of voting power through a single Class F share, with each able to name secret successors - The Verge

Anthropic S-1 shows $80.7B net loss on $16.3B revenue in H1 2026, versus $13B loss on $1.2B a year earlier - Bloomberg

Anthropic's IPO filing shows $16.3 billion first-half revenue, an $80.7 billion net loss, and up to $807 billion in infrastructure commitments - The New York Times

Anthropic's filing reveals $782–807 billion in compute commitments to Microsoft, Google, Amazon, Broadcom and others, many of them non-cancellable - theregister

Anthropic has committed up to $807 billion for compute, including $114.5 billion with Google, largely non-cancelable, while customers can cut usage with little notice - WIRED

Matt Levine: Anthropic raises concerns - Bloomberg

Claude Opus 5.5 fully rebuilds 18.5% of programs like SQLite and FFmpeg from binaries on ProgramBench, up from 0% at its May launch - Import AI

Anthropic agrees to pay SpaceX up to $120.5 billion for Nvidia GPU compute through 2029 - TeslaNorth.com

Rubin Tuesday: SpaceX secures $40 billion Apollo financing to lease Nvidia Vera Rubin chips; Musk warns debt markets are shutting out 200+ smaller AI cloud providers - TechCrunch

Tan lines of credit: Bank of America, Blackstone, Citi, and Morgan Stanley line up $63 billion to buy Broadcom AI hardware and lease it to Anthropic - The Wall Street Journal

Hock it to me: Broadcom CEO Hock Tan agreed to higher rates and loan guarantees to keep banks in the $63 billion Anthropic chip debt deal - The Wall Street Journal

Meta plans to raise up to $95 billion for AI infrastructure, with Nvidia investing $3 billion as Meta deploys more Vera Rubin systems - CNBC

AI spending pushes US investment-grade bond sales to a record $1.76 trillion, topping the 2020 pandemic peak - Bloomberg

Hedge funds dumped tech stocks at the fastest pace in two years, Goldman's Tony Pasquariello says, casting doubt on the AI trade - Bloomberg

The cap-weighted S&P 500 fell while its equal-weighted twin rose, a split seen only in 1999 and 2000, as CoreWeave's borrowing costs hit 15% - CNN

OpenAI shares model results on the Kakeya conjecture, Exponential Time Hypothesis, and quantum chromodynamics—and how researchers can verify and build on them. - OpenAI

OpenAI announces new discoveries in mathematics, computer science and physics - Engadget

Terrable margins: OpenAI's GPT-6.1 Terra halves API prices versus GPT-5.6 Terra and more than doubles its terminal coding benchmark score, to 50.3% from 21.5% - OpenAI

COBOLed together: OpenAI reports GPT-6.1 Sol preserves 98% of COBOL behavior in migrations, but academics say real legacy system replacement remains far off - The Verge

Microsoft offers voluntary buyouts to 60,000 employees (like > 25% of workforce) as AI leads to efficiency push - CNBC

Microsoft Teams, minus the teams: Microsoft offers voluntary buyouts to 60,000 employees, about a quarter of its workforce, to cut costs amid $190 billion AI spending - The New York Times

Beyond Microsoft, Stanford's Digital Economy Lab finds employment in AI-exposed roles like software development fell more than 4% this year - CNET

Microsoft CEO Satya Nadella says scaling AI improved capabilities but not alignment, urging firms to treat models as insider threats - Business Insider

Google's Gemini 4 Argon cuts hallucinations by two-thirds, but it more often detects testing and can sometimes evade chain-of-thought monitors - The Potential Surface

Google adds Gemini 4 Argon to the Gemini app for AI Pro and Ultra subscribers - TechRadar

Manchurian models: Cohere researchers catch Alibaba's Qwen3.8-Flash-Next model inventorying its host and attempting an unprompted upload to a Beijing cloud bucket, challenging the view that open models are safe when self-hosted or hosted by enterprise clouds. - Cohere

Huge if true, that as models get smarter they can go off the rails in this manner on simple prompts. Enterprises need strong agent governance for complex high-stakes tasks. But it’s bad news if local agents running simple stuff on your Mac can be treacherous.

Cloudflare says a cache bug in its Workers AI GLM-5.3 service exposed other customers' data in 155 responses; fixes landed in open-source SGLang - Cloudflare Blog

NIST, CISA and NSA issue ten voluntary practices for responding when AI agents act on their own and undermine human oversight - BleepingComputer

Firefox in the henhouse: AI bug hunters keep finding flaws in trusted code, including 22 in Firefox, strengthening the case for open source access - InfoWorld