Daily Brief Oct 28, 2026
How we checked every item today except "Also happened"
For readers who want to dig in. What happened, in brief, is on the Today tab.
This issue draws on material from October 27 to early October 28, plus an August retrospective in Research & models. Last night's collection took in 1,217 pieces, and this issue uses 23 outside sources with checkable links; the later sections also draw on material from the past week.
1. [This week] (filing, October 26) Anthropic's S-1 cash flow statement shows US$3.929 billion of first-half operating cash inflow, the first positive figure in five periods
Not proven yet： Baker's US$35 billion a year has no matching figure in the S-1
How we checked
The cash flow figures come from page F-12 of the S-1 itself, a primary document. The US$3.929 billion covers a single half-year and includes shifts in the timing of collections and payments, so it can't be read as recurring operating cash. The filing's risk factors also say operating and investing activities will keep using net cash for the foreseeable future. Baker runs an investment fund, and the posts don't say what related positions Baker holds. Free cash flow is the only one of Baker's three claims that can be checked against the filing. The US$35 billion a year has no matching figure there, and the October revenue claim rests on third-party usage data rather than a company disclosure.
2. [This week] (newsletter, October 26) Epoch tested six AI models on 11 of its own jobs; the two leaders reached 65% of the staff rubric, which Epoch calls a failure
Not proven yet： only two of six models are named, and 65% is their average
How we checked
The 65%, the 22% and "still a failed task" all come from Epoch's newsletter post, a primary source. According to the post, the full report is on Epoch's website; the post gives no per-model scores and doesn't say how the rubric items were set or who did the scoring. This is the second piece on the same study: Epoch's October 23 brief had already concluded that AI can't yet replace workers, at least not at Epoch (Epoch AI, 2026-10-23). Both pieces come from Epoch alone. The tasks come from Epoch's own work, with one attempt each and no human help, and don't extend to other knowledge work.
[Today] AMD says Meta has installed AMD's Helios rack-scale systems for testing, Meta will deploy a custom Instinct GPU, and AMD will put US$1 billion into Meta's equity offering.
How we checked
The press release gives no deployment scale, purchase amount, terms for the investment or total size of Meta's offering.
[This week] Stanford's Christopher Potts publicly questions a preprint's key claim, writing that the evidence for "subliminal learning" carrying across different models looks "weak at best."
How we checked
(Christopher Potts on X, 2026-10-26; arXiv 2610.30884, 2026-10-22)
[Today] In Anthropic's own randomized experiment, 86% of 50 professional analysts' choices on four analysis decisions followed a slant secretly given to an AI agent, against about 50% with no influence, and only 15% later said the argument came from the agent.
How we checked
[This week] (preprint, October 26) The VELA preprint finds that when websites open backend interfaces for AI agents to call directly, interface design mostly decides whether 44 author-built vulnerabilities stay exploitable; swapping among three commercial agents changes the exploitable share by at most 4.5 percentage points.
How we checked
The abstract doesn't say which three agents were tested or give counts by vulnerability type.
[Trend] (first published August 13, 2026) In the ICML 2026 community reproduction run by Hugging Face and alphaXiv, 1,221 people with coding agents checked 2,226 papers in 19 days, and the organizers say the most reliable results came from human-directed work.
How we checked
[Today] Oracle will sell OpenAI and Snowflake through its own channel. Oracle applications will come into ChatGPT, US customers can buy OpenAI with Oracle credits, and Snowflake will become a service Oracle Cloud Infrastructure (OCI) sells and bills directly. This comes from same-day announcements by OpenAI and Snowflake.
How we checked
[Today] Google made two physical-AI releases on the same day: DeepMind's world model Genie 4 launched as a research preview, and the robotics reasoning model Gemini Robotics ER 2 entered public preview on Google Cloud.
How we checked
DeepMind's post acknowledges that object permanence drifts in long simulations, and the post gives no evaluation figures (Google DeepMind, 2026-10-27). Google Cloud reports 91.3% accuracy at finding key-event frames; Google doesn't say which frames were tested or which models it was compared against (Google Cloud, 2026-10-27).
[Looking back] (SecondSource deep dive, October 21; rewritten October 27) If Anthropic's revenue can't keep up with its multiyear compute bills, Anthropic and its shareholders absorb the loss first, and the suppliers selling the compute come later.
How we checked
Still missing today. The gap that most affects judgment is still the year-by-year payment table and third-quarter revenue that the first S-1 lacks, which await the amended filing; as of early October 28, the company's EDGAR page shows only the first version. The three charts attached to Baker's posts, including the OpenRouter usage chart, weren't read for this issue. Per-model scores and the rubric on Epoch's full report page aren't in the newsletter post. Anthropic's blog post on the judgment-calls study says only that a full paper exists. How Amazon, Google, Microsoft and Broadcom account for these contracts isn't in the filings this issue drew on. For the three preprints, VELA, AgentMemRec and the subliminal-learning paper, only the abstracts were read.
Sources & accounting
The past 24 hours. 1,217 new pieces came in from October 27 to 28: 953 academic papers (the arXiv papers added last night, a different count from the 2,226 ICML papers in the Research & models reproduction item), 127 social-platform posts, 76 company and personal blog posts, 50 from a second academic-paper feed (counted separately from the academic papers above), 7 subscription newsletters, 2 company filings and 2 show transcripts. Of those 1,217, 278 were read in full, 141 were screened out as off-topic, 787 haven't been read yet, 11 were logged with a stated reason for not reading closely, and 1 was skipped before fetching. Of the academic papers, 220 were read, 95 screened out, 638 not yet read and 1 skipped before fetching. Of the blog posts, 18 were read, 11 screened out, 10 logged with a reason for not reading closely and 37 not yet read. Of the social-platform posts, 32 were read, 33 screened out, 61 not yet read and 1 logged with a reason for not reading closely. Of the newsletters, 5 were read and 2 screened out. Of the 2 company filings, 1 was read and 1, a pair of current reports from Broadcom and Vertiv, hasn't been read; the 287 filings listed as unread yesterday aren't on today's new-items table. Both show transcripts were read. None of the 50 pieces from the second paper feed has been read. Last night's collection picked up 461 original X posts; those 461 are counted differently from the 127 above, so the two can't be divided. One account, @teortaxesTex, accounts for 43 of the 461, followed by @RamaswmySridhar, @TheStalwart, @pstAsiatech and @PatrickMoorhead. Shows included Latent Space and Lex Fridman; newsletters included Latent Space, Exponential View, Stratechery and The Zvi. The three primary documents behind today's main line all came from material already waiting to be read: Anthropic's S-1, Gavin Baker's four posts and Epoch AI's newsletter. Five items in this issue center on Anthropic. The primary material on other companies comes from releases by Epoch, Hugging Face, AMD, OpenAI, Snowflake and Google.
One-time catch-up. No new one-time catch-up today. The August Hugging Face report in Research & models was fetched retroactively and isn't counted in the new pieces above.
Tracked sources. The long-term roster currently has 529 named sources: 302 on social platforms, 90 shows, 51 news outlets, 48 blogs, 48 paper authors and 46 newsletters, with the rest spread across earnings reports, keynotes and other channels. The same source may be tracked on several channels, so the channel figures add up to more than 529. These are counts of tracked sources and can't be added to the piece counts above. Representative names include Daniel Kokotajlo, Sergey Levine, Arvind Narayanan and Chris Lattner on social platforms; Dean Ball, Zvi Mowshowitz, Miles Brundage and Jamin Ball among newsletters; Terence Tao, Lilian Weng and Nicholas Carlini among blogs; Tri Dao and Song Han among paper authors; and Sam Altman, Dario Amodei and Demis Hassabis among show interviewees. This issue uses 23 outside sources in the body. Only outside links actually cited in the body are counted, and links to SecondSource's own site are left out; these 23 are a separate count from the 529 tracked sources. On the Taiwanese side, 7 new sources were added in September, bringing the total to 138; Taiwanese media, Taiwanese companies' official channels and local reporters in Taiwan count, while foreign outlets relaying Taiwanese media do not.