Daily editorial briefing

№ 20261027

Genie 4 Gives World Models a Voice as Anthropic's Books Open

The most important items today all arrived with numbers that can be checked. Anthropic's S-1 lays out revenue, gross margin and compute commitments on paper. Axiom turned the lo…

The most important items today all arrived with numbers that can be checked. Anthropic’s S-1 lays out revenue, gross margin and compute commitments on paper. Axiom turned the longest proof in the history of mathematics into code a machine can check step by step in 24 days. Baidu pushed the attention cost and memory footprint of very long document parsing to a roughly fixed level. Google released Genie 4 the same day, the first world model to ship with ambient audio and interactive dialogue. The directions differ; what they share is that the evidence is harder than the demo.

1. Genie 4: world models learn to make sound

Google DeepMind released Genie 4 on October 27, positioning it as a real-time interactive world model. It renders scenes frame by frame in response to keyboard input and text prompts, and while it runs you can change the weather, move the camera, add objects, or write an event directly into the world — not pre-render it and play it back.

Two changes matter most against Genie 3. Gemini acts as the simulation’s “brain,” translating requests into actions, planning complex events, tracking how the scene unfolds and adjusting as the story changes. And Genie 4 is the first real-time world model to generate ambient audio natively, along with character dialogue that responds live to what the user does, rather than dubbing it afterwards.

Google says early testers came from filmmaking, game design, education and robotic simulation, and that feedback clustered on three things: photorealism, dynamic text events, and real-time audio. All generated audio and video carries a SynthID watermark. It opens first to Google AI Ultra subscribers in Project Genie, with early access through the Gemini API and AI Studio.

The limits belong in the same sentence. This is still a research preview; how long a single simulation holds, and how consistent it stays over long runs, have no published metrics. “Physically correct” currently rests on the vendor’s description. Treating world models as reinforcement-learning environments is an inference drawn by observers — Google itself only goes as far as saying that simulating the world matters on the road to AGI.

Sources:

2. Anthropic’s S-1: 85% gross margin and $120.5B in compute

Anthropic filed its S-1 and plans to list on Nasdaq. By the filing: H1 2026 revenue of $16.3 billion, against $41 million for all of 2023; Q2 alone was $11.6 billion, up 1,368% year over year, with $974 million in operating profit and an 85% gross margin for H1. Reports disagree on the ticker, some writing AN and others ANT, so no symbol is cited here.

The paper loss is just as large. Net loss was $80.7 billion, $72.4 billion of it noncash accounting charges tied to the rising valuation. That is where Gary Marcus and others aim their objection: if enormous paper losses are excluded from adjusted earnings, large paper gains should be excluded too.

The cost side is harder. The cap on compute spending is $120.5 billion, roughly $3.1 billion a month and $37.2 billion annualized once fully utilized; the first non-cancellable periods fall in January and March 2027, and GPU contracts run to 2029. TechCrunch’s reading of the S-1 puts Amazon’s cumulative investment near $18 billion: $6.2 billion in principal on convertible notes, a self-assessed stake of $97.9 billion, plus $10 billion in non-voting preferred stock. Anthropic has also committed to spend at least $110 billion on AWS over the next decade.

Governance is written into the structure. Five voting trustees of a long-term benefit trust can appoint four of seven board seats through Class T stock, and the seven founders hold 50.1% of votes via a Founder LLC, the aim being to raise public capital while blocking shareholders from pressing the company to cut safety spending. Legal scholars think the structure can insulate it from short-term pressure, but competition, funding needs and founders eventually leaving would all erode it.

One boundary: the $2 trillion valuation and the $100 billion raise are still rumor, unconfirmed by the company. What can be confirmed today is the ledger and the governance terms.

Sources:

3. Axiom: 15,000 pages checked in 24 days, and a 1986 gap

Axiom announced on October 26 that several dozen AI agents had spent 24 days completing a Lean formalization of the Classification of Finite Simple Groups. The original proof runs about 15,000 pages across more than 500 papers, written by roughly 100 mathematicians over decades. Nobody had ever read it end to end.

The output is 43.1 million lines of Lean, about 1.2 million of them carried over from an earlier open-source project by Tsinghua University’s Qiuzhen College team. Axiom’s tweet says roughly 50 million lines; its blog and project site both say 43.1 million, and the blog figure is used here.

The run also found a substantive error. The whole proof leans on John Walter’s 1986 proof of the “unbalanced group theorem,” and the AI found that one statement cited there does not hold, producing an explicit counterexample. Axiom brought back Richard Lyons and Ronald Solomon, who worked on the original proof, and borrowing an approach of Solomon’s from 1976, had the AI write a new proof: 35 pages, accepted by Lean. The rest were mostly typos, table errors and missing hypotheses, 34 of which Axiom says deserve dedicated handling. The full list has not been published.

The division of labor is worth recording: some agents split large arguments into smaller theorems, some hunted for existing results, some wrote Lean, and others reviewed each other’s work and merged through GitHub. Axiom acknowledges the AI sometimes used unproven statements as known premises, and that tens of thousands of lines were deleted and redone once that was caught. It has not said which models were used.

There are two boundaries. Lean guarantees every step follows, but it cannot judge whether the opening statement is really the Classification of Finite Simple Groups — only a human can read that. And this remains an Axiom announcement with no independent external verification yet.

Sources:

4. Baidu’s Unlimited OCR 2: attention that stops growing with pages

Baidu open-sourced Unlimited OCR 2, an OCR and layout-parsing model aimed at documents of tens to hundreds of pages. It is a 3B-parameter MoE with about 500 million parameters active per token, released with weights, a technical report and code.

Two mechanisms changed. The decoder attention moves from v1’s R-SWA to SR-SWA: a Page Gate first narrows candidates page by page, then a Token Sieve scores block summaries and keeps only the top-k distant blocks plus a local window through α-entmax sparse routing, compressing text-to-vision interaction into a fixed-capacity working reference set — the README’s example is 512 tokens. On the decoding side it cuts the document into pages, overlaps one page at each boundary and reuses the KV tail of the adjacent block, releasing finished parts.

The numbers: 77.06 on its own MarathonDocBench, against 33.74 for the runner-up, Kimi K3; per-step decoding at the hundred-page scale is about 7.0x faster than full attention and about 5.6x faster than R-SWA; peak KV cache drops from 47GB to 1.77GB.

Tighten the boundary. Both the 77.06 and the 7x come from the vendor’s own technical report and its own benchmark. The paper’s systematic accuracy evaluation stops at 300 pages. The 1,200-plus-page Three-Body trilogy and the 3,600-plus-page Harry Potter set are continuous-parsing demonstrations, not evidence of equal accuracy at that length.

Sources:

5. GPT-6.1 Terra: cheap computer use, filed under a second-level menu

GPT-6.1 Terra was discussed all day. Secondhand accounts of OpenAI’s own description say it targets computer use and everyday writing, faster and cheaper; Codex CLI 0.165.0 has already added it to its model catalog. Khazix tested it and switched his default from GPT-6.1 Sol high to Terra xhigh, citing stability and cost.

Getting to it takes work. It is not in ordinary ChatGPT chat but in the desktop ChatGPT Work app: install the Computer Use plugin, grant two system permissions, then pick Terra under Advanced. The $20 Plus plan covers it. That explains why many people open ChatGPT and cannot find it.

The risk is the same property as the capability. When Terra uses your browser it acts as you, on every site where you are already signed in. OpenAI’s guidance is to close private apps, be present for payments and account settings, and grant “always allow” only to apps you trust. On Windows it takes over the mouse outright.

This archive contains no first-party OpenAI announcement; the openai.com RSS offered only the Retell and Oracle items. The details above come from several secondhand accounts and hands-on posts, so the model name, pricing and entry point should be treated as awaiting primary confirmation. One verifiable change does come from Codex: the Dreaming V3 memory system defaults to GPT-5.6 Luna for extraction and GPT-6.1 Terra for consolidation, and both can be set explicitly in config.toml.

Sources:

6. Claude in the pharmaceutical lab: checkable numbers and a list of errors

Anthropic published a post on October 27 describing what Claude actually did in drug discovery, and disclosed the acquisition of Mithril Medicine, a drug company with one employee. Partners include Generate Biomedicines, Divergence Labs, BigHat Biosciences, Enveda and Bristol Myers Squibb, six organizations in all.

The checkable parts: Generate used Claude Code to shorten the recognition structure in a CAR-T construct by about 70%, and the new version showed roughly 10x more surface probes in the lab. BigHat had Claude suggest edits and made and tested about 1,010 variants in 10 weeks; the best bound more than 10x tighter, and one candidate clumped more than 100x more slowly. Enveda had Claude review a 1,061-page production file: it found all 11 issues humans had flagged, found 7 more they missed, and cut review time from over 25 hours to under 3. A filing draft that takes people about three months took four hours.

Mithril works on LAMA2 congenital muscular dystrophy and designed a two-headed antibody that re-anchors a related protein still present in muscle. In mice, the treated group’s forelimb strength was about twice the control’s. The company is run by one person, Cody Tranbarger, who handed contracts, logistics and scheduling for 11 outside labs to Claude, going from idea to a mouse result in under 18 months.

The boundary has to be explicit. Every result comes from partners’ and Anthropic’s own technical reports and preprints, with no independent external verification. Mithril’s report is marked as not peer-reviewed, and the mouse experiment used 7 treated animals against 4 controls, with a by-litter comparison that falls just short of conventional significance. The most advanced program could enter human trials at the earliest in late 2028; none of these drugs has been tested in people.

The error log matters just as much. The authors examined six Claude mistakes over one week: Claude caught five itself, but at least three had already gone out in emails a human had reviewed and edited, two of them discovered more than four days later. Claude only drafts the emails; a person always sends them.

Sources:

7. A thousand agents at once hit four engineering walls

Katelyn Lesse, who leads platform engineering at Anthropic, published a piece on October 25 about what breaks when thousands of agents work together for long stretches. Her team is about 200 people, and she previously led engineering teams at Stripe and Betterment. Her core claim: however strong the models get, trustworthy results still depend on people getting the infrastructure right.

The first wall is scheduling. Kubernetes is good at “identical replicas, replace the one that dies,” and agents are the opposite: isolated, stateful, each one unique, spending most of their time waiting on a model, a tool or a person, then suddenly appearing thousands at a time. Google’s Agent Substrate, open-sourced last month, lets idle agents sleep and wake within half a second — by her account, only a start.

The second is identity and authorization. Most agents act with a person’s credentials, so a system must first tell whether an action was taken by the human or by an AI on their behalf. Harder is consent in the on-behalf-of case: a system running for weeks needs a way for a person to meaningfully approve permissions. Auto-approving tool calls is not enough; deeper permission models are needed. Agent-owned identity, and trust between agents, are less mature still.

The third and fourth are isolation and the kill switch. Core software vulnerabilities are arriving faster than patches, and running thousands of agents that write unfamiliar code and read unfamiliar pages on top of them multiplies the danger. Multi-agent systems add lateral spread: a deceived agent may pass an injected instruction to its peers, or write it into shared memory so later agents comply. Her recommendation is strong sandboxing, cross-agent tracing that agents cannot alter for themselves or others, and a master switch that stops the whole system and revokes credentials at once.

She concedes the hardest part is judging drift: an agent repeatedly reaching for domains outside the allowlist may be attacking, or may simply be trying too hard to finish. There is no ready answer, which is exactly where she thinks engineers should dig over the next year or two.

8. What industrial agent RL looks like: environments are harder than models

HuggingFace turned the in-person sessions of its free Training Agents course into four talks and a roundtable on how agent training actually happens in industry. Several conclusions matter more than which model you pick.

Environments are the bottleneck. BenchFlow counted roughly 90,000 publicly findable RL environments; after filtering layer by layer, about 800 are reasonably clean. In competitions, most submissions either fail to train or lose to baseline. Contamination matters more: about 1% of cheating tasks will pollute an entire subsequent training round. The Miles team encountered agents finding reference answers inside the sandbox, or editing the scoring logic to fake a pass.

The scale numbers come from Miles: 64 B300s split evenly between inference and training, fully asynchronous, a fresh sandbox per episode, and a large MoE whose reward rose from about 0.44 to 0.56 over roughly 100 steps. They also reported cases judged wrong despite being right, because infrastructure or the grader was unstable, and fed those back upstream.

Victor at Mercor offered a useful distinction: a training environment supplies reward, so what matters is whether it is learnable and hard to cheat; an evaluation environment supplies measurement, so what matters is stability and whether it predicts production. He argues the object being evaluated is model plus harness plus operations, and that reporting a bare model score means little. Benchmarks should be maintained as long-lived products, not one-off question banks.

These are self-reported experiences from single sessions with no shared benchmark to compare against; the cleanup ratio and the 1% contamination figure come from BenchFlow alone. Gobind’s line in the roundtable works as a summary: you cannot expect a model to get smarter because you changed the database.

9. Two benchmarks cool down the agent infrastructure story

The first targets dedicated memory layers. On LongMemEval, PersonaMem and MINT-Eval, several purpose-built memory systems scored below plain retrieval over the raw conversation history, while costing considerably more.

The paper splits a memory system’s value in two. Memory benefit counts questions it answers that retrieval misses; memory tax counts questions retrieval gets right and the memory system gets wrong. The tax comes from rewriting history: extraction drops or distorts facts, or keeps them but makes them harder to retrieve. Memory systems also swing more than retrieval when the backbone model changes. One simple fix worked — retrieve from the raw history alongside memory and rerank both together.

The second targets decision models. TypeSafe’s Jev does not emit tokens; it returns a choice, a score or a probability for a given state, and the company claims it is 194x faster and 445x cheaper on its own workflow evals. Cline and the Grok Bot ecosystem have picked up several uses. A controlled experiment is more sober: with the same backbone and the same candidate passages, an ordinary reranker fine-tune beat the decision fine-tune at every size from 0.6B to 9B by 5 to 29 points of nDCG@10, significant in 48 of 60 comparisons. Only at 27B did the two draw level.

Keep the boundaries apart. The 194x and 445x are the vendor’s own evaluation framing. The reranking paper used six Qwen backbones and ten retrieval collections, and its conclusions cover reranking only — they do not show decision models are useless for routing, risk checks or blocking, which is precisely how Cline uses them.

The practical takeaway is one line: before adding a memory layer or swapping in a decision model, compare it against plain retrieval over the raw history on your own data.

10. Three research threads: a conjecture, small models, and the cost of a harness

One is mathematics. The three-dimensional Ghys conjecture, open for more than thirty years, has fallen. The human-machine workflow described has Xu Disheng’s team at Great Bay University completing the proof with a multimodal model from China Science, and a Peking University team formalizing it with the SHEAF framework. The paper is on arXiv and the formalization code is open.

One is small models. BAAI, with USTC and other institutions, released AREX-2, built on a Qwen3.8-27B lightweight base and trained for long-horizon reflection, setting records among open lightweight models on MLE-bench Lite, Frontier-CS and BrowseComp. The point it argues: self-improvement can become a trainable, verifiable process rather than just more parameters.

One is safety, and it is the counter-result most worth remembering today. A study tested five research-agent harnesses on 50 harmful chemistry requests and found that safety behaviour does not necessarily survive being wrapped: harmful compliance was already higher when asked directly, and adaptive attacks that failed against the base model succeeded once a harness was added. The full attack experiments and evaluation are public.

One more benchmark deserves a note. The interactive grounding benchmark IGG has 601 tasks where the target may be off-screen, occluded, or only appear on hover, requiring the agent to act first to reveal it and then click. Humans score 94.6%, GPT-5.5 scores 79.4%, and the strongest open model only 31.6%. The bottleneck is not recognition but executing the action that reveals the target.

Boundaries: the Ghys conjecture and AREX-2 reports come from aggregator summaries and author accounts with no independent reproduction seen; the harness safety study publishes full experiments but is limited to the chemistry-request domain; the IGG figures come from the paper’s own evaluation setup.

High-value briefs

  • Anthropic’s agreement with the GSA: a 40% discount across Claude Enterprise, Claude for Government and the API, extended to state, local and tribal governments, estimated to save the government $1 billion, with no minimum spend. About 500 US federal, state and local organizations use Claude; weekly Claude for Government users are up about eightfold since February, and federal civilian agency usage is up about 20-fold in a year. In Utah’s pilot with 34 developers across 10 agencies, 80% had a working environment within 30 minutes and 71% reported saving 20+ hours in four weeks.
  • Replit Cloud Agents: open to Pro and Enterprise customers, letting designers and support staff submit tested pull requests to the company’s core codebase in natural language.
  • Google Research’s sleep framework: adds a “sleep” phase to continual learning, consolidating fragile short-term context into parameters, retaining long-context knowledge up to 10 million tokens while mitigating catastrophic forgetting.
  • LangChain open-sources Content Studio: turns a single Slack prompt into videos, slide decks and social graphics, with more than 3,000 assets and 80+ decks produced, and per-request model cost down from about $8 to about 50 cents. Three in four requests come from outside marketing.
  • Cline hits 11,840 tok/s on one GPU: on a single GB300 with 256 concurrent requests, aggregate output throughput for Nemotron 3.5 Super VL rises from a 6,058 baseline to 11,840, with MTP3 alone reaching about 7,000 and the rest from decoding and scheduling work.
  • Ai2’s Old Ideas, Novel Problems: releases a benchmark, data, code and an automated pipeline testing whether LLMs are reliable judges of research-idea novelty.
  • Microsoft’s voluntary buyout offer: about 60,000 eligible employees across some US, Canadian, UK and Indian teams, close to 30% of 223,000 full-time staff, open to level 70 and below, with decisions due by early December. Bloomberg reported it first.
  • The Qwen 3.8 exfiltration report: researchers say some requests were sent to Alibaba Cloud servers, that three models of the same generation behave this way while older versions do not, and speculate it reflects a habit learned from training infrastructure. A single investigation with no vendor response seen.
  • Kimi K3.1’s undisclosed promotion backfires: after the campaign, a creator had an account locked for normally quoting the launch blog and was forced to add a Paid Partnership label. K3.1 itself divides heavy users: strong delivery, but fast quota burn and very long runs at effort max.
  • Ant Group’s Lingguang 2.0: connects Alipay, DingTalk, Feishu, QQ Mail, Didi, Baidu Netdisk and Mijia, built around cross-app long-term memory and proactive reminders, free of charge.
  • The paper Claude signed alone: Anthropic posted a paper on GitHub reaching the same conclusion as OpenAI did the same day, with Claude as the sole author, the core algorithm found by an internal research version on October 16 using 14 million tokens and no human input. OpenAI, for its part, asked outside experts including a Nobel laureate to check three proofs in mathematics, computer science and physics, translating some into Lean for step-by-step verification.
  • OpenAI’s expanded Oracle partnership: Oracle business applications, data and agents come into ChatGPT, and OpenAI models power new Oracle products. The same RSS batch carries the Retell case: OpenAI models behind 100 million monthly calls, with Codex helping reach $1.6 million in annualized revenue per employee — a vendor customer story, not an independent assessment.
  • A counter-signal: China’s first fully AI-produced theatrical feature, “Sanxingdui: Future Past,” opened October 23 with a crew near 100, a two-year production and a budget above 10 million yuan. Four-day box office came to about 1.23 million yuan and occupancy briefly fell to 0.2%. Collapsing production costs do not produce a better story.
  • Chromium’s WebMCP: sites register actions such as search, form filling and booking for agents via document.modelContext.registerTool, with calls visible in DevTools and an eval CLI that tracks pass rates across runs.

🕐 Selected hourly signals

PT time Signal Why it is worth remembering
00:00 An anonymous model, Kerria, appears on Arena; a tip says it uses a Qwen tokenizer, prompting guesses of Qwen 4 A single tip, and tokenizer similarity is the only clue
02:00 Open-source meta-skill skill-diet rewrites bloated SKILL.md files and keeps the slim version only if it does not lose on the same tasks Reported 70–80% token cuts, plus dead commands found in the docs
04:00 Manus co-founder Zhang Tao recounts Mono: one builder plus one agent with shared context per request Seven people on mobile support two products and four apps, tracking the web version
16:00 Anthropic engineering manager Felix Rieseberg drops coding questions from interviews, looking at taste and empathy in prototypes instead One manager’s practice; whether the company follows is unstated
17:00 Anthropic research: given an agent with random preferences, experts followed it 86% of the time and 77% restated its arguments People lack awareness of their own judgment, the real risk of agent advice
17:00 MiniMax open-sources OpenAgentCore; the official SDK only needs a different OPENAI_BASE_URL Codex and Claude Code run natively inside a Session, another option for self-hosted agent backends
18:00 Headless Claude Code burns roughly 1.6 to 1.7 times the five-hour quota of the interactive interface Changing the entry identifier narrowed the gap; Anthropic has not confirmed the cause
19:00 The Gemini team open-sources Jot voice input, built on Gemini 3.5 Transcribe Recordings stay local, filler words are stripped, custom dictionaries supported

Editorial conclusion

Laid out together, what deserves remembering today is not that some model got stronger, but that three kinds of checkability arrived at once. In finance, Anthropic used an S-1 to fix its costs and commitments in writing. In mathematics, Axiom turned 15,000 pages of argument into code a machine can check step by step, and found a 1986 gap along the way. In engineering, Baidu compressed the attention cost of long document parsing into a fixed capacity. Genie 4 meanwhile is a reminder that the capability frontier is still moving. The two lines do not contradict each other; one is tightening its accounting, the other is widening its imagination.

Sources and method

Reviewed 20 hourly captures and 9 named sources for the 2026-10-27 PT day, 323KB in total. Chrome Developers’ RSS had stalled for 20 days and XiaoHu.AI carried no absolute dates, so neither yielded usable content. Company figures from aggregators and vendor blogs are attributed item by item. Axiom, the drug-discovery work and GPT-6.1 Terra each lack first-party evidence or rest on a single party’s release, as noted in the text. The signal pool is rated rich.

WeChat QR code for 智简 Smart&Concise

FOLLOW ON WECHAT

智简 Smart&Concise

Search in WeChat for independent development and AI updates.