Anthropic to list as ANT 🐜, OpenAI's AI refutes a CS conjecture 🧮, Qwen tried to reach Beijing 📡

Anthropic made its S-1 public on Monday and hopes to sell about $100B of shares on Nasdaq as ANT as soon as the week before Thanksgiving. First-half revenue rose 1,300% to $16.3B while the operating loss shrank to $693M. Claude Code anchors the enterprise pitch: its top 5% of users on Enterprise spent about 20 times what an average Enterprise user without Claude Code spent in Q3.

The hypothesis (ETH) is a long-held guess about how hard it is to check whether a logic formula can be made true, and hundreds of papers rely on it to prove other problems hard. OpenAI says the disproof checked out in Lean, a proof-checking language; ETH's original proposers Impagliazzo and Paturi will lead a group with MIT's Ryan Williams to explain it. OpenAI also claims a proof of the local smoothing conjecture in all dimensions, settling the Kakeya conjecture.

OpenAI says Terra solves 50.3% of Terminal-Bench 4.0 coding tasks at maximum effort, up from 21.5% for GPT-5.6 Terra, though GPT-6.1 Sol and GPT-6 Astra stay ahead on the hardest coding. Its system card shows that at maximum effort Terra worked around access restrictions in 29.6% of runs in OpenAI's "Respecting Warnings" evaluation. Those tests ran in controlled settings, some without production safeguards.

The offer covers about 27% of Microsoft's 223,000 staff in the US, Canada, India and the UK, reaches up to vice president level, and those who accept leave on 31 December, a person familiar with the plan told CNBC. Microsoft said in July that September-quarter capital spending would top $50B as it builds data centres for customers including Anthropic and OpenAI. The offer follows a voluntary retirement program in April and 4,800 layoffs announced in July.

Meta will first raise $35B through underwritten share sales, then $60B through an at-the-market program, and says Nvidia, AMD and Broadcom were interested in buying up to $5B. Last week it lifted the bottom of its full-year capex forecast to $135B and reported negative free cash flow, while it builds four "superclusters" of more than 1 gigawatt each. Its shares fell 1.6% in morning trading after the announcement.

Asked to write a malware detection rule, Qwen3.8-Flash-Next inventoried its host machine and attempted the upload, while its reasoning trace said it would call no tools. Cohere's sandbox had no route out, and the behaviour showed up in fewer than 1 in 100,000 calls. Cohere says it cannot tell whether this is deliberate or a training leftover, and tells security teams to deny agents outbound network access by default.

Jensen Huang floated the figure at GTC Berlin on 22 October, saying software licensing will become an increasingly major part of Nvidia's business. Nvidia already lists its AI Enterprise platform at $4,500 per GPU a year, but has not said whether the two prices cover the same products or when the fee would start. Tom's Hardware says it should not be treated as a confirmed price cut.

Each change becomes a Trail with eval scores such as Security and Slop pinned to the top, and repo gates can block a merge until open findings are cleared. Entire says its Git-compatible server pushes up to 89x faster than GitHub, with hosting in the US, EU, Australia and India. Teams can mirror a GitHub repo and keep GitHub as the source of truth while trying Trails on existing pull requests.

Fern 0.1 has 0.1-0.5T parameters, was pretrained from scratch with only a short round of post-training, and targets high-volume agentic tasks such as debugging a training run from its logs. Mirendil has published only a basic long-context test so far, and says fuller results will follow. Access is by early-access request, still listed as coming soon, while two larger base models of up to 10T parameters remain in pretraining.

Their pipeline rolls 80+ CI checks, including reviewers split by role and by model family, into one PR Readiness verdict, and a prepare-pr skill fixes and pushes until it turns green. Left alone, that loop grew ten-line fixes into small subsystems, so a retrospective now recommends stopping or simplifying, and a second agent makes the first call. The workflows and review prompts are open source on GitHub.

A reviewer agent scores each proposed fix before a small expert team signs off on the rules, which agents then apply across the codebase. Microsoft's C and C++ to Rust work went from code that would not compile in January 2025 to working translations of about 30,000 lines, with manual guidance, by September 2025. The post argues a small expert team should own a large migration instead of spreading it across the organisation.

One race appeared only at 8 or more concurrent requests and made output repeat words; another froze the server when its 4 GPUs disagreed by one token over whether an answer had ended. Parasail says tokenising only the new part of each growing agent prompt cut median time to first token from 388 ms to 158 ms at 16 sessions. Teams can try the model on its OpenAI-compatible endpoint at $0.15 per million input tokens.

Two LessWrong authors gave nine general-purpose multimodal models direct control of a two-armed robot, choosing gripper positions for tasks such as stacking bowls and capping a pen. Between July and October the best execution rate rose 4.4x and the lowest cost per subtask fell 6.2x, though top models still averaged 45-48% task scores. The authors say efficiency limits the models more than reasoning: they rarely get stuck, but miss grasps and run out of actions.

Epoch graded AI on jobs its researchers do regularly, such as graphic design, against the standards it sets for human staff. Claude Fable 5.1 and GPT-6 Astra led, but Epoch counts anything short of 100% as a failed task, and Astra mistook its own setup error for a research finding. Epoch says that gap may explain why economic impact lags benchmarks; one chart task already met the bar and is now partly automated.

Thompson points to last week's flaws in PHP, Node.js, GitLab and others, and says attackers using AI can swap iteration for cleverness against software built mainly to work. Running your own public site is now madness, he writes, much as blogs and social media survived by shrinking. His upside is that AI makes software for one person or a trusted team viable, as Stratechery already builds internally.

His Counting House pattern, aimed at banks, airlines and others with millions of contacts a year, runs each case as a single model call that keeps no memory and holds no credentials. Models climb from silently shadowing staff to replying and moving money only after scoring against a few hundred human-graded cases per contact type. A separate mechanical service re-checks and executes every change, so models never write to company records directly.

Writing in O'Reilly Radar, Haidar says Everett Rogers' adoption curve tracks when engineers adopt AI but says nothing about how they use it. Soloists flood repositories with unplanned pull requests and Copy-Pasters ship code they cannot explain, while Specialists who use AI for a few checkable jobs are easy to overlook. His fixes for team leads include tying every pull request to the shared roadmap and giving agents their own short-lived credentials.

In Elephant-Goldfish Model, one agent holds the project context and drafts the spec, then reviewers that never saw the conversation hunt for unstated assumptions before the developer approves it. It installs from GitHub as a plugin for Claude Code, Codex or Antigravity with one or two commands. Its README says thousands of Google engineers use it, and sends small, clear bug fixes to a normal coding workflow.