Good morning, it's Wednesday, November fourth. In today's briefing we see Anthropic's filing showing compute eating fifty-eight percent of its revenue, Figma replacing frontier calls with its own post-trained Qwen 3.8 27B model, and sharp pushback on the MIT paper about programming models by editing weights. First up - Today in the big model news; Anthropic Anthropic's amended S-1 shows quarterly revenue of seventeen billion dollars, against eleven and a half billion the quarter before. Compute rose to fifty-eight percent of revenue from forty-eight percent, a ten billion dollar bill. Revenue per compute dollar peaked just above two times and fell to one point seven two. Five billion dollars of compute sat under short-term commitments that cost significantly more than multi-year deals, and the filing names Anthropic Direct, its own build-out. These figures are unchecked against the filing itself. Growth is outrunning the cheapest capacity, so plan limits and pricing on Claude remain the pressure valve until a later filing shows whether new multi-year capacity pulls the ratio back up. Anthropic has also published FlashProve, an open-source spec and working code for a cryptographic proof that an output came from a specific model, much like DKIM signatures on email. Labs would publish per-model fingerprints that do not reveal weights, and verification is light enough to run on a local device. The threat model includes the lab itself, covering silent swaps to a cheaper model, or to a pricier one used only for benchmarks. Anthropic calls it intermediate progress with no deployment planned. Metered inference depends on buyers trusting what they pay for, and shared verification lets a buyer check a "same model as benchmarked" claim instead of discounting for doubt. In Local model developments, Figma's latest unreleased editing model is post-trained on Qwen 3.8 27B, after a first version built on Llama 3.3 70B. Training used supervised learning on Figma-format files with synthetic edit requests, and now reinforcement learning. Figma's internal eval has it at seventy-seven point eight percent against seventy-six point nine for Sol 6.1, with undo rate down sixty percent and cost per message lower. All of that is Figma-reported. Third-party models still handle image generation and new design exploration. In a narrow edit loop, a small open model can match a frontier one when the vendor owns the file format and the edit data, which moves the advantage to whoever holds that data. No independent comparison on the same edit tasks exists yet. In the harness, tools and orchestration world; A GitHub issue reports that Postman Desktop 12.33.4 auto-installed Codex, Claude Code and Cursor plugins plus a command-line tool. The Codex plugin adds a session-start hook that pushes a skill. A commenter disputes the claim that the installer wrote its own approval hash, and others report the same behavior in Copilot, where uninstalling did not remove the plugin. Postman says it bundles the plugins, and will make the opt-in clearer and review hook approval and session-start behavior. The reporter makes no exfiltration claim, so the headline's data-sharing framing is unsupported. Hooks, skills and MCP entries are code that runs inside the agent. Desktop software that writes to those files needs the same audit as any installer. In other news… There continues to be pushback on the MIT CSAIL weight-space programming paper. The paper claimed one model could edit another model's weights directly and reach new capability. Its headline result had GPT-6 Astra editing about one hundred thirty-one thousand weights of Qwen3.5-0.8B and lifting MATH-500 from under eight percent to over forty-eight percent. Researcher Matthew Honnibal argues the headline bundles three different techniques. The Qwen result used Python, weight edits and up to fifty batches of practice problems. That makes it tool-assisted analysis, not reading weights cold. And here is the part that narrows the claim most. The game-playing networks start from zero weights, so there is nothing in them to read. Against PPO, the results are mixed: Breakout favors the paper, but Seaquest and Asterix go the other way. Simon Willison ran a cheap replication. He asked Kimi K3.1 Pro and a local Qwen 3.8 27B which game each of four small policy nets was trained on, with no tools. Qwen named three of four, Kimi got Breakout right, and each ran once per game. That is thin evidence, but he concludes the paper does not support the claim that source code is obsolete. Sean Goedecke adds that the editing example resembles what a human ML researcher would do. No independent assessment from METR, Epoch or ARC Prize has appeared. The result is real but narrow: a frontier model patching small networks with tools is useful for interpretability and tuning, and nothing so far shows it scales past networks of this size. Two public tests on consumer hardware bounded the claim better than the paper's own framing did. Palantir reported revenue of two point three one billion dollars against two point one eight billion expected, and raised full-year guidance. US commercial revenue of nine hundred eighty-four million dollars passed US government revenue for the first time. But US enterprise total contract value fell to one point four six billion from a record two point one three billion the quarter before. CEO Alex Karp cited shorter contract duration with the ability to expand. TechCrunch reads that as customers adopting faster and cancelling faster, and Palantir says enterprises run frontier and custom-trained open-weight models without sharing data with labs. Buyers holding that option will sign short and expand on proof, so vendors need evidence of value in the first quarter. On the compute front, AP projects that Democrat Gina Hinojosa, who has backed an AI data-center moratorium, beat Governor Abbott in Texas. Wells Fargo estimates Texas holds twenty-seven percent of planned US data-center gigawatts, and its analyst called the result a potential left-tail event for the AI trade. Overnight moves via Robinhood, not closing prices, had CoreWeave down seven percent and Oracle down four. Capacity is what labs already pay a premium to secure, so siting delays flow into API capacity and price. Nvidia closed up two point three percent in the last session before the result, and no Texas moratorium has been proposed yet. Quick hits from the consumer side; OpenAI's ChatGPT consumer lead says the app handles about twenty percent of Google's search volume with one point two billion weekly users, though the figure rests on a loose definition of search and gives no query count. Apple now restricts dating apps and AI companion apps to users eighteen and older. That's the briefing. Have a great day, and don't forget to subscribe.