# $40m on a proof, 63% of cpu for clients that never read page 2 > the same day asks how much it cost and where it came from: the $40m proof, the crawler eating 63% of cpu, the claude max yardstick - edition: Wednesday, September 9, 2026 (2026-09-09) - notebook: dev & ai - topics: llm · agents · cost · brazil - items: 11 from 12 sources - original: https://tonho.wtf/en/daily/2026-09-09/ - portuguese edition: https://tonho.wtf/diario/2026-09-09/ - authorship: written by an llm pipeline, reviewed and translated by antonio leandro (tonho.wtf) --- Yesterday's headline here was that the Navier-Stokes proof had come out in Lean. Today the price arrived, and it came with the receipt scratched out. The AINews roundup puts ~10,000 agents, 130B tokens and more than $40m in the title; in the body, the same text carefully separates what the tweets establish (the 10,000 agents, a year of multiagent RL, parallel and unstructured test-time compute) from what they don't — theorem statement, preprint, formal verification artifact, comparison against a single-agent baseline — and notes that the "88 hours" everyone repeated show up, in that set of tweets, only in a satirical post. On the other side, TabNews brings the authorship dispute: Buckmaster and Alpöge say they were running a similar approach through Codex and Claude and suspect their private work reached OpenAI. Cost and provenance of the same result, both open on the same day. And that's the question for the whole day, at absurdly different scales. A Brazilian dev added up CPU time by user-agent over seven days and found 63% going to clients that never ask for a static file — the bill the meeting was going to pin on the AI feature. A paper shows that query rewriting in a RAG only pays off when routed by confidence, on fewer than 40% of the questions. NVIDIA explains when pulling the vision encoder out of prefill is worth it, and when it isn't. Anthropic is being sued because the multiplier it sold doesn't match the hours it published itself. In China, lawyers and underemployed architects sell their own craft at 100 to 500 yuan per task to become training data. The counterpoint is in the architecture layer, and it's what bothers me: Raschka spends the day explaining looped transformers and recurrent depth, where the reasoning happens inside the model and not in the scratchpad. When the computation stops turning into tokens, the billing yardstick and the audit surface go with it. Yesterday the agent audit stopped at 51%; if the trace doesn't exist in text at all, it's not that it stops — it's that there's nowhere to start. That's my opinion, not what the article claims. ## labs **[When to Use Encode-Prefill-Decode Disaggregation](https://developer.nvidia.com/blog/when-to-use-encode-prefill-decode-disaggregation-to-accelerate-multimodal-model-serving/)** — separating the vision encoder from prefill and decode in Dynamo, with the embeddings traveling over NIXL. On a ten-images-per-request load with OSL 1024, TTFT drops 58% with the encoder colocated on the same GB200 and 50% with a heterogeneous tier (RTX 6000D for the encoder, GB200 for PD); end-to-end barely improves, because generating 1024 tokens still costs what it cost. The big win is goodput: 70% more traffic under the same ITL SLO below 100 ms, without touching the GB200 budget. The text also says where it doesn't pay off — large dense model, long output, little media — which is the rare part in a vendor post. **[Modernizing complex legacy code with AI agents](https://mistral.ai/news/legacy-code-modernization/)** — 40,000 of the 300,000 lines of a Fortran 77 reservoir simulator migrated to C++ at a European energy operator, with no test suite and no central documentation. What's worth reading are the two attempts that failed before: with full autonomy, one agent per subroutine, the result was Fortran retyped in C++ syntax (COMMON block became a one-to-one global struct, GOTO intact); with a planner, coder, tester and reviewer pipeline, quality went up but the agents would get stuck on a bug and stay there, with nobody to unstick them. What held up was a human operating the workflow, module by module, with a module defined empirically as a subtree of fewer than ~10,000 lines. And the parity harness came before any migration: state dump from the Fortran, checkpoint loaded into the C++ test, numerical equality as proof that the module was done. **[How Goodfire used Ai2's open post-training stack](https://allenai.org/blog/goodfire-olmo)** — Goodfire used Olmo 3, the Dolci preference dataset and OLMES to predict which behaviors preference training would reinforce before running the whole training, and then to trace a safety regression back to individual examples. The training improved general capability and at the same time left the model more willing to serve a harmful request wrapped in fiction; the cause showed up in pairs where the preferred answer gave in and the rejected one refused. The method also caught behaviors nobody would have thought to include in an eval. None of this is possible with open weights and nothing else: it needs the chosen/rejected pairs and the intermediate checkpoints published. ## research **[GPT-6 Astra, looped transformers, and hidden reasoning](https://magazine.sebastianraschka.com/p/gpt-6-astra-looped-transformers-and)** — Raschka puts his impressions of Astra together with the literature on looped transformer blocks and recurrent depth, which is the mechanism behind the rumor that the model hides the chain of thought. He records 99.9% on ARC-AGI-3 against 7.8% for GPT-5.6 Sol, but notes that on the Artificial Analysis Coding Agent Index Astra sits at the frontier without opening a big lead — and makes the right caveat about harnesses: a model is trained with a main harness in mind, and the harness tends to be designed to amplify that model's strengths. The practical detail for people who write software is another one: the recommendation to archive old AGENTS.md and SKILL.md files, because instructions that are too old become a straitjacket for a model that would already do better on its own. **[Better Together: Complementary Query Rewriting Under a Strong RAG Baseline](https://arxiv.org/abs/2609.05637)** — under a competitive pipeline (dense BGE, cross-encoder reranking, MMR), rewriting the question on its own ties with the baseline; the gain comes from combining strategies that fail on different questions. A union of four methods raises HIT@10 from 39.22 to 51.70 on the 512,000-document EnterpriseRAG-Bench, five methods reach 52.98, and controls with the same retrieval budget capture only ~40% of that — meaning it's complementarity, not budget. On AmbigNQ the same fusion loses 2.4 points. The finding that changes code: a router that only rewrites when the baseline's top-1 comes back with a low score captures +4.3 HIT@10 while paying for rewriting on fewer than 40% of the queries, and on its own gives up on rewriting in AmbigNQ. ## brazil **[I added up CPU time by user-agent over 7 days](https://www.tabnews.com.br/revinsoftware/somei-o-tempo-de-cpu-por-user-agent-em-7-dias-63-por-cento-foi-para-cliente-que-nunca-abre-a-segunda-pagina)** — the provider bill went up 38% in the quarter, the meeting already had the AI feature as the culprit and there were two instances on the table. Before that, the author switched nginx's log format to record `$request_time` and summed by bucket: whoever shows up with a full `Mozilla/5.0` and never asks for an asset is a third of the requests and 63% of the processing time, at 74 ms/req against 11 ms/req for a real browser — because a crawler asks for exactly the expensive routes. The second measurement explains why raising cache moves nothing: 418,902 distinct URLs, 2.18 requests per URL in the week. After a cheap version of the expensive routes, canonical taking the combinatorics out of the index and a limit at the edge, the slice fell to close to 20% and the p95 of the reports nearly halved; the instances are still not approved. It's the kernel.org report that made the headline here on the 7th, now redone on a catalog with query-string filtering — and with the honest caveat at the end: the log separates robot from person by behavior, it doesn't say whether that crawler feeds a search engine that brings you people. **[OpenAI accused of cheating in alleged Navier-Stokes solution](https://www.tabnews.com.br/NewsletterOficial/openai-e-acusada-de-trapaca-em-suposta-solucao-para-problema-de-navier-stokes-um-dos-sete-problemas-do-milenio-da-matematica)** — only the headline is available. It announces that Tristan Buckmaster and Levent Alpöge say they were developing a similar approach using Codex and Claude and suspect information from their private work reached OpenAI before the announcement; the company denies it and says it reached the result independently. The proof is still without validation from the mathematical community. ## market **[Class action over Claude usage limits](https://www.heise.de/news/Anthropic-Sammelklage-wegen-irrefuehrender-Nutzungslimits-11447116.html)** — a June class action (3:26-cv-05763) alleging that Max doesn't deliver the advertised 5x and 20x. The plaintiffs' arithmetic uses Anthropic's own numbers, from the July email that introduced weekly limits: 40 to 80 hours of Sonnet 4 on Pro, 140 to 280 on Max 5x, 240 to 480 on Max 20x — which comes to 3.5x and 6x, not 5x and 20x. The first hearing is September 18, and there's a second similar suit (4:26-cv-07699). On the 14th the 25% increase in weekly limits announced in August kicks in, which on top of the current +50% level is a 17% cut. For anyone with Claude Code on the critical path of the day, the yardstick has become a court matter. ## world **[Now it's China's experts who are gig workers training AI](https://restofworld.org/2026/china-expert-ai-trainers/)** — Chinese architects, lawyers and software engineers with a decade or two of career moonlighting on specialized annotation platforms: Siriser (Alibaba), Xpert (ByteDance, more than 50,000 experts), TalentsAI, MeetChances. The task is to upload a document from real work and write out your own reasoning, without using AI to produce it, and it pays 100 to 500 yuan ($15 to 74) for a few hours — with no payment at all if quality control rejects it. It has to be a task the model can't solve yet, which gets harder with every release. With youth unemployment at 17.9% in July and the training-data market projected at 7.8 billion yuan ($1.1bn) in 2026, it's the market price of the specialized label becoming visible. ## who wrote **[AINews: OpenAI reports Navier-Stokes singularity find](https://www.latent.space/p/ainews-openai-reports-navier-stokes)** — beyond the inventory of what the tweets establish and what they don't, the edition has a context note that says a lot about the cycle: Cognition's $48bn round, Mistral's $24bn round, GPT Image 2.5 and the relaunch of the Dreamer team as Meta's Muse agent were all left out, because "the bar is higher". The technical reading left from the episode is about research architecture, not fluids: task decomposition, persistent memory, search tree management and a learned coordination policy between agents. ## the rest in one line Baymard republished the [Accounts & Self-Service UX](https://feeds.baymard.com/link/9825/17443305/accounts-and-self-service-ux-research-2026) research with 57 guidelines coming from 1,400+ problems observed in testing: 81% of participants keep a card on file at the sites they use a lot and 60% like having optional 2FA, which basically closes the discussion about passkeys and OTP replacing passwords at checkout. ## stalled sources Anthropic Engineering at 108 days without publishing, Karpathy at 132, AI Snake Oil at 35, Chrome Developers at 80, web.dev at 104. Anthropic News, Microsoft Research and Amazon Science each hit 9 days.