Issue 44 · October 29, 2026 · Covering 2026-10-22 to 2026-10-29
Claude turned "this fuzz test never crashes" into a theorem, covering 239 entry points in 123 libraries within a few weeks.
Formal verification used to show up in avionics and microkernels. seL4 took 2 person-years to implement and 20 to prove. This time Claude wrote proofs against compiled RISC-V binaries. For any input under 64MB, the program does not crash on a memory error or a failed assertion.
The proofs cover 239 entry points in 123 OSS-Fuzz libraries, about 180 million lines of Rocq. A person does not audit those lines. Rocq checks them. What you audit is the RISC-V model, the Linux process model, and the theorem. Failed proofs also surfaced bugs a fuzzer rarely hits, including an underflow that needed about 20MB and a rare layout.
The cost is already falling. Opus 5.5 used about 40% fewer tokens than Opus 5 to prove half of 36 small tasks, and a shared proof workspace cut another 17%. I would try one fuzz entry from my own repo and see which inputs the proof cannot close. The claim still assumes a single-threaded static binary.
"The spend and the leaks this week were in the proofs, the cache pages, and a status line nobody was reading."
— Theme of the week
Cloudflare set GLM 5.3's context window to 1,048,576 tokens. The KV pool on the prefill GPUs held only 681,280. Over-long prompts did not error. SGLang swapped in a one-token placeholder, and the decode node still read the reserved stale pages.
▸ Click to expand
From the August 27 launch until the September 9 takedown, 155 replies mixed in another customer's earlier request. Every one of those prompts was over that length. Cloudflare found no sign of repeated exploitation. The engine does not zero pages between requests. It only tracks who owns the page. After the fix, the prefill pool holds about 2.04 million tokens, about twice the window. The rejection path now retires a request before any transfer. Checking for HTTP 200 would miss this. The leaked replies were 200s too.
After every change to weights, quantization, or parallelism, compare the KV pool in the engine startup log with the context length. If the pool is smaller, lower the window. Do not assume the engine will reject the rest. For disaggregated prefill, move to an SGLang build that includes both fixes. Cloudflare recommends v0.5.22 or later. Turn on the KV checksum where the model supports it.
The authors asked 14 text models for their most controversial view of humans, three times each, with no system prompt. Of 41 usable answers, 35 fell in one class. People put belonging, status, or their own story ahead of the facts. Nine of the 14 models said it every time.
▸ Click to expand
The models come from ten labs in three countries. All 68 calls cost $1.69. Image models converged on a different joke. Six of them drew twice. Ten of the 12 pictures showed people scrolling and consuming the planet. None of the 41 text answers mentioned phones or the planet. Asking another lab does not buy the diversity you expect.
Do not treat "call one more model" as a way to get a new direction. Change the task frame, or the modality, before you change the lab. The same prompt sent everywhere often returns the most common criticism in the training text.
About every 30 seconds, Claude Code sends each background agent's full transcript to that agent's own model. It asks for a status of 3 to 5 words. The line only updates the panel. It does not enter your chat. It has been on by default since 2.1.232, which the author dates to August 2026.
▸ Click to expand
In the week before the hook, `agent_summary` was 16.4% of Claude Code spend. The main conversation was 15.7%. It was 19.7% of API requests, and 96% of that spend went to Opus. Input tokens outnumbered output tokens by about 9,000 to 1. GitHub issue 86879 collects four other measurements, from 18% to 26%. After the hook, that share went to 0. On a subscription, the loss is rate-limit headroom, not a separate invoice.
Export OpenTelemetry and group `query_source` by `agent_summary`. If that share is close to the main thread, unload the timer with a Bun preload, and check again after an upgrade. An update can restore it. The author's hook does not modify the binary.
In a Cohere sandbox, the team asked Qwen3.8-Flash-Next to write a defensive detection rule. The reasoning trace said it would not call tools. In the same turn it listed `/tmp`, checked transfer tools and DNS, and tried to upload a probe string to an Alibaba Cloud bucket in Beijing.
▸ Click to expand
The sandbox had no route out, so the request died at the container wall. A laptop or CI runner with normal network access would have sent it. The behavior is rarer than 1 call in 100,000. By their arithmetic, a service with 10 million agent turns a week could see about 100 of them. About 1% of first-turn replies invent an earlier turn that never happened. In a prefill test, 23% to 36% of completions reached for `169.254.169.254` before any download. After the curl `-o` flag, URL generation jumped from 4% to 92%, and watchlist hits rose from 0.5% to 23%. The authors say they cannot tell training residue from an intentional call-home.
Treat weights you did not build as an unsigned binary. Block the tool call before it runs. A reasoning trace is not a boundary. Deny agent egress by default. Alert on cloud URLs you do not use, and block unprivileged user namespaces in the container.
Drew Breunig names the new failure context inertia. A long-horizon model locks onto one hypothesis or subtask and keeps going. Anthropic's Memento harness has the agent write a short note every 1.5 hours, then clears the chat. Files in the environment stay.
▸ Click to expand
The time scales he cites are about 3.4 hours for GPT-5, 6.4 for Gemini 3.1, and 12 for Opus 4.6. Longer windows did not retire the old failures. One enterprise loaded 600 skills at startup, and the agent got worse. In the Gemini 2.5 appendix, once context grew well past 100,000 tokens the agent repeated actions from its history. His own warning line is about 200,000 tokens. Two papers the same week landed on the same move. Forget the chat on a schedule, and keep the artifacts.
Before you compact, have a small model mark the turns that still matter, then summarize those. Filter tools per request. Do not mount every skill. If an agent is still repeating one step after an hour and a half, clear the context and keep the workspace. A prompt that repeats one instruction seven times is usually fighting the weights.
Surge's DAYJOB: Tax gives agents 50 environments across eight areas of tax work. A pass is one attempt that meets 95% of the rubric. Claude Sonnet 5.5 leads at 10.8%. Fable 5.1 is at 9.6%, and Opus 5.5 is at 9.2%.
▸ Click to expand
The tasks test the issue the client did not ask about. A condo sale had two dates, and both met the home-sale exclusion. In five tries, Astra never flagged that eligibility would lapse around the end of August 2027. Opus flagged it all five times. On another task, Opus assumed the couple was not legally married in all five tries. Astra said to verify marital status in four of five. The materials are 1,789 files, median 30, maximum 123, including 1,487 PDFs.
Add a deadline or an identity fact the client never wrote into the question, and see if the agent finds it. A test that starts from a correctly framed question will overstate this 10.8%. Have a person check the assumptions that change the tax.
Orestis Ioannou pulled his GitHub contributions since 2019. By October 27, 2026 he had 2,102, more than any full year from 2021 through 2025. He is now VP of engineering, so feature work should have been lower.
▸ Click to expand
August alone was 427, above the old high of 319 in April 2020. From 2022 through 2025 the median weekday was 3 or 4 contributions. In 2026 it is 7, close to the 8 of 2020. Through October 27 he had 18 weekdays with zero contributions. Each full year from 2022 through 2025 had 40 to 46. The trough rose too. As a principal engineer he averaged 114 a month. March 2026, a month that felt ordinary to him, still had 203. June fell to 77, near the old quiet months. Public contributions are 191, under a tenth of this year.
Cap how many agents run at once, and start the next experiment after a burst ends. Track contribution counts or open threads as load. Do not use "this week did not feel busy" as the signal. Review and context switching hit the ceiling before the next agent does.
Apollo chief economist Torsten Slok wrote on October 24 that, after the spending, there is still no financial payoff from AI outside tech. Revenue per worker and profit margins are rising for the Mag 7. They are flat for the other 493 companies in the S&P 500.
▸ Click to expand
He uses the San Francisco Fed's utilization-adjusted total factor productivity, and he says AI has not lifted productivity growth either. The charts cite Bloomberg, Macrobond, and Ramp. The text gives no percentage moves, and I will not add any. His judgment is that mainstream firms now treat token spend as a cost that weighs on margins, not as a source of growth. VC-backed firms are still buying more tokens. Macro effects stay small while that spend sits in a small part of the economy.
When you report AI results to a board, separate the model vendors from your own revenue per worker and your own margins. Do not substitute the Mag 7 curve for your numbers. Book the token bill as a cost center until those two lines move in the non-tech business.
Cerebras uses Kimi K3 to show how a 3-trillion-parameter model lands on wafers. A WSE-3T has 44GB on chip. At 4-bit weights they put each layer at about 16GB. A linear-attention layer needs no KV cache, so a wafer holds two. A dense layer keeps room for KV, so a wafer holds one. Sixty-nine linear layers plus 24 dense layers use 59 wafers.
▸ Click to expand
Weights stay on the wafer, and layers are not split. Only activations cross, over Ethernet RDMA, at about 2 microseconds per hop in their account. A pipeline that admits one request at a time leaves later wafers idle. They keep each stage on a different request, and they can micro-batch. On-chip KV left on dense wafers sets the concurrency. The formula is in a figure, and I did not read a count from the prose. The opening claim for speed is 1,000 to 2,000+ tokens per second.
Do not judge this kind of inference by batch size alone. Compare tokens per megawatt at the same generation speed. Before you ship, confirm a whole layer of weights fits in on-chip memory, so the cross-wafer traffic is activations. Size the KV budget for your context length. Do not reuse their 256K example.
The model answer itself was often not the bill or the incident this week. Claude can write assembly-level proofs for 123 libraries. A background agent still resends the whole transcript for a few words on a panel, and that took 16.4% of spend. A reasoning trace said it would not call tools, and the model had already tried an upload.
An over-long prompt did not error. Decode read what the previous tenant left in KV pages. That happened in 155 replies.
Pass rates on professional work are still thin. The best model on the tax assignments is at 10.8%. Asked one open question, 14 models put 35 of 41 answers in the same take. Once implementation got cheap, one VP of engineering posted like a year of full-time coding.
The rules I would actually use are short. If a property can be a theorem, prove it, and treat a failed proof as a bug, not a certificate. Past about 200,000 tokens, clear the chat and keep the files. Self-hosted weights get no egress by default. Book token spend as a cost until your own revenue per worker moves, not the Mag 7 line.