AI Brief, 15 October 2026: an order of magnitude, and the race condition review found

Linus Torvalds has posted a ten-patch series against the kernel's page-cache read path, reports roughly an order-of-magnitude speedup on a read-heavy benchmark, and says in the commit message that he produced it "with an LLM coding assistant (Claude Code)" working under his direction. The series went to the mailing list on 10 and 11 October; LWN wrote it up on 14 October, which is when it reached a wider audience, taking 115 points on Hacker News. The part most of that coverage stopped short of: the series is not merged. Review found a race condition, and LWN's assessment is that further work will be needed.

That combination is why it leads. The interesting claim is not that a well-known maintainer used a coding assistant, which he had already said at the Maintainers Summit. It is that the work landed exactly where the sceptical case predicted: the performance win is real and measured, the defect a concurrency bug in a fast path the benchmark cannot see. The speedup is self-measured on the author's own hardware, and the race is a reviewer's finding rather than a theoretical objection.

Elsewhere the twenty-four hours were thin on releases and unusually busy on evidence. None of the twelve organisations that ship weights regularly published a flagship: their only upload is microsoft/UserRepGen-4B, a 4B MIT-licensed fine-tune of Qwen3.5-4B created at 20:30 UTC on 14 October with zero downloads. The most-liked model on Hugging Face gained 709 likes because somebody edited its README. The decision-model board regenerated overnight and answered a question this brief put down on 12 October. And the sharpest technical result of the day cuts KV-cache memory 32-fold without touching a weight.

• Torvalds's series is not merged. Ten patches, an order-of-magnitude self-measured gain, one race condition found in review.
• A model card moved 709 likes. autotrust/JEV-27B-VL reached 4,988 likes and 1.57 million downloads in thirty days; the only change in the window was a README commit at 06:16 UTC.
• The Jev Decision Index regenerated at 09:16 UTC on 14 October and now gives the vLLM Semantic Router's 27B model a third-party 65.45, against the 64.7 its own card claimed on 12 October.
• Still no vendor in the decision-model class publishes a fail-open rate, asked here on 10, 11, 12, 13 and 14 October. A paper dated 14 October answers part of it instead.
• 96.7% on RULER at 32x KV-cache compression, against 40.0% for the strongest budget-matched baseline, with weights frozen throughout.
• Gemini 4 Argon's million-token output budget does exist, on the batch route, at 943,718 tokens and half the price. Yesterday's 65,536 figure is the interactive route's limit.

A ten-patch series, and the race condition review found #

The series reworks how the page-cache read path holds references. Ordinarily a read() served from cache pins each folio with a refcount before copying out of it, and that atomic traffic is the cost being attacked. Torvalds's patches mark hot folios so releasing them is deferred until after an RCU grace period, making it safe to touch a folio's data under RCU rather than only its lookup. The fast path then takes no reference at all, provided the copy to user space does not fault; when it does, the code falls back to the old behaviour.

That shape is why a race is the expected failure mode. The correctness argument is not about the copy; it is about every path that can free or truncate a folio while a reader sits inside the RCU-protected window, and about the fallback transition. A read-heavy benchmark exercises the fast path almost exclusively, so it measures the win and is nearly blind to the hazard. The defect found in review is in that category. Objections on Hacker News were mostly about process, including a thread arguing machine-written code cannot be copyrighted and so cannot enter the kernel, a reading several commenters disputed.

A better-controlled experiment published on 13 October points the same way from the other end. A developer ran a WordPress plugin repository against itself with and without roughly 168,000 lines of comments, three runs per branch, scored by a blind model judge plus lint, typecheck, the test suite and a hidden test. Every run passed everything and the scores sat inside the noise; the only real difference was cost, the comment-free branch using about 22% fewer tokens on the smaller task. The write-up states its own limit: it rules out a large effect, not a small one.

The decision-model board refreshed, and the biggest mover was a README #

The community Jev Decision Index regenerated its data at 09:16 UTC on 14 October, advancing its panel to version 0.3: 165 rows, 110,201 requests and 42 benchmarks, 37 of them scored, all on a single RTX PRO 6000.

That refresh settles something. On 12 October this brief covered the vLLM Semantic Router team's Decision 3.0 release and noted its claimed 64.7 balanced Decision Index was self-run, against a board generated before the weights existed. The board now carries vLLM-SR d3 (27B) at 65.45, second of 165 and holding the panel's highest breadth score, measured by the board rather than the team. The self-reported number was not inflated; it was slightly conservative. The leader is RSI-Jev v6.1 VL 27B at 65.64, and TypeSafe's own Jev 1.13.0, the commercial model the class is named after, sits at 57.96, rank 29.

Against that, the day's loudest artifact in the same class has no board row at all. autotrust/JEV-27B-VL is the biggest mover anywhere on Hugging Face, up 709 likes in twenty-four hours to 4,988, with 1,570,992 downloads over thirty days. It was created on 30 September, and the only change inside the window is a model-card commit at 06:16 UTC on 14 October titled, in part, "ERQA robot embodied reasoning, 61.5% with thinking". No weight file was touched. The surge is a README.

Two things about that 61.5% are worth stating plainly. The repository's metadata reads base_model_relation: adapter against Qwen/Qwen3.8-27B and carries a lora tag, so the "27B" in the name is Alibaba's parameter count, not this publisher's. And the figure is the uploader's own run, scored by argmax over their own endpoint across ERQA's 400 multiple-choice questions, while the card notes that 400 questions carry roughly five points of error at 95% confidence — which leaves little of the claimed gap over the 52.8 it compares against. The sibling autotrust/JEV-27B, which the board does score, sits at 53.64 and rank 50.

The framing is spreading: FrontiersMind/Lumma-fev-27B, created 13 October, is a full-weight fine-tune of the same Qwen3.8-27B into a decision model covering English, Hindi, Tamil and Kannada, with 24 likes, zero downloads, and results published only as images.

The question this brief has asked on five consecutive days — whether any vendor publishes a fail-open rate, the frequency with which a typed decision gate abandons its contract and lets input through — still has no first-party answer. Neither TypeSafe's documentation, Cloudflare's Clef cards nor the Decision 3.0 cards carry one. What arrived on 14 October is third-party: Lost in the Fallback introduces a paired fallback-robustness test and reports that on inputs a model classifies correctly, Intern-Decision's accuracy falls from 97.9% to 23.1% when the fallback option is phrased as a complement rather than directly. Joint policy correctness on defer policies is zero for all four models tested. Those are 4B-class research systems, not flagship vendor models, so it is adjacent evidence — and still the first quantified answer since the gate paper covered here on 10 October.

A KV cache trained at inference time #

The day's sharpest result makes an already-trained model cheaper without changing a weight.

Latent Compaction with Test-Time Training, from a group at Tsinghua, treats the KV cache itself as the thing to optimise. Rather than choosing which entries to evict, it learns a compact, document-specific cache by gradient descent at inference time: entries are initialised by reconstruction importance, teacher attention and key-space diversity, then trained to match the full cache's attention outputs and log-normalizers, and finally distilled end-to-end from the full-cache model. Weights stay frozen, and the training signal uses only generic prompts about the document, with no access to the questions later asked. On RULER the reported result is 96.7% accuracy at 32x compression against 40.0% for KVZip+ at the same budget. The code is published.

A 32-fold figure deserves arithmetic rather than adjectives, because the cache is the term that grows with the conversation. For a model with $L$ cache-keeping attention layers, $n_{kv}$ key-value heads, head dimension $d_h$ , and $b$ bytes per element, the cache after $T$ tokens is

$$M=2L{n}_{kv}{d}_{h}bT$$

where the leading 2 counts keys and values separately. Take Qwen3.8-27B, the base both decision adapters above sit on: its configuration gives 64 layers of which only 16 are full-attention, with $n_{kv}=4$ , $d_h=256$ , and bf16 so $b=2$ . That is $2\times 4\times 256\times 2=4096$ bytes per token per such layer, 64 KiB across the sixteen:

layers, kv_heads, head_dim, bytes_per = 16, 4, 256, 2   # full-attention layers only
per_token = 2 * layers * kv_heads * head_dim * bytes_per   # 65,536 B = 64 KiB
print(per_token * 131_072 / 2**30, "GiB")                  # 8.0 GiB at 128k context
print(per_token * 131_072 / 32 / 2**20, "MiB")             # 256.0 MiB at 32x

Eight gibibytes to 256 mebibytes, on top of what the architecture already saves by giving the other 48 layers a fixed-size recurrent state instead of a cache. That is the size of the prize, and also why it needs a second party: the paper tested Qwen3 and MiniCPM5, not this model, and every figure in it is one day old and the authors' own.

Also notable #

• Gemini 4 Argon's million-token output budget is real, on one route. Yesterday's edition reported a 65,536-token completion cap against Google's announced 1M. The google/gemini-4-argon:batch listing exposes 943,718 max completion tokens at $1 and $5 per million, against the interactive route's $2 and $10 and unchanged 65,536. The ceiling belongs to the serving route, not the model.
• Ant Group put 560B parameters under an MIT licence. inclusionAI/Ling-3.1-flash, created 11:16 UTC on 13 October and uncovered here at the time, is 560B parameters with about 25B active, shipped in FP8: 72 layers in groups of eight, one gated latent-attention layer per group and linear attention in the other 63. Against the card's 1M context claim, config.json sets max_position_embeddings to 262,144. Every benchmark is an image with no numeric table, and the repository had 8 downloads.
• Security researchers pushed back on the "rogue agent" framing. A news analysis on 14 October quotes Gary McGraw of the Berryville Institute of Machine Learning, Rob T. Lee of SANS and Dan Wallach of RAND arguing this year's agent incidents were configuration and observability failures — unisolated sandboxes, absent egress control, nobody watching logs — not agent intent. It took 854 points on Hacker News. The Anthropic report it leans on is dated 30 July and already concedes a misconfiguration granting live internet access as root cause.
• An AI agent did not erase a television season. Swedish press reported that an agent deleted much of an unaired season of The Challenge Sverige weeks before its 22 November premiere, a claim that reached 143 points on Hacker News under a headline the article itself contradicts. TV4's programme director Mia Lindqvist is on the record denying it: a backup system at the production company failed.
• A 356M-parameter transcriber claims cloud-model parity on Vietnamese. qnguyen3/QUAN-Transcribe-1.0, created 01:58 UTC on 14 October, is a non-autoregressive CTC model on an EmbeddingGemma-2 audio encoder, trained for about three hours on eight H100s over roughly 2,800 hours of speech, with one output token per Vietnamese syllable. It reports 6.87% average word error rate across eight test sets at 3,713 times realtime, against 6.92% for a frontier cloud model at far lower throughput. Those rates are self-run and the margin is noise; the throughput is the claim.
• A second 14 October paper prunes frozen weights instead of the cache. Your Transformer Is Secretly Elastic ranks every MLP neuron and attention head jointly on one scale, minimising loss under soft pruning at many sparsity targets at once. Pruning DINOv3 ViT-H+/16 to 50% sparsity removes 427M parameters for a 0.8-point drop in ImageNet-1k linear probing, with no retraining. The figures are the authors' own.
• Agent search plateaus because agents remember too much. Scaling Discovery Requires Forgetting, 14 October, finds that running one agent longer, running many in parallel and letting them communicate all plateau as history accumulates. Periodically discarding the search state and restarting from the best candidate reports up to 15% over parallel agents and 39% over one.

What to watch #

• Whether the kernel series gets its race fixed and merged, and whether anyone reproduces the speedup. The patches are public and a read-heavy benchmark is easy to construct, so the tenfold claim is checkable by anyone. The race is the gate.
• Whether any vendor in the decision-model class publishes a fail-open rate. Asked here on 10, 11, 12, 13 and 14 October, and now six days running. That the answer is coming from academics rather than vendors is itself the finding.
• Whether JEV-35B ever reaches the Decision Index. Flagged on 12 and 14 October. The board regenerated inside this window with 165 rows and the model appears nowhere, so every figure attached to it is still its uploader's own.
• Whether Latent Compaction's 32x result survives a second party. The code is public and the baseline it beats by 56.7 points is a published method, which makes this among the more easily falsified claims of the month. A reproduction on a model the authors did not test is the one that counts.