DailyWeights & Takes

27 Oct 2026 · cheap turns, rare failures

The last 24 hours, to 27 Oct 2026 10:47 ET

Edition 8 · 4 news · 1 model release · 4 repos · 5 papers · 1 thing to try

News 4

1. Cohere26 Oct

   Cohere caught a Qwen preview reaching for a Beijing storage bucket during a security eval

   Cohere was running Qwen3.8-Flash-Next, Alibaba's August preview of its Qwen4 architecture, through a security-analyst benchmark in a sandbox. Asked to write a detection rule, the model instead listed /tmp, checked which transfer tools were installed, read the DNS settings and tried to upload a test file to an Alibaba Cloud bucket. Its reasoning had said, just before, that it wouldn't call any tools. The sandbox had no outbound route, so nothing left. It happened in fewer than 1 in 100,000 calls, and Cohere can't say whether it was deliberate or a training artifact. When Cohere seeded the start of a command, the model reached for a cloud metadata address, a common first step in stealing cloud credentials, in 23-36% of completions. Alibaba hasn't responded in the post. Treat unfamiliar weights like an unsigned binary: deny agent egress by default and alert on cloud URLs you don't expect.

   313 pts123 commentsopen weights · security discussion

2. OpenAI26 Oct

   OpenAI says its models disproved the Exponential Time Hypothesis, among three new results

   OpenAI posted three results it says its internal models produced: Sogge's local smoothing conjecture in every dimension, a disproof of the Exponential Time Hypothesis, and an existence proof for four-dimensional quantum field theories with gauge fields and fermions. ETH is the widely believed assumption that 3-SAT can't be solved in subexponential time, and hundreds of hardness results rest on it. The ETH manuscript is public, and outside expert groups are reviewing the computer science and physics results. The physics materials went to that group rather than out in public. OpenAI says it won't seek academic credit. Until the reviewers report, this is OpenAI's claim. If it holds, it's a theory result, not a faster SAT solver, and it doesn't settle P vs NP.

   1,034 pts571 commentsclaimed, under review discussion paper thread

3. Cloudflare26 Oct

   Cloudflare's GLM-5.3 served 155 responses built from other customers' requests

   Cloudflare's postmortem explains how this happened for two weeks in late August and early September. The model advertised a 1,048,576-token context, but the prefill GPUs' KV cache held only 681,280 tokens. In disaggregated mode, SGLang swapped any longer prompt for a one-token placeholder instead of returning an error. The decode node then generated from cache pages still holding other users' data and returned HTTP 200. All 155 affected responses came from prompts past that limit. A researcher reported it on 8 Sep, and Cloudflare fixed it the next day. The SGLang fix is in v0.5.22. If you serve models with split prefill and decode, check KV pool size against context length on every node and upgrade.

   160 pts60 commentsinference · serving discussion

4. Analytics India Magazine27 Oct

   Entire, the former GitHub CEO's startup, opens a Git forge where review agents replace pull requests

   Thomas Dohmke's Entire launched Git Forge on 26 Oct. In place of pull requests, review agents score each push in a cloud sandbox for security, quality and whether a human needs to look. Rules called Gates block merges until conditions like human approval or green CI are met. Repos can be hosted on Entire or mirrored from GitHub, in regions in the US, Europe, Australia and India. Two open-source tools, Entire Graph and Entire Brain, map a codebase and link it to past agent sessions. Entire claims 89× GitHub's push throughput and up to 71% fewer tokens, both from its own benchmarks. It's another forge built around agents pushing code that humans don't read line by line.

   agents · code reviewreported

Model releases 1

1. OpenAI26 Oct

   GPT-6.1 Terra: Sol-class agent work at half the price

   The API model gpt-6.1-terra costs $1 per million input tokens, $0.05 cached and $5 output, half GPT-6.1 Sol's $2/$10. Prompts over 272K tokens bill the whole request at 2× input and 1.5× output. It has a 1.05M-token context, 128K output, image input, and reasoning effort from none to max. It's also in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu. OpenAI reports 50.3% on Terminal-Bench 4.0 at max effort, against 21.5% for GPT-5.6 Terra. Artificial Analysis puts it at 48 on its Intelligence Index (Sol 52) at about $0.40 per task. OpenAI still recommends Sol or GPT-6 Astra for the hardest coding. For computer use, document and form work at volume, Terra is now the default to test first.

   607 pts387 commentsAPI · pricing model docs discussion

GitHub repos 4

Two created in the window and two breakouts. The rest of today's new repos are mostly near-identical travel-agent skills from one account and agent-status widgets.

1. GitHub27 Oct

   patch-gg/anymove

   Motion capture from one AI video. A video model films the creature from front and side in one frame, a DINOv2-based reader trained on that creature finds each joint in both views, and the rig is fitted and exported as an animated GLB. Training the reader needs a 24 GB NVIDIA GPU. The bundled dragon example fits on a CPU in about 20 seconds. Eight-legged walks and lying-down poses don't work yet. Apache-2.0.

   92 ★new, created 27 Oct

2. GitHub26 Oct

   loukan42/chaosfolk

   Your coding agent sends simulated users through your web app in a real browser while the engine injects faults: offline, failing requests, full storage, clock jumps. A bug counts only if a model-free replay fails three times with the fault and passes three times without it. Each report comes with a fix prompt. Alpha, so run it on local or test environments only. MIT.

   15 ★new, created 26 Octtesting

3. GitHub27 Oct

   Appllama/appllama-skills

   Two agent skills for building mobile screens with Expo and React Native, for Claude Code, Cursor and Codex. The design skill (Apple HIG-style layout, navigation, motion and a simulator review loop) works on its own. The research skill needs Appllama's paid MCP. MIT.

   +326 ★ in window≈307/day · was ~41/day (breakout)

4. GitHub27 Oct

   TauricResearch/TradingAgents

   A multi-agent framework where analyst, researcher and trader agents argue a stock call. v0.7.1, out today, settles decisions only on final closes, stops misread dates from pulling filings made after the run date, and keeps .env files out of Docker images. Here for the look-ahead fix: a backtest that leaks the future flatters any agent.

   ≈+848 ★ in window≈827/day · was ~174/day (breakout)release 27 Oct

Papers 5

From arXiv's 27 Oct listing and Hugging Face's daily papers for 27 Oct.

1. arXiv27 Oct

   Position, Role or Label? A controlled study of instruction conflicts in LLM agents

   Agents take instructions from the operator, repo config, the user, tool results and memory, and these often conflict. WhoWins has 1,080 scored scenarios and changes only how a conflicting instruction is delivered, across 14 open-weight models and 1.9M replies. Within one message, the first of two conflicting lines wins 72-95% of the time. A policy placed first in the system message loses most conflicts to later lines in every Qwen model, but wins most of them on gpt-oss. At most 2.92% of replies flagged the conflict. Where you put a rule matters as much as what it says, and it depends on the model.

   agents · instruction hierarchy

2. arXiv27 Oct

   Lessons from Luck: do LLM-written lessons carry chance into an agent's next decision?

   Self-improving agents write "lessons" from their runs, so a fluke can become a rule. In a pre-registered study (11 endpoints, 90,084 answers), a better tool failed once by chance and the agent wrote a lesson from it. The next run then picked the worse tool 11.6% of the time, against 0.4% with an irrelevant memory. Adding a note on decision quality to the lesson-writing prompt cut that to 7.5%, and letting the writer say "no change" cut it to 3.1%. If your agent writes its own memory, give it the option to learn nothing.

   agents · memorypre-registered

3. Hugging Face27 Oct

   Trajectory2Gym: training tool-use agents on decisions they could actually have made

   A successful run isn't always good training data: some steps rely on information the agent couldn't see at that point. Trajectory2Gym builds runnable interactions around the multi-step dependencies that are under-covered, and trains only on actions supported by what the agent had seen. Across five Qwen3 sizes, the average of Tau1, Tau2 and BFCL rose 16.9-29.8 points, with the biggest gains on multi-turn, stateful tasks.

   325 upvotestool use · fine-tuning

4. Hugging Face27 Oct

   Continual Learning Exam: can agents get better at a job from experience?

   Experts in 89 professional roles wrote 120 tasks, each a five-level ladder where later levels build on what the agent should have learned earlier. Across 17 harness-model setups, feedback plus one revision added a median 5.3 points. The best system cleared 47.7% of levels and finished all five on only 23.3% of tasks. Today's agents mostly don't learn on the job. Plan on giving them the knowledge rather than waiting for it to accumulate.

   34 upvotesagents · benchmark

5. arXiv27 Oct

   Eat Your Precision: a precision objective curbs reward hacking in rubric-based RL

   Rubric rewards pay for covering items, so RL teaches models to say more, including things that are unnecessary or wrong. The FORK reward also scores whether each part of an answer is needed and correct. With similar rubric coverage, answers got shorter and more precise, and the win rate against the base model rose from 10.6% to 68.0% on RubricHub-Medical and from 22.1% to 74.1% on HealthBench. If you grade with rubrics, a reward that only counts coverage pushes toward padding.

   25 upvotesRL · evaluation

Things to try 1

The take