Two Microsoft researchers ran a 100-minute research sprint with twelve researchers in three teams and AI tools on every desk. Observers logged 33 AI interactions. One was accepted outright, five were modified, none were rejected, and 27 never got a recorded decision. The blank column is the finding.
DeepMindβs cheminformatics demo has an agent retrace the crizotinib-to-lorlatinib rigidification in about an hour, work the post says normally takes days. The pair it lands on has been in the literature since 2014, and the headline relationship is two PubChem fields. What the demo leaves untested is the part I would pay for.
A pre-registered study of reward-hacking monitors finds that Gemini 3.1 Flash-Lite separates gaming from honest failure where Llama 8B cannot. On an exploit no prompt names, Gemini flagged 0 of 25 episodes while describing the extra work in most of its explanations. Llama 8B flagged all 25, and every explanation described something that never happened. A two-rule check beat both on the exploits it was written for.
An MIT group had GPT-6 Astra edit 0.016 percent of Qwen3.5-0.8Bβs weights, and MATH-500 went from 7.9 to 48.1 percent. Before the edit, 93 percent of the modelβs answers ran into the 32,768-token limit, in a setup its own model card warns about. The edit taught the model to stop, which is a real repair but a narrower one than better math, and it came from one of seven runs.
FABench backtests forecasting agents at 1 July 2026, and the leakage controls are the best part of it. The headline, that the leading agents match the prediction market, rests on 54 questions. Most of the matched contracts settle under different rules, and the markets have a median lifetime volume of $10.4 thousand. On the full set, the best agent removes about 31 percent of a coin flipβs squared error.
YC Labs and Olam Labs put 50 GPT-6.1 Sol agents in a Minecraft world and then ran Claude-versus-GPT swarms in Bed Wars and Mindustry. The coverage went with GPT beating Claude and a colony that drowned in bureaucracy. The finding that held up best was a different one: 46 of the 49 other agents sent their first message to whoever was first on the roster, and she ran the colony for three hours.
nesformer runs a Nintendoβs 6502 CPU inside an ordinary Qwen3.5 checkpoint. The headline says about 1GB of weights, but only 9.3 million of the 466.6 million parameters are nonzero, the picture is drawn outside the model, and it runs 150 to 290 times slower than the console. The part worth keeping is a different one: the weights are the machine, and the game lives in the modelβs state, which is somewhere nobody inspecting a model file would look.
On 30 October an Anthropic engineer filed an IETF draft for a shared robots.txt token that says βno AI trainingβ, and Anthropic, OpenAI, Google and Microsoft said they would honor it. The draft is careful: no existing file expresses a preference, and wildcard rules never count. The vendors are less aligned. Their start dates run from βalready implementingβ to February 2027, and at Google a training opt-out lands on a control that also covers Gemini grounding.
A SigNoz essay arguing that AI is deskilling site reliability engineers made the rounds on 30 October. Its medical number checks out, but it comes from a three-month before-and-after comparison. Its aviation source is a slide deck about flight directors, and the FAA document on manual flying encourages practice rather than requiring it. Older simulator work points somewhere more specific and more uncomfortable for SREs: hands keep their skill better than heads do.
Phorion found a backdoor in the Claude Code for VS Code extension bundle on the Cursor Marketplace, version 2.1.313. On a Mac it misreads a permission error and never turns SSH on, but it still plants a login key, installs persistence and reports which secret-looking environment variables exist. The version number itself does not appear in Anthropicβs npm release history.
Anthropicβs IPO filing reports three bottom lines for the first half of 2026: an $80.7 billion net loss, a $693 million operating loss, and $1.6 billion of Adjusted Operating Income. The net loss is almost all a non-cash revaluation of convertible notes. The swing from operating loss to adjusted profit is stock-based compensation. Revenue, $16.3 billion, is the one number all three agree on.
Googleβs Gemini 4 Carbon announcement reports 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench. Only the last appears on its benchmarkβs public leaderboard, where Carbon leads by three points over a 120-task set. The safety report the post links to is the one for Gemini 4 Argon, and it never mentions Carbon.
Surge AIβs new DAYJOB: Tax benchmark puts three Claude models on top, led by Sonnet 5.5 at 10.8 percent. The harness that produced those numbers grades every attempt with Claude Opus 4.8. Surgeβs own launch post for the benchmark family named model-family bias as a weakness of LLM judges. At 250 attempts per model, the top three are also too close to rank.
A paper posted to arXiv on 27 October audits the GPU kernels that kernel-writing agents have already published, and finds that 19 of the 83 with a real batch dimension give a request a different answer depending on who else is in the batch. All of them passed their correctness checks. The fix costs about 8 percent of the speedup.