Microsoft put a date on the shared robots.txt rule that lets a site refuse AI training, and split training from grounding so publishers can choose each separately. For builders, crawl rules for AI are moving from one rule per company to one rule for many.
The Biological Computing Co. launched mischief 1, an image-to-video model built on Wan 2.2 that it says matches the base model's quality while generating video 5x faster. TBC grows living neurons in its lab, studies how they process information, and turns findings into software that adds less than 0.1% to the base model's parameter count; no neurons run when you use it. On fal and OpenRouter, a five-second clip costs 2.5 cents.
Google says its Project Suncatcher prototype, launched October 1, returned the first AI response from a data center class TPU in orbit. Engineers uplinked a prompt during a ground-station pass, and the onboard Gemma 4 open model replied with a haiku about being in space. Next, Google will test how the TPUs cope with thermal stress and radiation, on the way to twin interconnected satellites in 2027.
ICALL-BENCH tests whether models can follow how functions call each other inside compiled binaries, using the targets of indirect calls as the probe. It covers 50 real C/C++ programs averaging 74,746 lines of code and 941 functions each. Across five recent LLMs and their agents, program-level success ranged from 23% to 81%, and giving agents tools or internet access did not consistently help on the hardest cases.
The authors built a set of USAMO 2025 candidate solutions graded by experts and tested 20 models as proof graders, finding considerable room for improvement. Their MABGrader treats grading as choosing among the possible scores, the way a multi-armed bandit picks which option to try. It beats the previous best, ProofGrader, by 26.4% in Quadratic Weighted Kappa, a measure of agreement with expert grades, and reduces grading error by 14%.
A new paper splits the gap between small and frontier models on terminal tasks into three parts. Sampling more attempts narrows the gap in reaching a correct solution but does not close it. Given the same pool of attempts, frontier models are far better at picking the one that worked, and small models are more overconfident about their own chances of success. So extra tries alone cannot close the gap.
EpiAgent splits an agent's decisions into two steps: update its beliefs about other players' private information from the interaction so far, then plan an action using those beliefs. Both steps are trained by self-play, with a reward for how well the beliefs predict what happens next. Across mixed and competitive scenarios, it beat prompting-based and reinforcement-learning baselines.
The ChatGPT help page says account switching works on ChatGPT web and is rolling out on iOS and Android. Only one account is active at a time: chats, memory, files and settings follow the selected account, and switching does not move chats or share a subscription. A second account costs nothing to create and starts on the Free plan. Organizations may turn it off for managed workspaces.
Microsoft says it will read a shared AI Training preference in robots.txt, taken from an IETF Internet-Draft, as it crawls from February 1, 2027, and lets single pages override it. It adds a separate page-level AI Grounding preference: a new DisallowExtended value limits use to Microsoft's own experiences such as Copilot and excludes grounding services it provides to third parties. NoArchive and NoCache stop counting for training or grounding after March 1, 2027.
TRAI's draft guidelines would make telecom operators classify each AI system as high, medium or low risk, under an internal AI policy approved by their board. Every AI incident would get a severity rating, and critical- and high-severity incidents, including material malfunctions and recurring failures, would be reported to TRAI. Operators would stay accountable for AI systems they buy from third parties. The draft is open for public comment.
DeepSeek published dsh-superagent, a research prototype of a proactive coding agent for one GitHub repository, packaged for DeepSeek Harness. When an allowed person asks for work in GitHub or, optionally, Feishu, it opens a work item, works in its own session and directory, and answers in the same conversation, for example with a pull request. A web panel shows what needs a person. DeepSeek says it is not production-ready.
agent-desktop, from Vercel Labs, is a command-line tool and C library for macOS, Windows and Linux. A snapshot returns JSON with a ref for every element an agent can act on; the agent acts on a ref, then snapshots again to check the result, with no pixel matching. When the UI changes, an action returns a stale-ref error instead of hitting the wrong element. By default, actions don't move the pointer or take focus.
Stageshow is a skill that agents such as Claude Code, Codex and Gemini read before making mock product interfaces and demo animations. The agent starts from the closest example, builds the demo, films it, then fixes what it finds, and the result is an ordinary React component. A playground shows the same demo done badly and done well side by side. Its author says most demos still take a few rounds of feedback.
You’re caught up.
That was all 12 stories in the 1 November issue.