GitHub blocks pull_request_target by default from December 2, AWS distills a shopping agent into a 4B model that matches its teacher, DeepMind's PlanAssist enters a UK council beta
No major model launches over the weekend; today is about the next step of running AI in production: safer CI defaults, small models taking over repetitive agent work, and AI drafts that a person signs off.
From December 2, 2026, pull_request_target events are blocked by default in all public repositories unless an applicable Actions event policy explicitly allows them; enforcement was previously scheduled for November 2
GitHub's reason: the trigger runs in the base repository's context with access to its secrets, so untrusted fork code could compromise the pipeline or steal secrets
Two brownouts on November 9 and November 16, 16:00–18:00 UTC: blocked runs fail with an error annotation and do not execute; outside those windows the rule stays in evaluate mode and workflows run with warnings
Applies to public repositories on GitHub Free, Team and Enterprise Cloud; private and internal repositories and GitHub Enterprise Server are not affected
Builder's takeAI review bots and agent workflows that comment on fork PRs are the ones most likely to sit on pull_request_target, so before November 9 I'd run Insights and list every run the rule would block. Where you really need it, allow specific workflow files rather than the whole trigger, and keep secrets out of any step that touches fork code. The brownouts land at midnight to 2 a.m. Beijing time on November 10 and 17, so warn whoever is on call.
GLM-5.3 on Bedrock generated 850 trajectories on ShoppingBench's 426 training problems, keeping only the solved ones; Qwen3.5-4B was fine-tuned with LoRA (rank 64) through serverless SFT on SageMaker AI
Average success on 324 held-out problems: base model 0.297, distilled student 0.541, teacher 0.540 greedy and 0.511 sampled, a medium-sized frontier model 0.546
General instruction following drops: IFEval prompt-strict falls from 0.819 to 0.741 and instruction-strict from 0.872 to 0.813
Keeping the reasoning alongside the actions is what carried the student most of the way; it serves on a single G-family instance (A10G/L40/L40S), and AWS publishes no per-task cost
Builder's takeThis is the clearest recipe I've seen for moving a fixed-format agent task off a frontier API: log trajectories, keep only the successful ones with their reasoning, and fine-tune a 4B model. I'd try it first on PandaClaws' highest-volume background steps, keep a general-ability regression check like IFEval, and only switch when traffic is steady enough that an hourly instance beats per-token billing.
Built with Gemini, together with the UK government, Google Cloud and Faculty; it validates documents and drawings, checks heritage constraints and planning policies, reviews past objections and drafts cited reports for officers to edit
Early pilots in Barnet, Camden and Dorset cut officer assessment time by up to 58%, and more than 70% of generated reports were approved with no edits
Twelve more councils join in three cohorts, for 15 in total; if testing succeeds, the government plans to offer it to all councils in England in 2027
Builder's takeThe pattern worth copying is AI drafts with citations, a person signs, and the metric is the share approved without edits. Our AI Interview product fits the same shape: the evaluation report should cite the exact answers it relies on, and I'd track the no-edit rate as the main quality number instead of a model score.
The Cosmos Framework, an open modular toolkit for training, reinforcement learning, optimization and inference, now covers NVIDIA Alpamayo, GR00T, Qwen, pi0.5 and other compatible Hugging Face models, with new DPO, GRPO, PPO, OPD and DiffusionNFT pipelines
Released checkpoints on Hugging Face include Cosmos3-Nano-Transfer-Auto, Cosmos3-Nano-Sim-Auto, Cosmos3-Nano-Sim-Transfer and Cosmos3-Nano-Sim-Bimanual; Transfer-Auto generates up to 400 frames at 480p or 720p
Cosmos3-Nano needs 58 GiB of GPU memory for generation in BF16, 44 GiB in FP8 and 31 GiB with FP8 plus offloading; speculative decoding gives roughly 2–3x mean speedups on H200
Builder's takeIf you don't build robots or driving systems, this one can wait. If you do, start with the Nano checkpoints in FP8 and read each model card first: the post itself states no license for the weights, the framework or the NIM containers.
LifeVal measures how well AI helps people complete everyday tasks
Examples range from finding childcare to managing finances and government paperwork
Builder's takeI've only seen OpenAI's summary, so I'll hold off on any scores until the full page is readable. The direction is right for products, though: judge an assistant by whether the errand got done. For AI Cloud Drive I'd keep a small eval set of real user errands, such as finding a contract and filling in a form from it.
Researched and drafted with AI assistance; editorial standards and views set by Darius.
Every Monday, the week in AI in one email
Every Monday morning: the 6 AI stories that mattered most in the past seven days, with my take, in your inbox.
Your email is only used to send the weekly and is never shared. Every email has a one-click unsubscribe link. See the latest issue →