Monday, November 2, 2026 · Europe/Bucharest · Sources: Hacker News, X, tech press, vendor blogs (YouTube not checked today)
A 1.6T-parameter MoE (about 48B active per token) trained with large-scale RL on long task sequences, and it turns real agent failures into reproducible training tasks. It's live in Meituan's chat and API now, with GitHub and Hugging Face releases promised.
This multimodal 9B model runs a shared block of 24 decoder layers twice to get extra depth without extra weights, with 262K-token training sequences and tool-use data. Weights, inference and training guides are all published.
A hidden Claude Mythos 5.5 label showed up in Anthropic's site source and gpt-6.2-astra appeared in OpenAI CLI telemetry. Chatter points to releases after the Nov 4 US midterms; nothing is confirmed by either company.
The paper argues different models converge on shared representations and shows how JEPA-style training strengthens that effect. It's one of the more-discussed research threads on HN today.
GitHub data cited on X shows agent-authored PRs rose from 17% in March to over 40% in September, about 110M that month. The practical takeaway: give coding agents their own limited-permission GitHub accounts so every change is a reviewable, traceable PR.
Nolan Lawson looks at how framework choice affects how well coding agents perform and how many tokens they burn. A useful read if your team leans on agents for frontend work.
Experiments with 50-agent teams trying to build a civilization stalled out in coordination overhead. It's a fun but real data point on the limits of naive multi-agent setups.
Instead of force-closing a slow watcher that fills its buffer, the apiserver marks it unsynced, catches it up from shared watch-cache history and resumes it on the same connection. It targets alpha in v1.38 behind the WatchCacheStallResume gate, with no client changes needed.
Plan upgrades accordingly: the monthly patch drop is on Nov 10, the week KubeCon NA opens in Salt Lake City (Nov 9 to 12).
Each workload gets its own microVM and kernel, with live snapshot/restore and copy-on-write forking of running VMs for parallel agent or RL runs. It targets teams that want managed-sandbox features without leaving their clusters.
An attacker reached environment variables for a subset of orgs on the background-jobs and agent platform; the exploit is blocked and affected owners were emailed. If you run agents or tasks there, rotate the secrets stored in it now.
Robert Haas says AI-assisted bug hunting pushed the release to November or later, and features like SQL/PGQ and FOR PORTION OF were reverted late in beta. A concrete look at how AI bug-finding is changing big open-source release cycles.
Tao argues the field needs to adapt its practices to AI and formal tools now that they're producing real results.
Status monitoring logged an ongoing GitHub services disruption and about 6h18m of Copilot problems on Nov 1 (UTC). Worth checking status before blaming your CI.