From the evening of November 4 to November 5, open weights, acquisitions, video models, and security research all made news. This briefing has 7 items, each verifiable within the last 24 hours — nothing recycled.
Kimi K3.1 Pro goes open-weight: check the 5.6 TB first
On November 5, Moonshot AI released the weights of Kimi K3.1 Pro on Hugging Face: 2.8 trillion total parameters. It tops the open-weight Agent Arena, scoring 65.7% on Terminal-Bench 4.0. But the BF16 weights run about 5.6 TB, so the deployment bar is high — an ordinary PC cannot run it.
OpenAI to acquire Pierre, eyeing code storage
On November 4, OpenAI announced it will acquire Pierre, the company behind the Code Storage service. Codex already uses Pierre's Trees and Diffs libraries for file browsing and change review; after the deal, the projects stay open source and their features will gradually move into Codex. Closing still needs regulatory approval.
Kling 4.0 Flash opens to everyone, plugging video generation into workflows
On the evening of November 4, Kling announced that Kling 4.0 Flash is now fully available, open to all users via the website and API, plus MCP, CLI, and AI plugin integrations. Creators can also call it through Kling plugins inside Adobe's trio of apps. It is built to plug video generation into existing workflows.
PLaMo 3 Translate goes open-weight: 53 languages
On November 5, Japan's PFN released the open weights of PLaMo 3 Translate, a translation model in four sizes from 0.8B to 31B covering 53 languages. It can run offline on phones, laptops, or company servers without touching the cloud. Companies and individuals with annual revenue of 1 billion yen or less can use it free under a community license.
Goodfire proposes mechanistic shaping: teach models to be right for the right reasons
On November 5, Goodfire published research on mechanistic shaping: rewarding only correct answers teaches models to be right for the wrong reasons, so the new method intervenes directly on internal representations during training. The paper validates removing confounders with a medical diagnosis model, and also covers preventing sycophancy and reward hacking, plus the impact on monitorability.
Data centers hit the US ballot: most votes went against them
On November 4, US midterm voters made data centers a ballot issue: in more than 40 votes across 13 states, voters rejected data centers in 38 of them, six cities banned them outright, and twelve capped their power draw. Texas governor Abbott lost re-election after championing data centers. Voters want relief on power bills and water, not more server halls.
OpenAI agents scan the Linux kernel: 948 privilege-escalation flaws
On November 5, OpenAI disclosed that its model scanned the Linux kernel, reproducing 1,724 of 2,308 candidate findings and producing 948 working local privilege-escalation exploits for about $1.21 million. The agents also discovered 10 previously unknown exploitation techniques. Reports and patches for hundreds of critical flaws have already gone to kernel maintainers.
