Today’s brief covers Anthropic rejecting the claim that it wants open weights banned, Nvidia making its second open-weights move in days, Taiwan detaining an Nvidia employee over allegedly forged chip paperwork, and AI arriving on both sides of security, plus the agent plumbing and model architecture actually driving the gains.
The Open Weights Fight Escalates: Anthropic, Nvidia, and Moonshot Stake Their Claims
Anthropic pushed back hard on claims it wants open models banned, Nvidia rallied dozens of companies around a security argument for openness, and Moonshot backed openness with real infrastructure, not just weights.
- Anthropic Rejects Ban Claim, Counters Nvidia’s Open-Weights Letter. Dario Amodei said Anthropic never asked regulators to ban open-weight models and laid out its actual asks: chip export controls, limits on distillation, and mandatory safety testing.
- Nvidia’s Third Open-Weights Push in a Week: A Security Alliance. Nvidia and 36 other companies formed the Open Secure AI Alliance, arguing openly available models make better cyber defenders than closed ones, its second open-weights move in days.
- Moonshot Ships Kimi K3 Weights, Then Opens the Machinery Behind Them. Moonshot released Kimi K3’s weights and technical report along with the kernels and libraries needed to actually run the 2.8 trillion parameter model.
AI Turns to Security, in Both Directions: Defenders Build Models, Attackers Get Agents
Today’s security news split cleanly between offense and defense: Microsoft and Vercel tried to make scanning better, while Cogent built an agent that automates the attacks scanners are supposed to catch.
- Microsoft Bets a Purpose-Built Model Beats General AI on Vuln Hunting. Microsoft says its MAI-Cyber-1-Flash model now handles most vulnerability work inside its internal security tool, beating general-purpose models on cost by Microsoft’s own benchmarks.
- Vercel’s DeepsecBench Shows Top Scanners Still Miss Most Bugs. GPT-5.6 Sol topped Vercel’s new security benchmark, but even the leading model caught under a third of the known code vulnerabilities in the test set.
- Cogent Built an AI That Chains Enterprise Attacks on Its Own. Cogent says its VR-1 model more than doubles rival systems at completing multi-step intrusions, though the company graded that claim on a benchmark it built itself.
Nvidia’s Two Faces: Backing a Trillion-Dollar Bet While a Chip Probe Widens
The same week Nvidia poured billions into Ilya Sutskever’s stealth lab, Taiwanese prosecutors detained one of its own employees in a widening chip smuggling investigation.
- SSI Takes Nvidia’s Billions, Tests Its Own Purity Pitch. Ilya Sutskever’s Safe Superintelligence struck a multibillion-dollar Nvidia compute deal, one Bloomberg values at $5 billion, without a disclosed product or timeline attached.
- Taiwan Detains Nvidia Employee, Widening AI Chip Smuggling Probe. Taiwanese prosecutors detained an Nvidia staffer surnamed Chang and searched his desk at the company’s Taipei office, the first such action against Nvidia staff in the probe.
Building the Agent Stack: The Infrastructure Underneath Catches Up
Away from the headline fights, four teams shipped the unglamorous infrastructure that decides whether agents and fine-tuned models actually work in production.
- Ramp Labs Open-Sources PorTAL, a Portable LoRA Framework. PorTAL trains one shared task representation and exports ordinary LoRA adapters across several frozen base models, aiming to end the retrain-on-upgrade cycle.
- Google Lets Gemini Flash Inherit Pro’s Reasoning, Not Just Its Text. Google Cloud’s new Distillation Service trains cheap Flash models on Pro’s internal reasoning, though the transfer currently works only within Gemini’s own model family.
- Nvidia’s Molt Trains AI Agents as Code, Not Config Files. Nvidia’s NeMo team open-sourced a PyTorch-native reinforcement learning framework that writes agent behavior in Python instead of YAML, scaling to trillion-parameter models on large clusters.
- Agent Autonomy Is a Property of the Task, Not the Model. PostHog argues how much you can safely delegate to an agent depends on a task’s verification cost and reversal cost, not on which model is running it.
Model Architecture Notebook: What’s Actually Driving Capability Gains
Three technical threads today looked past benchmark scores to ask where model capability actually comes from, from Kimi K3’s memory design to diffusion language models crossing 100 billion parameters.
- A Baseten Engineer Traces Why Kimi K3 Beats Scale Alone. An inference engineer’s lineage from GPT-2 to Kimi K3 argues the model’s gains trace to its memory architecture rather than raw parameter count.
- Looped Transformers Decouple Parameter Count From Capability. An explainer breaks down looped transformers, which reuse the same layers repeatedly, trading fewer stored weights for more compute per pass.
- InclusionAI’s LLaDA2.2 Pushes Diffusion Language Models Past 100B. Ant Group’s InclusionAI team shipped an editing-focused update to a diffusion language model family it had already scaled past 100 billion parameters.