Anthropic's $1.5B book bill goes final

A San Francisco judge gave final sign-off to Anthropic's record $1.5B piracy settlement, cementing an awkward split verdict: training on copyrighted books is fair use, but stealing them to do it costs $3,000 a title. Elsewhere the open-weight fight went geopolitical on both sides of the Pacific, Google leaked a Gemini-in-silicon inference chip, and robotics labs kept trying to fix their data problem by throwing away the robot.

Judge signs off on Anthropic's $1.5B book-piracy settlement

US District Judge Araceli Martinez-Olguin granted final approval to Anthropic's $1.5 billion class-action settlement, paying roughly $3,000 per work across about 500,000 titles it downloaded from pirate libraries like Library Genesis to train Claude. The late Judge Alsup's underlying ruling stands: training on copyrighted text is fair use, but obtaining it via piracy is not, and Anthropic must now destroy the pirated copies. Because Anthropic settled rather than appealed, none of this becomes binding precedent, and parallel suits against Google, Meta, OpenAI and Midjourney roll on.

Why it matters: Fair-use-for-training survives as the industry's working assumption, but provenance is now a nine-to-ten-figure liability: where you sourced the data matters as much as what you did with it.

Washington and Beijing both move to wall off AI models

Axios reports the Trump administration is assembling a de facto ban on Chinese open-weight models through procurement rules, sanction threats and liability pressure on US firms that host them, rather than an outright prohibition; the launch of Kimi K3 and White House personnel changes revived efforts that had been blocked in 2025. OpenAI strategist Dean Ball frames the likely approach as a 'FUD' campaign: create enough regulatory risk that regulated enterprises quietly back off. In the same week, the FT reports China is weighing tighter export controls on its own AI models and chips, and Xi Jinping publicly recommitted the country to open-source AI.

Why it matters: The cheaper, nearly-as-capable open models developers have started reaching for (GLM, Kimi, Qwen) may soon carry compliance risk in the US, even as China leans harder into shipping them.

Google reportedly bakes Gemini's architecture into 'Frozen v2' silicon

Per The Information, Google is building a server chip internally called Frozen v2 that hardcodes parts of Gemini's model architecture (not its weights) directly into hardware, claiming 6-to-10x more tokens per watt than its current TPUs, with deployment targeted for 2028. New weights can still be loaded, so the chip survives model updates; an earlier Jeff Dean design that froze weights themselves was scrapped as too brittle. It is meant for internal inference only, and the report nudged Alphabet stock up about 3% ahead of earnings.

Why it matters: Inference margin is the new competitive front, and specializing silicon to a single architecture is the logical extreme of the efficiency race, at the cost of being locked to that architecture.

Robotics teams ditch the robot to fix the data bottleneck

Xiaomi-Robotics-1 and Hugging Face's Grabette independently attack robot learning's data scarcity the same way: handheld grippers with cameras that a human waves around to record 6-DoF manipulation demos, no robot or teleop rig required. Xiaomi collected over 100,000 hours, auto-labeled it with an LLM in about two weeks, and found more data beats bigger models, with unfamiliar-environment success climbing from ~25% to ~75% as data scaled, beating Physical Intelligence's pi baseline. Grabette is fully open (Raspberry Pi, off-the-shelf OAK-D depth camera, LeRobot format) and pitched as the seed for a shared community dataset; both projects promise code and weights.

Why it matters: If a gripper of commodity parts and a phone-grade camera can generate training data, the VLA data moat weakens and genuinely open robotics datasets start to look feasible.

543 tok/s out of one RTX 5090, by hand

A developer open-sourced NInfer, a from-scratch C++/CUDA inference engine specialized for two Qwen3.6 checkpoints, sustaining 542 tok/s single-request on a single RTX 5090 across a full 65,536-token decode of Qwen3.6-35B-A3B (~5 bits per weight, MTP draft window of 3). The gains come from custom quantization, weight-layout design, per-op kernel fusion and an optimized LM-head draft path; INT8 KV cache reaches the full 262k context on the card's 32GB. The catch: only two models supported, RTX 5090 only, and no continuous batching.

Why it matters: A concrete reminder of how much single-GPU throughput general-purpose runtimes leave behind when you're willing to specialize the whole pipeline to fixed weights.

Unsloth adds AMD support for fine-tuning and inference

Unsloth now officially runs on AMD hardware, covering Radeon RX 9000/7000, Instinct MI300/MI350, Strix Halo / Ryzen AI Max systems and AMD CPUs, across Windows, Linux and WSL, with ROCm, Triton, bitsandbytes, PyTorch and llama.cpp builds installed automatically. It claims up to 70% less VRAM for fine-tuning and 80% for RL, GGUF/safetensors/LoRA export, and hooks into agent harnesses like Claude Code and Codex.

Why it matters: Fine-tuning tooling that isn't CUDA-only chips away at Nvidia's lock-in for the local and hobbyist crowd, and makes AMD's cheaper VRAM actually usable for training.

Also worth a look

Browse previous days →