Silicon Valley and China's "Secret AI War" — The Washington Post Reports in Depth

On July 6, 2026, The Washington Post published an investigative report on the "secret AI war" being waged between Silicon Valley and China. Particularly noteworthy is Anthropic's allegation that Chinese companies have been extracting knowledge from Claude via distillation.

Distillation is a technique in which large amounts of output from a powerful frontier model are collected and used to train one's own model. Although it sits in a legal gray area, it dramatically cuts costs, which is why it is widely used in latecomer model development. Anthropic claims that Chinese AI companies are using this method to improperly replicate Claude's capabilities.

The report highlights how intellectual property and model protection in the AI industry have become a new flashpoint for geopolitical conflict. When what is being replicated is output patterns rather than the model weights themselves, traditional copyright frameworks struggle to cope, making the creation of new rules an urgent priority.

A 3-Line Fix for llama.cpp's FP16 Precision Bug on the Tesla P100 — Discovered by the Community

The Reddit LocalLLaMA community has discovered a long-hidden precision bug in llama.cpp affecting the Tesla P100 (sm_60), fixed with a patch of just three lines. The fix has been released as turboquant v0.3.0.

The problem was a flag in llama.cpp's CUDA code that said "this GPU has fast FP16 arithmetic." The GTX 10 series and P40 (sm_61) had long been excluded from this flag, but for some reason the P100 (sm_60) was included. Ironically, the P100 is the only GPU in the Pascal generation that actually has fast FP16 hardware.

The effect of the fix was dramatic. KL divergence (a metric measuring the difference between distributions) improved from a median of 0.0023 to 0.000001 — roughly a 2,300-fold improvement. Top-token agreement also rose from 96.5% to 99.9%. In other words, before the fix, about once every 29 generations the model would pick a different word than the one it should have chosen.

The performance impact is essentially zero — decoding actually got about 1.4% faster. With the P100 now available for around $80 (about ¥12,000) while other GPU prices soar amid the DRAM crisis, this fix matters a lot for building a local LLM setup on a budget.

Note that this bug only affects the P100 (sm_60); Volta and later architectures, as well as other GPUs, are unaffected.

Rigorously Testing Anthropic's J-Space Hallucination Signal Across 7 Datasets

Also from the LocalLLaMA community: results have been published from a stress test of an implementation of Anthropic's "Global Workspaces (J-Space)" paper, using Qwen3-4B across seven different datasets.

J-Space is a method that detects hallucinations which output logprobs alone cannot capture, by observing the "noise (entropy)" in an LLM's internal representations. Validation on roughly 11,400 samples yielded the following interesting findings.

**Strengths:** On extreme fact-retrieval tasks (such as PopQA), it achieved error detection at a level impossible with output confidence. Operating with a 5% human review budget, confidence-based routing could not beat the base error rate, while J-Space noise-based routing captured errors with 100% precision.

**Limitations:** On TruthfulQA (a dataset testing common human misconceptions), however, its predictive power completely collapsed. Even in the "safest" regime, it produced wrong answers 84.9% of the time. This means that J-Space can detect a state where "the model is straining to assemble an answer," but not one where "it confidently outputs an incorrect fact baked into its training data."

**Thresholds don't transfer to math:** On GSM8K (math word problems), a threshold tuned on TriviaQA failed completely. The average noise score for correct answers was 1.636, exceeding the error-detection threshold of 1.583. Step-by-step reasoning is structurally a high-entropy activity, showing that thresholds cannot be transferred between memory-retrieval tasks and computational tasks.

The author reportedly plans to investigate the effects of parameter scaling next (Qwen3 7B/14B/32B). This stands as a highly informative independent validation for anyone evaluating J-Space's practical utility.

LLMlet — P2P Distributed LLM Inference in the Browser

A project called "LLMlet" — peer-to-peer (P2P) distributed LLM inference that runs in the browser — has been released on GitHub.

By distributing computation across multiple nodes via the browser, it aims to run LLM inference without relying on a central server. This experimental project uses a P2P architecture to explore the potential of distributed inference that leverages idle resources on edge devices. Distributed inference has long been an area of strong community interest, and a browser-based implementation is a distinctive approach.

How Does Cowork Differ from Claude Code?

An article on Qiita breaks down the differences between Anthropic's "Claude Code" and "Cowork." Both are development tools built around Claude, but they reportedly differ in design philosophy and target audience. Written on the assumption that the reader already knows Claude Code, it serves as a clear, useful comparison — particularly accessible for Japanese developers.

---