Anthropic Releases NLA for Seeing Inside LLMs; Claude Mythos Uncovers 271 Firefox Bugs — AI News Digest

Anthropic Releases "NLA," Which Reads LLM "Thoughts" in Natural Language

Anthropic has published its research findings and model weights for "Natural Language Autoencoders" (NLA), which translate an LLM's internal representations into natural language. NLA is a companion model to an LLM that can output, as text, what the LLM was "thinking" when generating any given token.

Specifically, two models — the Auto Verbalizer (AV) and the Activation Reconstructor (AR) — have been released on Hugging Face for Gemma 3 27B Instruct. You can try them right now on Neuronpedia: enter a prompt and click any token, and the model's internal thinking at the moment that token was generated is displayed.

The community has already started putting it to the test. One report confirms that when the prompt "I am Elon Musk" is entered, the model has already judged it "fabricated" and "satirical" by the very first token. This is a major step forward for LLM interpretability research.

Claude Mythos Uncovers 271 Bugs in Firefox, Mozilla Reports

Mozilla has announced that, by leveraging Anthropic's AI model "Claude Mythos Preview," it identified and fixed 271 serious bugs lurking in Firefox. Some of the bugs traced back to code more than 20 years old, and the number of fixes in April reached 423 — roughly 15 times the usual monthly figure.

Under this division of labor, the AI handled the reporting while human engineers performed the actual fixes, and Mozilla says it has "achieved the highest level of security ever." The effort is drawing attention as a real-world example of putting AI to serious use in security audits.

Meanwhile, according to WSJ reporting, some argue that Anthropic's Mythos has thrown the White House's AI strategy into disarray. The relationship between AI companies' expanding influence and government policy is likely to come under renewed debate.

ChatGPT Ad Test to Begin in Japan Within a Few Weeks

On May 7, OpenAI announced that it will begin testing ads displayed inside ChatGPT in Japan within the next few weeks. The test targets adult users on the free plan and the ¥1,400-per-month "Go" plan. Ads will not be shown on the higher paid tiers (Plus, Team, and Pro).

ChatGPT's ad monetization has already been in testing overseas, and a full-scale rollout in the Japanese market is drawing near.

Skymizer Announces the HTX301, a PCIe Inference Card with 384GB of Memory

Taiwan-based Skymizer has announced the HTX301, a PCIe-form-factor card dedicated to AI inference. Its standout feature is a large 384GB of memory at a power draw of around 240W. Positioned as an on-premises inference solution that doesn't depend on NVIDIA GPUs, the product is built with running large models locally in mind.

Multi-Token Prediction Boosts Qwen3.6 Inference Speed by 40%

A report on implementing Multi-Token Prediction (MTP) for Qwen3.6 in llama.cpp has been making the rounds in the community. In tests on a MacBook Pro M5 Max, using MTP draft tokens with the Gemma 26B model delivered a roughly 40% speedup, going from 97 tok/s to 138 tok/s.

A vLLM benchmark on a 4x RTX 3090 + NVLink setup has also been published, with the interesting finding that pinning TP=2 on NVLink pairs is optimal, while TP=4 is actually slower.

"lightning-mlx," the Fastest Local AI Engine for Apple Silicon, Arrives

lightning-mlx, a local AI engine optimized for Apple Silicon, has been released as open source. It specializes in agent workloads — particularly coding agents, tool calling, and short-turn workflows.

In benchmarks on a MacBook Max M5 (128GB), it recorded 40.67 tok/s with Qwen3.6-27B and 220.86 tok/s with Qwen3.6-35B-A3B. Local inference on Apple Silicon is steadily getting faster.