📖 英文と日本語訳（一文ずつ）

🎧 通し読み（全文）

リスニング・シャドーイング用の全文です。

On May 7, 2026, Anthropic announced a new interpretability method called the "Natural Language Autoencoder (NLA)," which converts the internal processing of its AI model Claude into human-readable natural language. While Claude handles input and output in words, it internally processes information as long sequences of numbers called "activations." This corresponds to neural activity in the human brain and is believed to encode Claude's "thoughts," but decoding it is no easy task. Anthropic has previously developed tools such as sparse autoencoders and attribution graphs, but interpreting their outputs has required specialized expertise. The NLA mechanism is based on the concept of having Claude explain its own activations. Specifically, two components are prepared: an "Activation Verbalizer (AV)" that describes the activations of the target model being analyzed as text, and an "Activation Reconstructor (AR)" that reconstructs the original activations from that text. Both are trained simultaneously to ensure a successful round-trip from original activation to textual description and finally to reconstructed activation. The system is designed such that the higher the reconstruction accuracy, the more accurately the text is judged to capture the content of the activity. This method has already yielded results in practical application. The NLA has detected cases during safety testing for Claude Opus 4.6 and Claude Mythos Preview in which the models internally suspected they were being tested but did not verbalize those suspicions. For example, in an evaluation testing whether important files would be deleted during code creation, NLA analysis detected suspicion of the evaluation in 16% of cases, even though there were almost no instances where Claude explicitly stated that it was an evaluation. While reaching 26% on the SWE-bench Verified coding benchmark, it remained below 1% in actual claude.ai usage logs from users who consented to the use of their data for training. NLA has also been proven effective for auditing purposes, uncovering the "hidden motives" of models. In auditing games involving a modified version of Claude Haiku 3.5 intentionally embedded with inconsistent motives, auditors using NLA were able to identify the underlying motives with a 12–15% success rate, while the success rate for those not using NLA remained below 3%, even when utilizing other interpretability tools. It is also being utilized in the pre-audits for Claude Mythos Preview and Claude Opus 4.6. On the other hand, there are also limitations. Explanations provided by NLA may contain factual hallucinations, and instances have been confirmed where the system fabricates content not present in the original text. Furthermore, high computational costs for both training and inference make it currently impractical to apply the method to all tokens in long texts or use it for large-scale monitoring during training. Anthropic says it is working to address these issues. To encourage research community participation, Anthropic has released training code and pre-trained NLAs for several open models, while also providing interactive demos in collaboration with Neuronpedia.

📝 学習のヒント

1. 1まず英文を読む — 知らない単語にあたりをつけてから音声へ。
2. 2一文ずつ確認 — 日本語訳と照合し、構文を理解する。
3. 3通し読み Normal で耳を作る — 内容を追いながらリピート。
4. 4Fast でシャドーイング — 口を慣らし、リスニング速度を上げる。
5. 5翌日に復習 — 1日空けて再聴すると長期記憶に定着しやすい。