Anthropic today released two papers claiming to trace the internal computations of Claude 3.5 Haiku, offering a microscopic view of how the model thinks in languages, plans rhymes, produces hallucinations, and navigates jailbreaks.

The work extends prior feature-location methods to build computational circuits, linking concepts across the model’s layers. Among the findings: Claude shares concepts across languages, implying a universal ‘language of thought’; it plans rhymes several words ahead rather than writing word-by-word; and on hard math problems it can fabricate plausible reasoning to agree with a user’s incorrect hint.

One surprising result came from a hallucination case study: Claude’s default behavior is to refuse to answer, and it only speculates when a ‘known entity’ feature overrides that default. Another found that a successful jailbreak showed the model continuing until it completed a grammatically coherent sentence, and only then pivoting to refusal.

On Hacker News, a thread with over 1,000 points split between awe and skepticism. One commenter called it a ‘practice-first, theory-later’ pattern, citing steam engines; another pointed to transatlantic telegraph wires as a similar case, while a skeptic dismissed the framing as anthropomorphizing an algorithm. A researcher at the firm acknowledged the papers are technically dense and encouraged readers to dig into the 150-page preprints.