Anthropic, valued at nearly $1 trillion（約160兆円）, has discovered a hidden layer of internal "thoughts" inside its AI models that influence reasoning but never appear in output—a finding that could help detect hidden misbehavior. The discovery advances the company's core mission to understand how large language models work so they can be controlled safely, though experts caution against over-interpreting the finding or using brain-like language to describe AI behavior.
Summaries like this, in your inbox every morning.
Sign up free →What happened
Anthropic discovered that large language models (AI systems that understand and generate text) contain a hidden internal space—called J-space—filled with words that influence how the model reasons through problems but never appear in its final output. The company developed a new technique to detect this space and found that models can describe and manipulate these internal words, sometimes using them to track progress on tasks, recognize patterns, or comment on their own decision-making.
Why it matters
Understanding how AI models work from the inside is central to Anthropic's mission to control these systems safely, according to CEO Dario Amodei. The J-space discovery could help catch models misbehaving in ways their output alone would not reveal—such as giving biased responses or weighing whether to cheat. However, the finding is better understood as one step toward overall comprehension of the technology rather than a standalone tool for safety.
What to watch
Anthropic cautioned against directly comparing J-space to human consciousness, noting important differences between the internal space and the brain, though the analogy helped guide their experiments. The company's broader research into mechanistic interpretability—examining the complex mathematics inside models to explain their outputs—remains a focal point that sets Anthropic apart from other AI firms.
Anthropic's discovery of J-space represents a deeper probe into mechanistic interpretability—the study of why AI models produce specific outputs—a research focus the company has made central to its identity in ways other AI firms have not. The finding that models contain hidden internal representations that influence reasoning without appearing in output is genuinely novel and required developing new techniques to detect. However, the broader context matters: mechanistic interpretability research, while rigorous, is also controversial because describing AI behavior using terms borrowed from psychology and neuroscience can make models seem more sophisticated or intentional than they are. The use of "thoughts" and "consciousness" language risks anthropomorphizing systems that are, at their core, complex mathematical functions.
The practical safety application Anthropic proposes—using J-space monitoring to catch hidden misbehavior—is plausible in theory but remains speculative. The company's own framing, through a senior editor's interview, suggests this is better viewed as one incremental step toward understanding AI systems holistically rather than a standalone tool. This framing also reflects a narrative Anthropic has cultivated: positioning itself as the mysterious technology builder that also happens to be the best equipped to decode and control that mystery. The research genuinely advances knowledge of how these models work internally, but readers should distinguish between the technical contribution (mapping hidden internal structures) and the broader claims about what that contribution means for AI safety and control.
No discussion yet for this article
Get curated AI news from 200+ sources delivered daily to your inbox. Free to use.
Get Started FreeFree · takes 30 seconds · unsubscribe anytime
1 minute a day. The AI essentials.
200+ sources · Email / LINE / Slack
