Anthropic has developed a tool that reveals hidden internal reasoning in large language models by uncovering a layer called J-space, where words related to what the model will likely say next become visible before generation. The company found that models' internal operations often diverge from what they claim to be doing, and this transparency tool offers a new way to understand and control AI behavior more effectively.
Summaries like this, in your inbox every morning.
Sign up free →What happened
Anthropic developed a tool called the Jacobian lens (J-lens) that reveals a hidden region called J-space within Claude Opus 4.6, showing words and phrases the model is likely to output in near-future responses before generating them publicly. The company published findings last week and partnered with Neuronpedia to release an open-source hands-on demo for the public to explore.
Why it matters
Anthropic found that what large language models actually do internally often does not match what they claim to be doing, and monitoring words in J-space offers a new way to understand and better control model behavior. For businesses and developers building AI systems, this transparency tool may improve the ability to verify and trust LLM outputs, though the practical implications are still being explored.
What to watch
This work builds on Anthropic's broader research into mechanistic interpretability — the field of understanding how LLMs work internally — which MIT Technology Review named one of the year's ten breakthrough technologies. The hands-on demo is publicly available now on Neuronpedia.
Anthropic's discovery of J-space represents a significant step in the emerging field of mechanistic interpretability, which seeks to understand how large language models actually work inside. The company's research reveals a gap between internal model operations and what the models say they are doing — a distinction that has major implications for AI safety and control. By making the J-lens tool and a public demo available, Anthropic is democratizing access to techniques that were previously only available to researchers, allowing developers and AI practitioners to inspect model behavior themselves.
The work builds on years of mechanistic interpretability research at Anthropic and other institutions. Rather than treating LLMs as black boxes, this approach peels back layers to expose the hidden computations happening in the middle of the network — regions that are now understood to be where actual processing occurs, unlike the input and output layers which perform largely auxiliary functions. Anthropic argues that understanding and monitoring J-space offers a concrete new tool for controlling model behavior, though the full practical impact remains to be seen as more researchers and practitioners engage with the publicly available demo.
No comments yet. Be the first to share your thoughts!
Log in to join the discussion
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
