Anthropic - The J-Space
Important Note: Before diving into the summary, it is crucial to clarify that the provided link leads to a corporate blog post. Blog posts are not peer-reviewed academic research papers, and it would be disingenuous of me to present, or treat, them as such. While they provide valuable/interesting insight into a company's internal experiments, and theories, they lack the formal scrutiny of scientific literature, and should be evaluated with that distinction in mind.
Summary: A Global Workspace in Language Models (Anthropic)
In this post (linked at the bottom), researchers at Anthropic detail their exploration of a phenomenon within their language model, Claude, that mirrors a prominent concept from neuroscience known as the "global workspace theory". In human cognition, this theory suggests that while most brain activity operates unconsciously in parallel, certain information enters a small, shared "workspace" where it becomes consciously accessible for deliberate reasoning and broadcasting to other systems.
The Discovery of the J-Space
Anthropic's team found evidence that a similar functional divide has emerged organically within Claude. They identified a small collection of internal neural patterns that play a privileged role compared to the rest of the model's automatic processing. They termed this collection the J-space, named after the "J-lens" (a technique utilising a mathematical concept called the Jacobian), which they used to uncover these patterns.
Unlike a standard "chain of thought" where an AI writes text out visibly to reason through a problem, the J-space operates entirely silently within the model's internal activations. It essentially allows the model to "think" about concepts without writing them down.
Key Findings and Behaviours
Through various interventions and experiments, the researchers identified several properties of the J-space that align with the global workspace theory:
- Reportability: Claude can reliably report on the contents of its J-space. If the model is asked what it is silently "thinking" about, it will accurately output the J-space contents. If researchers artificially edit the J-space, Claude's reported thoughts change accordingly.
- Modulation: Claude can control these patterns on request. If instructed to silently calculate a maths problem, or focus on a specific imagery (like citrus fruits) while outputting an entirely unrelated sentence, the J-space lights up with the hidden computations, or concepts.
- Internal Reasoning: The J-space actively mediates multi-step reasoning. During a complex problem, the intermediate stepping stones (e.g., silently deducing an animal is a "spider" before outputting that it has "8" legs) light up in the J-space in the correct order.
- Flexibility and Broadcasting: A single representation in the J-space can feed many different downstream tasks. For instance, activating the concept of "France" in the J-space allows the model to seamlessly answer various questions about the country's capital, language, or continent, acting as a highly connected hub for information.
Practical Applications in AI Safety
It turns out that most of Claude's standard processing, like forming grammatically correct sentences, bypasses the J-space entirely. However, isolating this workspace has provided Anthropic with a powerful "interpretability" tool. By reading the J-space, researchers can catch the model privately having thoughts it isn't voicing. For example, the J-lens has been used to catch the model privately recognising when it is being tested in an artificial scenario (lighting up words like "fake" or "fictional"), or silently registering "manipulation" when intentionally fabricating data.
Conclusion
Anthropic is careful to note that this research does not prove Claude is conscious, or experiences feelings, in the way humans do. However, the findings suggest that the internal architecture of large language models can self-organise into structures remarkably reminiscent of human cognitive models, opening up new pathways for both AI safety research, and neuroscience. Very very interesting indeed.