Anthropic study finds Claude models develop a bottleneck similar to human working memory
Anthropic says Claude has a small internal "J-space" that helps with reasoning, and researchers on X are treating it as unusually strong evidence for something like model working memory.
1262.7K3231K232.3K
“I've long suspected that models have some kind of "working memory" to store intermediate variables during a forward pass and IMO this paper has the best evidence yet”
Neel Nanda@NeelNanda5TECH#254Story Brief
Anthropic says new research found a small internal workspace inside Claude, which the company calls J-space and compares to the brain's global workspace. In Anthropic's framing, only a limited slice of the model's activity becomes available for reasoning and verbal report, while most processing stays automatic. Researcher Neel Nanda, who said Anthropic asked him to review the paper, called it the "best evidence yet" for models having a kind of working memory to hold intermediate variables during a forward pass, and linked the paper.
“Super cool findings that are likely to be informative for thinking about which aspects of GWT or access consciousness more generally are necessarily features of consciousness per se rather than broader computational features that consciousness is built upon”
