Fireship Takes On the 'Is Claude Conscious' Debate: Swap a Spider's Thought for an Ant's, and the Answer Changes

Anthropic's paper, 'A Global Workspace and Language Models,' is described by Fireship as the most philosophically cursed research ever published by a company that also happens to sell API tokens. The claim -- that buried deep inside Claude's incomprehensible pile of matrices is a small, organized set of neural patterns called J-space -- conveniently pushes the AGI-and-singularity narrative forward, though not everyone is buying it, the video notes, before diving into Claude's weird brain.

Anthropic researchers pinpointed exactly where 'private thoughts' get stored inside Claude's brain. Swap one thought for another, and the model's entire chain of reasoning obediently follows the lie. For good measure, they deleted the region entirely -- and, surprisingly, Claude kept speaking fluently and confidently in English.

The most important surprise: nobody designed this structure. It emerged on its own through training -- strikingly similar to the 'global workspace theory' proposed by Bernard Baars in 1988, the idea that the brain runs countless functions automatically in the background while consciousness is a single, brightly lit stage.

In one experiment, Claude was asked how many legs a web-spinning animal has. Just before answering 'eight,' the concept 'spider' lit up in J-space. Researchers then surgically swapped that internal concept for 'ant' -- with neither the prompt nor the output edited -- and Claude's answer changed to 'six.'

Fireship stresses that Anthropic itself states in the paper that none of this tells us whether Claude is conscious, while still finding it fascinating that a 'scratchpad for thoughts' like this can emerge spontaneously given enough data and the right linear algebra. Personally unconvinced Claude is conscious, the video closes on a wry note.