Is consciousness something AI might discover on its own?

We asked @hamandcheese, chief economist at @joinFAI and former neuroscience researcher.

"There's this notion of universality in machine learning... you train different models and they converge in very similar representations. LLMs show tight functional alignment to the language centers of the human brain."

"As long as you accept that consciousness is something the brain does... if neural networks learn similar representations under similar training pressures, they may discover consciousness as a solution for a particular problem."

"Even a little bit of instruction tuning in an LLM improves alignment between the model and the cognitive control regions of the brain, parts we use for directing our attention."

"As models become more agentic, is consciousness load-bearing for being coherent over time? If so, it may re-arise as an efficient solution for AIs as well."