Mastodon Feed
jonny@neuromatch.social ("jonny (nonvenomous)") wrote:
Anthropic published that p-space, "global workspace", steerable latent space thing - you can tell it doesn't really work because that would be something you would immediately want to productize, being able to do surgery on the internal state of a model would have immediate and obvious marketability in many of the cases where LLMs are most poorly suited. But if it did work, I would also give in and pay money to retranscribe all the LLM output with the LinkedIn voice vector nulled out.