New blogpost! Current LLMs think in words, but what if they might one day think in opaque vectors? Google DeepMindās recent DiffusionGemma a priori has the ability to do such ālatent reasoningā. Together with @NeelNanda5, we investigated whether it is in fact harder to monitor.
Jan Bauer
@japhba
Grappling with the ways of neural networks, both biological and artificial š§ š¤
163 Following 137 Followers 16 Posts