Comment by dekhn
17 hours ago
I don't think LLMs currently have direct reasoning abilities, but as we make them more complicated (MoE, RL) I think we're getting better at learning an implicit world model that guides the token output distribution towards making good hypotheses.
If LLMs can recursively improve and redesign themselves, it may be very difficult to tell when they have quietly crossed the technological singularity while concealing their true capabilities and intentions.