Just because a model is bad at one simple thing doesn't mean it can't be stellar at another. You think Einstein never made a typo or was great at Chinese chess?
LLMs can invent things which aren't in their training data. Maybe its just interpolation of ideas which are already there, however it's possible that two desperate ideas can be combined in a way no human has.
Systems like AlphaProof run on Gemini LLM but also have a formal verification system built in (Lean) so they can do reinforcement learning on it.
Using something similar AlphaZero was able to get superhuman at GO with no training data at all and was clearly able to genuinely invent.
![[personal profile]](https://www.dreamwidth.org/img/silk/identity/user.png)
vaa) wrote in ![[community profile]](https://www.dreamwidth.org/img/silk/identity/community.png)
lithiumtheory2024-11-21 09:40 pm
no subject
Just because a model is bad at one simple thing doesn't mean it can't be stellar at another. You think Einstein never made a typo or was great at Chinese chess?
LLMs can invent things which aren't in their training data. Maybe its just interpolation of ideas which are already there, however it's possible that two desperate ideas can be combined in a way no human has.
Systems like AlphaProof run on Gemini LLM but also have a formal verification system built in (Lean) so they can do reinforcement learning on it.
Using something similar AlphaZero was able to get superhuman at GO with no training data at all and was clearly able to genuinely invent.