Building information perpetual motion machines? Reinforcement learning! Why not in 2023?

@xwang_lk Thanks! NLP/LLMs are very different from AlphaGo.

@garrytan LLMs are different from AlphaGo.

@xwang_lk RLHF is inverse RL, so imitation learning. Diff from "learning from demonstration" though, which is supervised learning, cf "supervised fine-tuning". RLHF is not SL. RLHF is a principled approach to problems w/o a reward function. Most NLP problems w/o objective objectives.

Paper by John Schultz, Jakub Adamek, @MatejJusup, @sharky6000, Michael Kaisers, @sarah_perrin_, Daniel Hennes, Jeremy Shar, Cannada Lewis, @anianruoss, @TZahavy, @PetarV_93, Laurel Prince, Satinder Singh, @ericmalmi and @weballergy

@MatejJusup @GoogleDeepMind @ADarmouni

@ADarmouni @GoogleDeepMind

Title: Mastering Board Games by External and Internal Planning with Language Models Speaker: John Schultz, Deepmind Time: Jan 16, 2-3 pm EST Pls mark your calendar!

@zdhnarsil I think PRM is a misuse of terminology: No need to differentiate PRM and ORM, just RM. It should be value function or may be reward shaping. Something wrong in the way to define PRM, like (1,0,-1). A short blog hybrid in Chinese & English.

@TonyZQin Title: Building Task-driven Conversational Agents for Business Phone Operations

Title: Building Task-driven Conversational Agents for Business Phone Operations Speaker: @TonyZQin Time: Jan 8, 5:30pm PT Welcome!

@natolambert Seems no exact solution (from the answers so far and AFAIK). Can we say the numbers reported are "heuristic"? BTW, passing several tests, like with HumanEval, can not guarantee code correctness. So, many (all?) code generation papers are reporting "heuristic" results?

Title: Building Task-driven Conversational Agents for Business Phone Operations Speaker: @TonyZQin Time: Jan 8, 5:30pm PT Welcome! Please mark your calendar.

@denny_zhou Any LLM can *guarantee* accuracy, not mention optimality? Any LLM can beat AlphaZero on chess (without training data from an AI like AlphaZero)?

@omarsar0 Shouldn't the title be "LLMs are not good enough?" or "LLMs are not good enough for building (autonomous) agents"?

Reflection 2024, Guesstimation 2025

@sh_reya I talked about it. I am not an influencer though...

@denny_zhou @aidan_mclau Gradient decent is search, in continuous spaces.