Yann LeCun is making the case that generative models are fundamentally divergent: at every token, there is a probability of getting something wrong, and errors accumulate exponentially over the number of generated tokens.<p>I tend to agree with the premise, however, what if the generative process is overlaid with an "inner debate", as a substitute to having the model play against itself, ala AlphaGo Zero?<p>The sequence of prompts would go:<p>1. Please explain X<p>2. Criticize your explanation for X, use reason and logic.<p>3. Based on your own critics, improve your explanation of X.<p>I have manually toyed with this approach (the prompts are longer, you get the gist), and it gives very interesting results. This could lead to GPT re-create, on its own, a better high-quality corpus to learn from.<p>Is anybody pursuing this approach for LLM?
The thing with AlphaGo Zero is that there is a clear external arbiter of which side of the internal debate wins, so the algorithm can learn.<p>For LLM to use the technique on the kind of reasoning you talk about, you need a human in the loop to explain it why it's wrong or right, otherwise it just hallucinates random stuff.<p>That's basically what RLHF[0] is, which was used to great success in training ChatGPT.<p>[0] <a href="https://huggingface.co/blog/rlhf" rel="nofollow">https://huggingface.co/blog/rlhf</a>