2 ms·
Google DeepMind CEO Demis Hassabis Says Its Next Algorithm Will Eclipse ChatGPT
- isaacfrond 3y agohttps://archive.is/XFwF6 https://archive.is/XFwF6
- sylware 3y agoIt more about the size and speed of the connectome, the size and richness of the training dataset, than the algorithms (unless major breakthru ofc).
- isaacfrond 3y agoAlphaGo had three components of note: - MonteCarlo Simulation. - A PolicyNetwork and a Value network. - Improvement through self play The MonteCarlo simulation I could imagine. Instead of doing multiple full samples, you do one MC evaluation. The PolicyNetwork would play the same role as the LLM we use no. The PolicyNetwork predict the next move, just like an LLM predicts the next token. The Value network should put a 'worth' on the text you generated so far. No such network is used in current LLM I think. Finally, what would the improvement through self play look like. A rap battle against itself?