OpenAI's GPT-2 secret life as a pawn star: Boffins discover talkative machine-learning model can play chess GPT-2, OpenAI’s giant text-generating language model, can play chess – despite having no prior knowledge of the game’s rules. That may seem pretty odd, at first. The system is mostly well-known for spitting out passages of text after receiving a sentence or two as a prompt, after all. Trained on around eight million newspaper …
AlphaZero famously is really really good at learning to play chess, but proprietary from Google and runs on specialist hardware.
Leela Chess Zero is a hobby/OS project based on the same principles which has gone from being (relatively) terrible to beating many advanced engines in the space of a year or so, which is pretty fascinating.
StockFish represents the old guard of chess AI.
Everything GPT-2 has ever produced (like fake news articles) is quite convincing at the start then quickly wanders off into garbage.
It just isn't that good, and definitely struggles to stay on track. A convincing opening can just be cribbed straight from training data but getting past that...
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
