Get the latest tech news
The era of exploration
This post explores the idea that the next breakthroughs in AI may hinge more on how we collect experience through exploration, and less on how many parameters and data points we have.
The tougher question is to generalize these capabilities to fuzzier domains (e.g., generate a research report or write a novel) where feedback is sparse or ambiguous, and large-scale training and data collection are difficult. In supervised learning (or unsupervised pretraining), this axis covers data collection, synthetic generation and curation: gathering and filtering raw documents, images, or code, each of which corresponds to a “world”. Many thanks to Allan Zhou, Sam Sokota, Minqi Jiang, Ellie Haber, Alex Robey, Swaminathan Gurumurthy, Kevin Li, Calvin Luo, Abitha Thankaraj, and Zico Kolter for their feedback and discussion on the draft.
Or read this on Hacker News