WHAT DID DEEPSEEK FIGURE OUT ABOUT REASONING WITH DEEPSEEK-R1?

Summary from the TLDR newsletter

Chinese AI lab DeepSeek recently released a reasoning model trained on a GPU cluster a fraction the size of any of the big western AI labs that is better than the current best reasoning models. Reasoning models are able to 'think' about questions, which results in better answers. DeepSeek used a reinforcement learning approach, as opposed to a fine-tuning approach, to train its reasoning model, which resulted in cost and quality benefits. This approach can only really train the model on coding and mathematics, but this might not matter as superintelligence in coding/mathematics might transfer to other domains - this has yet to be demonstrated in practice.