Books, papers, and articles I’m exploring, with a few thoughts to return to.
PaperZhihong Shao et al.
Shows how training on carefully selected math data improves a language model’s math skills. It also introduces GRPO, a reinforcement learning method that compares groups of answers and uses less memory than PPO.
Open sourcePaperDeepSeek-AI
Explains how DeepSeek-V2 lowers the cost of training and running a language model. It activates only some of its expert networks for each token and compresses the stored attention data to use less memory.
Open source
PaperRafael Rafailov et al.
A simpler way to train language models to follow human preferences. It learns directly from pairs of preferred and rejected answers, without training a separate reward model or running a reinforcement learning loop.
Open source
BookMohsin Hamid
Nadia and Saeed flee a city at war through mysterious doors that lead to other countries. A story about migration, love, and how people change as they search for a new home.
Open sourcePaperAshish Vaswani et al.
Introduces the Transformer, which uses attention to connect words across a sentence. It processes words in parallel rather than one at a time, making training faster and improving translation quality.
Open source