The week ahead: 20251124

• [Tweet] Sergey’s interview.

• [Tweet] Sam Altman’s interview. Hints on a new work suite that doesn’t “create fake work”.

• [GitHub] Diffusion LM stack for, e.g., chatting with BERT.

• [GitHub] Karpathy’s LLM council.

• [Paper] Predicting a summary of the future tokens shown to be more effective than multiple-token prediction.

• [Paper] WorldTest - test for world-learning models. The test phase is separated from a reward-free interaction phase in a different but related environment, mimicking an open-ended modeling setup.

• [Paper] Task decomposition and red-flagging as remedy for LLM error-compounding.

• [Paper] Music Flamingo. General Music LM, driven in part by an extensive labeling process.

• [Utility] Andrew Ng’s paper review agent. Relevant tweet.

• [Tweet] Karpathy on the nature of intelligence.

Have a great week!