The week ahead: 20251229

Industry

• [Blog] Karpathy’s EOY review. Verifiable rewards as a better deal for fixed compute, leading us to all sorts of philosophical questions this year.
• [Company] MVP Studio. Builds your MVP for you.
• [微信] 醫療數據的商業邏輯。

Paper

• [Paper] ReaSeq. Recommendation systems based on in-log IDs have two limitations: (1) the noise and sparsity of co-occurrence statistical signals and (2) invisibility of user behavior outside of the logs. This work uses a LLM reasoning setup based on logs to uncover user demands and product attributes to combat the first limitation, and uses a diffusion-based approach to predict invisible user actions to compensate for the second.
• [Paper] JustRL. Simple, single-stage RL training without any so-called fancy tricks. This shown as transferrable between models, uses less compute, and matches SOTA on math reasoning tasks.
• [Paper] Estimate human body pose from Wifi signals. Relevant GitHub implementation.

Code

• [GitHub] Banana-slides. Nano Banana for slide deck generation.
• [GitHub] Jaaz. Prompt enhancement or multimodal generation workflows.
• [GitHub] ShadeOfColor2. Encode any file as an image.

Have a great week!