• [Announcement] ChatGPT Pulse. ChatGPT can now proactively push personalized content.

  • [Paper] RL’s Razor. KL divergence between trained and base models on the trained task is a strong indicator for forgetting; On-policy training causes less forgetting than SFT because it implicitly favors KL-minimal solutions by sampling from the model’s own distribution.

  • [Paper] Continued Pretraining with agentic abilities in mind.

  • [GitHub] Lots of algorithms in Python.

  • [GitHub] Spec-kit: YANC (Yet Another No-Code).

  • [Paper] Is In-Context Learning learning?

  • [Article] Investment lessons from the past - should one swim upstream or fish downstream?

  • [Blog] Onboarding best practices.

  • [Tweet] David Haber: The best opportunities are usually the hardest to classify.

  • ［採訪］Plaud的成功：是市場驅動產品，還是產品驅動市場？

Have a great week!

Updated:

Comments

You May Also Enjoy

The week ahead: 20251006

[Blog] Thinking Machines Lab: LoRA is better to be applied to all weights, including MLP and MoE. On RL, LoRA performs on par with full training even with...

The week ahead: 20250915

[Blog] First glimpse into the work of Thinking Machines Lab. Inference systems can be made deterministic by using “batch-invariant” GPU kernels (GitHub im...