• [News] Trumps launches the Genesis Mission, a government-wide effort to use federal scientific datasets to train models.

  • [Tweet] Super long Gemini 3 Pro system instruction, reported have a significant impact on various agentic benchmarks.

  • [Company] Conductor - Run parallel coding agents in isolated environments. YC company page.

  • [Tweet] DeepSeek-V3.2 & DeepSeek-V3.2-Speciale.

  • [Paper] In RL, penalizing mistakes seems to contribute more than rewarding successes.

  • [Paper] Imitation-based interactions on the individual level can be re-interpreted as collective RL at the group level. In other words, bandit RL can be implemented as a collection of very simple parts.

  • [Blog] Feedback descent: use textual feedback to learn instead of a single score. When model weights are too entangled, this allows learning to continue in semantic space. Relevant paper.

  • [Blog] Best practices for AI code reviews.

  • [Blog] What does it mean to ship a project?

  • [News] Claude is dead.

  • ［專欄］郵件工具的PMF。

Have a great week!

Updated:

Comments

You May Also Enjoy

The week ahead: 20251201

[Blog] HOPE. A model as a set of nested sub-optimizers, each with its own context flow and hyperparams. Relevant paper.