@hwchase17 How do you hillclimb? Are optimization loops guaranteed to move system prompt / agent.md / skills in the right direction à la SGD for traditional ML?
A thing I hate about codex and claude code equally is that the name of the thread is generated based on the first few messages. I am continuously context switching inside the same thread and hate having to manually rename it so I don't lose track of the work done @trq212@jxnlco
@hwchase17 Doesn’t post-training partially specialize models to certain harnesses? e.g., Opus may perform better in Claude Agent SDK than in deepagents, OpenCode, or pi
@sandhya Skills that are honed via hill-climbing using a real life eval set are a particularly interesting idea. Somewhat of a DSPy approach maybe. I’ve heard the term “Skill Harvesting” being thrown around. Definitely a moat if done at scale and maintained imo
ألا سيف من الإيمان يبرى السيف مسنونا
يجلى عن سما الأوطان هذا الذل والهونا
يقود إلى جنون المجد أبطالاً مجانينا
بقلب يحمل الآمال والآلام والدينا
يهز القوم بالذكرى وقد ينسى الفتى حينا
إذا أعطيت وعد الحر كان الوعد مأمونا
Layer10 is hiring an Executive Assistant based in Lebanon. Role includes managing calendars, handling general EA responsibilities, and potentially supporting some marketing work. If you’re organized, proactive, and interested in working with a fast moving team, reach out!
https://t.co/cVEitYTegh