Zhu Liang

i don’t think running 10 agents in parallel and shipping 50 commits a day will get you to success in today’s ai climate.

but rather, spend more than 8 hours just thinking, pondering without distractions or specific goals.

let the ideas come to you in the most natural and intuitive ways. i guess the analogy would be using random walks to escape from local maximum.

> "we just got rid of titles"

> "that includes me and the leadership team"

ai companies have made 11 discoveries on open math problems in 2026 so far.

9 of them credited an internal unreleased model.

this week in ai (nov 2 - nov 8)

i have tried 3 methods for agent identity and continuity (claude code cloud sessions):

1. ask coordinator to create fresh sessions for each task or work to be done. the benefit is that i can always control what's in context, the skills and docs via coordinator. downside is implicit lessons and experience by each session is gone by default, and there needs to be some methods to document the learnings and reconcile them with existing docs frequently.

2. have "standing sessions" that are long-lived for a specific type of work, much like an employee, the employee takes tasks from the coordinator and focus on that part of the work. it keeps its own notes and does compaction automatically, so no knowledge is lost except auto compaction. benefit is that in theory the session note and post-compaction hook reminder can help agent remember in long-term (read skills and docs after compaction) and not lost too much knowledge compared to the fresh sessions. downside is you get lower quality output when context build up beyond 500k or 700k, and the hook reminder might not work if the agent compacts in the middle of something (no mechanical guarantee to load docs and skills after compaction, it relies on agent to follow the instruction).

3. ask the sessions to do handover on their own when the context is close to full (700k). this didn't work reliably on it own when i tried, as the agents don't seem to be aware of its own context window around that threshold. not sure if can we set up some hook to trigger it, but even so, it might cause problems if the agent is in the middle of a task.

not sure what others are doing. please share if you have some better setup or ideas.