This is a special post for quick takes by Daniel Tan. Only they can create top-level comments. Comments here also appear on the Quick Takes page and All Posts page.
Effectively multi-tasking with AIs seems to require a state of mind that is inimical to deep work / thinking. When I'm switching between AI panes I'm constantly distracted, disoriented, frazzled. I get tired out having to pick up and put down large amounts of context each time, etc. etc. I've observed that doing large volumes of work with AI coding agents mostly destroys my ability to do deep thinking for the rest of the day.
I don't know what the right way is to resolve this. What's worked reasonably well for me so far is to do most of my deep thinking in the morning(s) when my mind is pretty fresh. And then do AI coding in the afternoon once I'm reasonably sure I've hit the limit on how much deep thinking I want to do that day. But this isn't really perfect either since it doesn't address the root cause.
Tentatively I feel like the problem is that "deep work" and "AI work" exercise two very different modes of thinking, which are a bit hard to switch between.
The deep work exercises a low-bandwidth, single-threaded, highly sequential mode of thinking. Deep work requires emptiness and focus so I can get into flow state. Deep thinking in particular requires deliberately taking time and space for listening to my own curiosity, so I can hear the inner thoughts which are easily drowned out by background noise.
The AI work part exercises a high-bandwidth, multi-threaded, parallel mode of thinking. This requires loading and unloading context, making quick 80/20 decisions, communication, and constantly monitoring (and remembering to monitor!) the situation for updates. Though it seems very different, it somehow also drains the mental energy needed for doing deep work.
Other observations.
"Deep work" vs "AI work" is a bit like "focusing on the big picture" vs "focusing on the day to day". In many successful companies, these roles are split across different people. A similar specialization of roles might be important for properly getting the benefits of both.
The people who are the most successful at using AI to do work seem to be able to use it in ways which don't require much cognitive load. Example #1 - repeatedly telling the agent "don't give up" until it gets to a SOTA math result. Example #2 - eliminating the human review step for software PRs in favour of "full self-driving" codebases.
Basically, convince your AIs to do Scrum (with one AI acting as Scrum Master), and you become the Project Owner. Your task is to specify the project backlog and priorities, and the AIs will work on them, and provide a demo at the end of each sprint. (During the Retrospective they will discuss how to convert you to paperclips.)
I desperately wish there was a way to iterate fast on high-quality thinking.
3Jan Betley
For me the solution is to split the work into chunks that take e.g. 3h-12h and the model can do fully autonomously. So you can just start it and not look at what it's doing. Ideally, fire several in parallel and check on them only e.g. 3 times a day. I think the "deep work mode" is destroyed only when all the time there might be something requiring your attention, and it's on you to make sure there's nothing : )
So e.g. there is a giant difference between working with an AI that needs your attention 2 times per hour for 30 seconds and one that never needs your attention at all.