tl;dr: LLMs rapidly improving at software engineering and math means lots of projects are better off as Google Docs until your AI agent intern can implement them.
2. AI's being bad at research ideation is just an elicitation issue that's going to get solved soon; writing Google Docs with good ideas might also be much faster in a year. [I have no idea how true this is.]
3. The space wait calculation thing is not an accurate intuition pump for singular projects, except for the pretraining example. [Correct. Any particular ongoing research project benefits from being started earlier. The field as a whole (or my own overall research output) benefits from deferring ideas that are not temporally privileged and are more easily automated later.]
After this circulated for a while, I'd like to address common objections and cruxes:
1. Many ML experiments may not be bottlenecked not on software-engineer hours, but on compute. See https://www.lesswrong.com/posts/auGYErf5QqiTihTsJ/what-indicators-should-we-watch-to-disambiguate-agi?commentId=kNHivxhgGidnPXCop. [Interesting point. It has been communicated to me that researchers inside labs are bottlenecked by compute surprisingly often.]
2. AI's being bad at research ideation is just an elicitation issue that's going to get solved soon; writing Google Docs with good ideas might also be much faster in a year. [I have no idea how true this is.]
3. The space wait calculation thing is not an accurate intuition pump for singular projects, except for the pretraining example. [Correct. Any particular ongoing research project benefits from being started earlier. The field as a whole (or my own overall research output) benefits from deferring ideas that are not temporally privileged and are more easily automated later.]