AI is great when I'm a noob, but not when I'm a world expert

AI is great when I'm a noob and close to useless in domains where I'm a top 1000 domain expert.

Nicely put, Vitalik.

This may explain why some of my sharpest friends and acquaintances are disappointed with LLMs and don't use them much: they're asking for help in domains where they are top 50 global experts.

If you're new to using LLMs, and you want easy wins, you should probably start with other high leverage areas, e.g. "help me plan my week", or "help me do a personal finance review", or "teach me about $UNFAMILIAR_AREA".

Once you get good at LLMs, you'll get a better sense of the strengths and weaknesses, and find ways to get notable helps even when you're working on domains where you are a top expert.

For example:

• If you're doing cutting-edge AI safety research, you may need to do a bunch of fairly mundane scaffolding or other technical tasks. Proficient use of Claude Code could take that from days of work to minutes. Your default might have been to delegate to an engineer, but removing dependencies can speed up your iteration cycle, which can be a very big deal.
• Google Deepmind is accelerating science.
• People say that LLMs are bad at writing. But I know a world class copywriter, proficient with LLMs, who gets great value out of them. Skill issue!

But sure, if you're trying to think through tradeoffs between zk-SNARKs and zk-STARKs, then maybe not. But also: LLMs are now getting IMO gold. So, watch this space...

What is the most efficient way to help world experts attain the kinds of AI fluency that'd accelerate their core endeavours? I don't know, but it's a key question for me over the next weeks.