Learning in the AI era
Anthropic’s latest research has pretty much confirmed what I thought before, especially combined with learning science. Put those together and we find that learning something new does need some “desirable difficulty”. If everything goes completely smoothly, the thing probably isn’t in your learning zone, or you aren’t learning at all.
Take the people who depend on AI completely and never step into the process themselves. It’s easy and the efficiency is high, but that’s efficiency in the short term only. Over the long term, serious problems show up. Or you go to fix a bug and have no clue how the code was written, which is completely normal, because the AI wrote all of it. Reading AI code is no different from reading someone else’s code, and you don’t know the structure of the project well enough either. So fixing a bug takes forever.
By the tests in the research, how much you take in and how well you understand it this way doesn’t match someone who went through the struggle and difficulty of fixing bugs themselves. So the key to using AI, I think, is to get familiar enough with a project first, know its structural patterns well enough, and then use AI to gain some efficiency bit by bit, instead of depending on it completely, which is guaranteed to go wrong.
Translation note. I wrote this in Chinese. This English version is an LLM translation, so the wording is not mine even though the thinking is. Original: ai时代的学习.