There’s been much debate of late about how generative AI (read LLMs) is taking over our thinking. Which, of course, has caused me to think about what’s crucial for us, as thinkers. And, arguably more importantly, as learners. In short, we’re asking what is learning, and what is it that we can’t abrogate if we want to keep doing so.
So, there’s increasing evidence that when we use AI to just get answers, it’s at least keeping us from learning, if not making us stupider (warning: PDF). The point is apt, to develop, we need what’s termed ‘desirable difficulty’. That is, we need to challenge ourselves. You can think of it as keeping an exercise regime, except we don’t plateau, we keep learning. There is, effectively, no limit to what we can learn; we run out of time before we run out of new things to learn!
What happens when we ask an AI for an answer is that we aren’t exercising our own minds. It’s been made clear that we need to do the work ourselves – tackle the task – so that we activate and leverage our cognitive representations for them to strengthen. To be clear, there are things we don’t want to learn. For instance, I’m fine not having learned how I repaired my dryer; I’m not a dryer repairperson. However, I don’t want some arbitrary summary that might be wrong, I prefer to find a document created by an expert who guides me through the steps to diagnose and repair it. And then I’m fine not learning, because I hope to never have to do it again!
The interesting question then becomes, what do we do? Markus Bernhardt has suggested there’re five steps, from working with a contractor. These are briefing, judging, composing, selecting, and building. Only the first two are before you talk agents (I’m not, for reasons). However, I want to simplify it, and suggest that there’s essentially one step that’s critical. As suggested by the above.
The main thing I think one needs to do is first generate our own ideas. I’ll suggest that an LLM can be a good partner for thinking. It can generate ideas you haven’t, though of course it may also generate ideas you reject. So evaluating the output is important (requiring an expert). But what I now think is that you should generate your ideas first. Then see what you’re coming up with. It’s an extension of the idea of self-evaluation of free form responses, with a rubric. Particularly when learning to learn on one’s own, you should be generating responses, but you need feedback. Without an instructor to do so, you need a way to evaluate your own output. Comparing your answers to someone else’s is good! Particularly if you’ve done the work beforehand to know what a good answer should contain or consists of.
Let’s be clear; you don’t have to flesh it all out. What you do need to do is get concrete enough that you know your own ideas. It’s one thing to think you’ve got an answer, and another to actually try to capture it. It’s a pretty common phenomena to realize, when you try to get concrete, that you need to tweak a few things. That thinking is part of the learning. But a quick phrase, a sketch, whatever captures your thinking quickly can be enough, it’s just that you have to generate the idea.
So for any area you want to grow in, you can’t abrogate your responsibility. That’s a negative, not a positive, path. For other things, I prefer the best expert advice, not a potentially wrong summary generated by a stochastic parrot (yeah, I said it ;). But when it’s important for your learning, and that is something you can and should be doing, you need to do the work. Sure, get assistance, but do the work and then get feedback. And evaluate the feedback to understand how you can improve. But, do the work. That’s where the learning happens. That’s what is learning: action and reflection. So, as the Nike ad says, Just Do It!