On Memory, Learning & The LLM Era
It's a widely held idea that learning is linked to memory. Reflecting on the years before LLMs, I have to confess: I didn't really remember a thing.
When I started out as a coder, memory was everything. Memory, repetition, and a rapid feedback cycle. This is how I learnt.
For some time, my stack has been PHP, vanilla JS and AWS, with a particular interest in system operations and security. Whether this is too wide or shallow is debatable, but I clearly remember a tipping point where I could not hold it all in my head anymore. I couldn't tell you what I'd coded yesterday, never mind last week.
What I believe you fell back on at that point, was the foundational principles, applied in whatever syntax happens to be in front of you. And this transformation was great, as I truly started thinking about systems and fitting what I was doing into the wider picture. Not just coding to solve whatever problem was under my nose.
So if LLM's are eroding memory, and memory is how we learn. I have to ask: which developers is this change hobbling?
As I have this feeling that most people I have worked with stopped remembering the specifics a long time ago. I do not think this can be attributed to any single tool, and I am not convinced that this outsourcing of thinking we worry about started with the birth of ChatGPT in 2022.
Stackoverflow and Google? Nobody has ever been able to hold all that information in their head. And even before then, people referenced code in books (the horror!). Granted, you did have to look a little harder, but I don't think it was ever called a crisis. Unless of course your were that dev that copied and pasted rm -rf * without thinking.
What I am less sure of, is if what happened to me was compression or decay? Its entirely possible I stopped storing detail because I had folded it all into neat patterns, which, I hear is what expertise looks like. Its also plausible that I just got worse at remembering and good enough at everything else that is stopped mattering. From inside my head, they both present as 'I do not recall the specifics, but the work gets done.'
Claims that LLMs are damaging the next generation imply that you can see the difference between using a tool and depending on it.
I can't see that difference in my own memory. But, I think I can see it overall in the work. It's the difference between thinking in systems and not; between code that lasts and code that doesn't. And this, was a problem I had with developers of every seniority long before the proliferation of LLM's.
Which is what makes me suspicious about the idea we are going to have real trouble with the next generation of junior developers and their ability to learn.
I am not saying there is nothing in the argument. The first stage of learning to code, the repetition, the fast feedback etc. That is real and it is the bit we all benefited from, and I can't be sure juniors are getting it. But that's a question about that specific stage, not about memory and learning in general.
I don't have an answer on if LLM's will effect our ability to learn. But I have a decent reason to think a lot of the people who confidently claim that they do, perhaps are not looking with enough context.
