…
Unlike Alice, who spent the year reading papers with a pencil in hand, scribbling notes in the margins, getting confused, re-reading, looking things up, and slowly assembling a working understanding of her corner of the field, Bob has been using an AI agent. When his supervisor sent him a paper to read, Bob asked the agent to summarize it. When he needed to understand a new statistical method, he asked the agent to explain it. When his Python code broke, the agent debugged it. When the agent’s fix introduced a new bug, it debugged that too. When it came time to write the paper, the agent wrote it. Bob’s weekly updates to his supervisor were indistinguishable from Alice’s. The questions were similar. The progress was similar. The trajectory, from the outside, was identical.
Here’s where it gets interesting. If you are an administrator, a funding body, a hiring committee, or a metrics-obsessed department head, Alice and Bob had the same year. One paper each. One set of minor revisions each. One solid contribution to the literature each. By every quantitative measure that the modern academy uses to assess the worth of a scientist, they are interchangeable. We have built an entire evaluation system around counting things that can be counted, and it turns out that what actually matters is the one thing that can’t be.
…
The strange thing is that we already know this. We have always known this. Every physics textbook ever written comes with exercises at the end of each chapter, and every physics professor who has ever stood in front of a lecture hall has said the same thing: you cannot learn physics by watching someone else do it. You have to pick up the pencil. You have to attempt the problem. You have to get it wrong, sit with the wrongness, and figure out where your reasoning broke. Reading the solution manual and nodding along feels like understanding. It is not understanding. Every student who has tried to coast through a problem set by reading the solutions and then bombed the exam knows this in their bones. We have centuries of accumulated pedagogical wisdom telling us that the attempt, including the failed attempt, is where the learning lives. And yet, somehow, when it comes to AI agents, we’ve collectively decided that maybe this time it’s different. That maybe nodding at Claude’s output is a substitute for doing the calculation yourself. It isn’t. We knew that before LLMs existed. We seem to have forgotten it the moment they became convenient.
Centuries of pedagogy, defeated by a chat window.
…
This piece by Minas Karamanis is excellent throughout, and if you’ve got the time to read it then you should. He’s a physics postdoc, and this post comes from his experience in his own field, but I feel that the concerns he raises are more-widely valid, too.
In my field – of software engineering – I have similar concerns.
Let’s accept for a moment that an LLM significantly improves the useful output of a senior software engineer (which is very-definitely disputed, especially for the “10x” level of claims we often hear, but let’s just take it as-read for now). I’ve experimented with LLM-supported development for years, in various capacities, and it certainly sometimes feels like they do (although it sometimes also feels like they have the opposite effect!). But if it’s true, then yes: an experienced senior software engineer could conceivably increase their work performance by shepherding a flock of agents through a variety of development tasks, “supervising” them and checking their work, getting them back on-course when they make mistakes, approving or rejecting their output, and stepping in to manually fix things where the machines fail.
In this role, the engineer acts more like an engineering team lead, bringing their broad domain experience to maximise the output of those they manage. Except who they manage is… AI.
Again, let’s just accept all of the above for the sake of argument. If that’s all true… how do we make new senior developers?
Junior developers can use LLMs too. And those LLMs will make mistakes that the junior developer won’t catch, because the kinds of mistakes LLMs make are often hard to spot and require significant experience to identify. But if they’re encouraged to use LLMs rather than making mistakes by hand and learning from them – to keep up, for example, or to meet corporate policies – then these juniors will never gain the essential experience they’ll one day need. They’ll be disenfranchised of the opportunity to grow and learn.
It’s yet to be proven that more-sophisticated models will “solve” this problem, but my understanding is that issues like hallucination are fundamentally unsolvable: you might get fewer hallucinations in a better model, but that just means that those hallucinations that slip through will be better-concealed and even harder to identify in code review or happy-path testing.
Maybe – maybe – the trajectory of GPTs is infinite, and they’ll keep getting “smarter” to the point at which this doesn’t matter: programming genuinely will become a natural language exercise, and nobody will need to write or understand code at all. In this possible reality, the LLMs will eventually develop entire new programming languages to best support their work, and humans will simply express ideas and provide feedback on the outputs. But I’m very sceptical of that prediction: it’s my belief that the mechanisms by which LLMs work has a fundamental ceiling – a capped level of sophistication that can be approached but never exceeded. And sure, maybe some other, different approach to AI might not have this limitation, but if so then we haven’t invented it yet.
Which suggests that we will always need experienced engineers to shepherd our AIs. Which brings us back to the fundamental question: if everybody uses AI to code, how do we make new senior developers?
I have other concerns about AI too, of course, some of which I’ve written about. But this one’s top-of-mind today, thanks to Minas’ excellent article. Go read it to learn more about how physics research faces a similar threat… and, perhaps, consider how your own field might need to face this particular challenge.
