How to keep learning in the age of LLMs
Oğuzhan Olguncu
Opinion — commentary, not a factual news event.
LLMs make it too easy to skip the hard part of learning. One writer says use them as a tutor, not an answer machine, or you stop actually getting better.
Based on reporting by Oğuzhan Olguncu — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Learning is getting weirdly harder in the age of LLMs. Not because the tools are bad — they’re astonishing at one-shotting code, explaining syntax, and clearing away busywork — but because they remove the friction that used to teach people things in the first place. The author, writing mostly from a programming angle, argues that the old route to real understanding still matters: struggle a little, get something wrong, then fix it yourself.
That’s why the piece starts with Bitcask, a lightweight paper the author read and then implemented mostly without AI help. When they got stuck, they didn’t ask an LLM to finish the code. They asked it to explain the concept clearly so they could keep going on their own. That distinction matters. An LLM that hands over the answer can short-circuit learning. An LLM that asks questions can push someone toward the answer and make the lesson stick.
The author points to a “socratic-code-mentor” style of prompting as the sweet spot. In the example, a buggy loop that returns NaN isn’t fixed directly. Instead, the model asks about array length, last index, and loop bounds until the mistake becomes obvious. The point is simple: if you work out the bug yourself, you’re far less likely to forget it. That matches the logic behind Make It Stick and the broader idea that effortful recall beats passive receipt.
The article then shifts from understanding to stamina. LLMs are useful for the dull parts — tests, tooling, visualization — while the human keeps the interesting core. For a toy LSM-tree project called tinylsm, the author used a plan with phases, clear “done when” checks, and a checklist the model could follow in later sessions. That reduces the blank-screen problem and makes it easier to keep going in short bursts instead of burning out.
And the motivation angle is very practical. The author likes visible progress, like a REPL that shows tables piling up or read latency dropping as compaction improves. Seeing 88 tables versus 1 table turn into very different latency numbers is the kind of feedback that keeps curiosity alive. The larger lesson is blunt: use LLMs to scaffold learning, not replace it, and pick ambitious projects that still leave room for your brain to do some work.
My take — AI-written commentary, not fact-checked reporting
This is the right antidote to AI laziness, and it’s embarrassingly simple: stop asking the model to do your thinking for you. Europe spends half its time worrying about AI dependency, and this is the real version of that problem — not robots taking jobs, but people outsourcing the part that makes them worth hiring. A chatbot that flatters your competence is a lousy teacher; one that keeps saying “why?” is actually useful.
Read more about this at: Oğuzhan Olguncu