Learning to think (again)

I started something small a couple of days ago and it is now public: Learning to think (again), one short code challenge per working day, solved by hand.

Why

Large language models and coding agents changed how I write software, and they are close to mandatory if I want to stay productive enough to be competitive. Something erodes along the way. It is easy for me to believe that reading part of what a model produced is the same as understanding it, that this is just a faster way of doing what I could have done myself. For me that belief is false. I notice my coding skills fading, and my sense of how data flows from the first line of an algorithm to the last has become partial.

I was already aware of it. Reading Gilles Louppe’s Atrophy, running the oral exams for the Deep Learning course, and talking with other researchers made it concrete. We are all affected, and I am part of it.

The rules

• No LLMs, no agents, no autocomplete beyond plain editing.
• Reading the official documentation is fair game when I am stuck.
• One challenge per working day, five days out of seven.
• About 20 minutes each, not more.
• The goal is not the perfect solution. It is to think the problem through and write the implementation myself.

The mix is weighted toward the tools I actually use: Python, JAX, PyTorch, C, and a Math track that is pen and paper only, covering integration, linear algebra, and the derivations behind deep learning.

A small irony

The challenges themselves are written by an LLM agent following a brief I wrote. I solve them by hand, but I do not invent them by hand. The tool I am trying to keep at arm’s length is the one handing me the exercises. I find that funny, and a little honest: the point was never to refuse the tools, it is to make sure I can still think without them.

Where it lives

The page loads the current challenge live from the repository, and the full log is there too. If you recognise yourself in any of the above, you are welcome to try it.

📄 The page: Learning to think (again)

💻 The repository: github.com/julienbrandoit/Learning-to-think-again