⌨️

Learning by doing

A story about AI tools, and how they will change the world.

Sep 30 - 2024

A good friend recommended me to try Cursor.sh, so I did.

I was not ready for what happened next.

But first let's start with a little idea I had for a while. Every time I needed to check a text, I needed to find my tab with ChatGPT, login if disconnected, start a new chat, type "fix typo:" then paste my text to fix, then copy the text back.

Not efficient. I knew that using OpenAI APIs I could do a simple website that would do the same, but without the need of writing "fix typo:" every time. A simple hack. Of course I never found time to start this pet project. Until I decided to give it a try with Cursor.sh.

Cursor.sh is yet another IDE, but it's a fork of Visual Studio, which is (almost) everyone's favorite. Visual Studio's best extension is Copilot, a tool that autocompletes the code you're writing.

Copilot is so shockingly good that it feels like it's reading your mind all the time. It literally guesses the next function you're going to write. With just a press of the tab key, it writes it for you.

There are also ChatGPT and Claude. Both are excellent at writing code. However, you have to copy and paste their output into your editor.

But Cursor.sh is another level. It optimize the integration of LLMs to the next levels. The LLMs have access to more context like your full codebase or the error message if your code is crashing. You can ask Cursor.sh things like: Move the business logic of this fonction to the backend, and it will suggest modifications in multiple files, so you can review before accepting them.

The point is not that the AI cursor is perfect: it’s not. Many time there are mistakes

The point is that the AI knows enough to get you started. You get your first reward faster, and it helps you at every step. You'll continue to get rewards along the way. It reduces frustration because you don't feel stuck alone. You can always ask the AI for help, or at least for explanations about what's going on, what was expected, and how the code works. Every failure of the AI is an opportunity to learn something. It takes "learning by doing" to the next level. Have the will, and the rest will follow. It has never been so easy.

Here is the result of this experiment: fix-typos.com

An AI generated 95% of the code.

Polishing the UI consumed 80% of the development time.

But most importantly, I've never felt more empowered.

So here are multiple predictions based on what I learn last week.

  • In 5 years, if you ask someone on the street about their AI "Aha" moment, it won't be ChatGPT. It will be tools with a narrower scope that empower them in a much deeper way. ChatGPT giving better answers than Google is nice, but when you suddenly feel you can do 10 times more of what you love, it's something else entirely.
  • The tech industry will be the sector most disrupted by AI, by far.
  • Software engineering schools will have to reinvent themselves. Learning by doing will become the norm.
  • A new generation of bootstrapped startups will rise, with co-founders no longer hiring full-time employees.
  • The efficiency gap between good and bad software engineers will widen dramatically, not narrow.
  • UX will become the main differentiator for many products.