Focus of the letter 28

57 articles in the “Application” category on the site, the first in April 2023, and three new ones this week: Flames, Cloudflare Build and Medo.
It’s not entirely a coincidence: I’ve been exploring a lot about “vibe-coding” and “co-development” in the last few weeks and two readings of the week deal with the subject, one on the education side, the other on the business side with Arthur Mensch from Mistral AI on “replatforming”.
Two very different scales, but the same idea is taking hold: a bit like in Fablabs, we experiment with making our own tools with the machines available.

I generated a certain number of applications for the site’s tests (you may have seen my famous multiplication farm pass by or even replay…), for the site itself (the “shared readings” page, the “help in choosing” AI tools, the database) but also for my daily use (a custom RSS feed aggregator is in progress for example…).
The access to languages that were almost inaccessible to me until recently is quite fascinating, the only one you seem to have to master now is natural language…
But once the first productions are over, something quickly fails: as with the 3D printer, it’s captivating to see your idea take shape. But as with the 3D printer I just felt like I was putting a file in a software, my .stl file being my idea and my Cura connected to my printer the chatbot that generates and executes… Understanding why this button doesn’t refer to the right place, why this menu is half hidden, is quickly necessary. We find the cyclist and the bike of last week: understanding how the machine works remains the only way to decide if and how to use it. That’s what agency is all about, certainly delegating but deciding to do it in “conscience”.

The avenue I explored is the one that we have been able to discuss here several times: forcing the machine to help with understanding or even to accompany learning. This will not make me a developer but at least it will allow me, and now most of the time, to understand what is being co-constructed. It has even become an open window on new possibilities: I was talking about it this week with one of you, I went from “wow” to “oh yes?”.

However, a few limits quickly impose themselves.
Security first: what does this generated code really contain? I understand the contours, I have it checked in another conversation, by another model: I know despite everything that I don’t control everything I delegate. So my choice is to share (on Github for example) but to use the results only on my tools, I produce proofs of concept not finished products.
Then the energy footprint: what I co-produce doesn’t already exist elsewhere? If so, do I really need to just change the buttons on a page with generative AI when a tool already does exactly what I want?
And finally, there is time: clearing, putting in order, defining the method, testing, understanding, starting over, it’s a long time. And that’s perhaps the most important thing: the most useful skill is not to generate faster, it’s to learn to distinguish the “wow” that opens up a path from the “oh yes?” that deserves to be stopped.
The “wow” is a good way to get in. But it’s certainly the “oh yes?” that makes us something other than spectators.