Focus of Letter 50

Fifty letters, and a framework that has hardly moved: the focus, the publications, the readings. These three pieces were there from the beginning. Everything else came down board after board, as you joined this Saturday sharing, by dint of trying, correcting, following new desires and ideas, and then starting again the following Saturday.
An issue is prepared from Saturday to Saturday, as it happens. The tools are published every day on the site, so the list of the week exists by itself: you just have to gather it for those who don’t come by every day. I put the readings aside when they seem interesting to me, without yet knowing what they will say together. The focus is born from a reaction: most often a tool of the week, sometimes a reading that challenges me.

With the machine, I work in two very different ways. I write “the machine” on purpose, even if Fred Cavazza offers this week a dictionary to distinguish chatbots, agents, assistants and teammates: it’s a shortcut, and I’m aware of it, but it’s still and, I hope, will always remain a machine for me, without anthropomorphism.
For applications and readings, the machine writes and I resume. One pictogram and one sentence per tool, one short summary per article: summarizing seven articles and a dozen readings each week would take me far too much time without this help. There is always proofreading, modifications and corrections, because these summaries are sometimes partial, and sometimes erroneous. A poorly summarized reading is a poorly transmitted reading. This is what I liked about Veridive, which answers a question by searching YouTube videos and cites the passages with a link to the exact moment: an answer that can be verified at the source, that’s what I expect from a summary. The e-Enfance study, in this week’s readings, echoes this: more than one in four teenagers uses an AI to verify information, and four in ten rarely question the generated answers. It would certainly have been interesting to be able to compare lower, higher, equivalent figures with an adult audience? Place your bets!

For the focus, it’s the opposite: I write, the machine accompanies. The idea is mine, the unfolding too, the first draft as well. The confrontation comes next. The machine offers augmentations, it doesn’t impose anything, and most of its tracks fall quickly, too far from what I had written or gone elsewhere. I keep what extends my text without replacing it. It does have a quality that I don’t have: it finds what I’ve already written. It tells me when I’ve already addressed a point, and it’s the one that helps me to weave the links of the “Echoing” part, at the end of the focus.
Then there is the proofreading: content, form, spelling, typography. I do the first one and then I entrust the rest to another conversation and often even to another model, who did not participate in the writing and therefore offers a new analysis. A precise instruction sets the procedure to follow. It is the machine that checks, it is I who decides.

On the technical side, a little structuring. A project dedicated to the letter, with a project instruction that applies to each conversation, a method file that lays down the structure and the forms to be respected, and each week the published letter is added to the archives. This is what gives the machine its memory: no memories, a file that grows. And each week brings its extension, its improvement, its strengthening: the letter is built without ever ceasing to be inhabited.

But in the end, why write this letter? First of all, to stop publishing the applications already on the site every day on the networks, a very time-consuming task despite automation. Second, to share and archive my monitoring. Finally, to have a freer space for expression than that of the site, which is not the function of the site. The letter has also made the site itself evolve: the Focus and Shared Readings sections were born from it, and I ended up developing a search and export engine for these readings (documented on GitHub). And for those who prefer to listen rather than read, I generate a podcast version from the audio produced by Substack.

The one-AI adventure continues to evolve: since September an English version, more than 1,300 applications published, 4 million page views, 2 million unique visitors, 4,500 subscribers, with 50 to 100 new ones every week, and 100,000 reads on Substack. It’s a source of pride, of course, but I don’t allow myself to conclude much other than that I enjoy it and that apparently it’s appreciated.

I don’t share this approach here so that it can serve as a model: it’s simply my way of doing things, my “testimony”, and so much the better if it is reused or remixed, it’s also what I’ve wanted since the beginning of the adventure by publishing all the content under a free license. In Rochane Kherbouche’s eve on AI in schools, I find a common thread that follows me: distinguishing what the student produces from what he learns. For this letter, it’s the same question: what it produces every Saturday, and what I learn from it. The day when the machine will be able to do everything on its own in this method, I may not have much left to contribute. But it seems to me that despite the endless race for the most powerful model, our desires, our thoughts and our agency will remain immutable foundations no matter what.

Echo:

• Letter 30, “Echo: when making a game about delegation becomes a delegation experience”
• Letter 36, “What Remains When the Tool Has Done the Work?”
• Letter 40, “In the rear-view mirror: 39 words for 39 letters”