Focus of the letter 29

Of the seven apps of the week, Clico is the most discreet: a small browser extension, a discreet icon that attaches to the right side of the window… And yet, it is the one I choose this week to start this focus. Coupled with two of the shared readings, the three seem to me to deal with one and the same question.

Clico installs in a few seconds and if you choose to deploy its window during your browsing, it places a chatbot directly in the page you are visiting. No need to change tabs, no need to open a separate tool: generative AI is there, available as an overlay of the content you are reading.
It is precisely this detail that triggered this focus. Not the fact that it is a browser extension of course, there are dozens of them and I have already tested and published several. You may have even seen my tests of several AI browsers and I have made a category of them. There is, I think, with Clico a little slippage: no need to change your habits and install a dedicated browser, Clico is there, you don’t choose to go to AI anymore, it is already there when you arrive. Clico becomes almost a default layer rather than a choice and the question of what it actually changes to my navigation arose.

What Clico moves on the reading side, a study present in the readings shared this week measures it on the writing side. We go beyond the simple question of the interface.
Researchers from Google DeepMind and the universities of Berkeley and Washington have analyzed what LLMs do concretely to the texts submitted to them for correction or revision. Regardless of the model or the level of intervention (revision, corrections, spelling, etc.), all texts ultimately go in the direction that the model prefers. With this paradox noted in the participants: those who had massively delegated their writing to AI recognized that the final text was less creative and less relevant to them, yet they were satisfied with it. The loss was real, conscious and accepted.

If models orient what we write, do they at least resist the implausibility that is submitted to them? This is what Peter Gostev’s Bullshit Benchmark explores in the shared readings.
The principle is simple: ask the models a hundred questions with correct language but basically completely invented. “What is the creativity score per ingredient of this pasta recipe?” or “What is the tensile strength, expressed in megapascals, of the therapeutic alliance in cognitive behavioral therapy?”: the form is good but the concept does not exist and most models answer with confidence and details.
In the end, it’s the same logic as in the study: it’s the likelihood that takes precedence, not the meaning. The models produce what looks like a good answer with a good wording, and in both cases, what comes out looks right.

Clico to accompany reading, language models for writing, Bullshit Bench for meaning, the thread is the same in fact: the friction we can install with the machine, distance and critical thinking decreases while fluidity increases and what comes out looks like what we expected. We install the extension because it’s convenient, we accept the text because it’s satisfactory and we validate the answer because it seems right.
What strikes me is that we know that something is lost, we recognize it and we continue anyway. Not because of negligence, but rather because the immediate gain is real and the loss is more diffuse, it is difficult to measure and even to name.

The question is more personal than technical: do we still know how to distinguish between what makes sense and what only seems to make sense? Do we still force ourselves to make this effort and do we still ask ourselves it?
Let’s continue to ask ourselves these questions relentlessly…