Focus of the letter 45

Fal.live broadcasts continuously generated video channels, each based on the MiniMax H3 Max model. Anime, sitcom, soap opera, storybook, chaos: several universes run in parallel, each with its own style of image, voice or music, produced at the same time and live by the machine. Regularly, a vote is opened: four instructions are proposed, the viewers choose, and the stream goes back in the winning direction. During the test, one of the channels was protected by a password as a discreet, and perhaps unwanted, reminder that these channels belong to someone.

While we have the impression of deciding, we are in fact choosing between four proposals that a model has itself generated, within an aesthetic and ethical framework that its publishers have already set before we even open the page. Voting gives the feeling of control. It concerns the continuation of a broadcast, never the rules that make it possible.

We almost never see these rules, and they come from many places at once. The developers of an application choose to integrate one model rather than another, and this choice already includes data and training methods decided far from us, by teams whose names we don’t even know, with filters and refusals that we sometimes discover by accident while testing a tool several days in a row. We have already come across it with Ecosia and its choice to back its AI features with Mistral Small: the choice of a model is not neutral, it carries a coherence… or its absence if this choice is linked only to cost or performance.

There are also the choices of the company that decides to keep a feature free or to reserve it for a subscription. To make a feature a marketing tool like fal.live to show that MiniMax H3 Max is available among the models offered in the fal.ai subscription. To withdraw a product that does not find its business model, or on the contrary to continue to invest in a segment that others are abandoning (we discussed this in letter 35 with the closure of Sora).
There is also hosting, an issue that is far from being purely technical: a model running on an infrastructure powered by carbon-free energy as in Euria does not have the same environmental impact as a model hosted in a datacenter built in the middle of the desert.
There are geopolitical tensions, such as those between China and the United States that were recently revived, which are shaping two ecosystems of models, each with its own very different commercial and opening logics.
There is regulation, when Europe imposes safeguards that other continents do not put in place or at least that they do not yet have.
There are also choices that are closer to us, those that a professional collective makes for all its members by deploying this or that application.
And at the level of the user, there is again the proposed vote: a gesture that gives the impression of deciding, but which probably silently feeds the model that proposed it.

This list is certainly not complete, and I am not an expert on all these subjects to draw it up in detail. It just shows that the decisions that shape a tool and its use are numerous, dispersed, often invisible to each other, and that none of them is ultimately made by the person who, at the end of the chain, each and every one of us, clicks on one of the four options proposed in fal.live.

But nothing on this list is entirely incriminating, each point must be qualified. A filter can protect, a regulation can set useful safeguards, an industrial choice can benefit the user. The problem is not that these choices exist, it’s that the user doesn’t know them all the time or doesn’t have access to them.
All of these decisions are made before the tool gets into our hands. We choose an instruction, a color, a tone, a direction in a video stream. We almost never choose the framework in which that choice takes place, or the people who shaped it for us.

Some of these choices even remain invisible even in the answer itself. A model decides, in each generation, what it says explicitly and what it implies, what it develops and what it eclipses. This is the subject of “Journey to the Land of the Unwritten”, the application built from an article by Arthur Sarazin to identify what the responses of generative AIs do not say: the implicit, the implied, the blind spot, the detail that is supposed to be acquired without ever stating it. These are also choices, they are perhaps more difficult to see because no one is asking us to vote on them.

What if knowing these rules, precisely, was already a way of acting? Individually, it changes what we agree to choose without thinking about it. Collectively, it changes what we are able to demand: more transparency on a model, a regulation, a mode of hosting. It will certainly never make each level of decision-making visible, but perhaps getting into the habit of asking the question, rather than believing that voting among four options is a decision entirely, will end up weighing on what we demand in return for transparency?
Do we still have to ask ourselves if we keep control of what we produce with a tool, or if we still know how to identify, in the four options that are offered to us, the trace of the choices that we have not been allowed to make?

Echo:

• Letter 39, “Research after research?“,
• Letter 35, “Video Generation, a Geography That Is Taking Shape?“.