Focus of the letter 32

This week I’m bringing together two applications that seem to have only open-source in common. One is a lightweight multimodal language model, installable on a phone and available without a connection. The other monitors the world in real time and is fully configurable. Two seemingly opposite directions, the power that is compressed on one side and the flow that opens up on the other, and yet the same question implicitly: what does this double scale bring us?

Gemma 4 runs on a phone, no server, no subscription: a multimodal, open-source model that can be installed on your device. In the end, it reverses the direction of what we are now used to following: not bigger, more powerful or with more parameters, but smaller, lighter and more autonomous.
This “miniaturization” movement is not new, we have already talked about it here with the small frugal models of letter 21 or Thaura and Public AI in letter 30.
What changes with Gemma 4 is the technical scale achieved: multimodal, capable of generating text or voice and analyzing images, quickly and consistently, in the pocket, without depending on an external infrastructure.
The promise of access for all seems to be taking concrete form and offline uses are becoming possible.

On the other side of this focus, World Monitor takes the opposite path, it opens up to the immensity: geopolitical, economic, technological, environmental data and information flows in real time, on a single fully configurable interface. Open source, installable on one’s own machine: there is no longer an editorial intermediary between Le Monde and us.
It is a monitoring tool but not quite like the others: where a traditional aggregator offers a pre-built selection, World Monitor requires you to decide for yourself what to monitor, how and with which filters. Setting up becomes an editorial gesture in its own right: choosing your sources, geographical areas or alert thresholds is already a form of looking at the world even before reading the results.
I find this reversal interesting: we no longer receive a feed pre-filtered by publishers, we build our own. It’s a real freedom but it’s also a new responsibility, that of knowing what we’re looking for, where, why, and how to read what’s coming up on this giant painting.

These two tools are “available to all” but available is not synonymous with “accessible to all”. Gemma 4 requires a recent phone, powerful enough to run a multimodal model locally, it requires an understanding of how a generative AI works and mastering machine instructions. World Monitor requires you to know what you are installing, to know the sources chosen, how to configure it, and then to know how to read raw streams without drowning in them. Even if open-source and free are necessary conditions, they are not enough to reduce the possible divide.
In my opinion, the digital divide does not disappear with these tools, but it is transformed. It is moving from technical and financial access to literacy: understanding what a language model is, what we set when we choose our sources, what instructions we transmit, how we analyze the results. And these are skills that cannot be acquired in a few clicks…
The model and the world in the pocket, yes. But in whose pocket and for what concretely? This is perhaps the most important question…

The two seem to me to point in the same direction: not to make AI more impressive, but to make it closer, more personalized and to make us more autonomous. This is certainly an essential step, but technical proximity does not create understanding, and availability does not guarantee access.
What if what remained to be built was no longer the question of the availability of applications, but what will allow each and everyone to decide if, why and how to take advantage of them?