AI: “Free Software” is essential to save humanity
With AI-augmented humans becoming a reality, the wrong choices could lead to the enslavement of humanity.
The current debate on the “openness” of AI models
For about 25 years, the concepts of both Open Source and Free Software have been mostly grouped together — even going as far as sharing a combined acronym (FOSS or FLOSS). And yet, the philosophy behind each movement is quite different. With the AI revolution in motion, one of the current AI public debates revolves around whether AI models should be Open Source or proprietary — with Yann LeCun (currently at Facebook) and Sam Altman (OpenAI) at opposite ends of the spectrum. This is not merely a “debate of the day” that will simply come and go, it has both much deeper roots and consequences.
In this post, I make the point that the current “Open Source” side of the debate offers a stance that does not go far enough, in particular by not providing proper protection to individuals.
The raison d’être of the FSF
Note: I won’t go deep on the FSF here, you can find more on this blog post I wrote 5 years ago part of a short series on Open Source.
Philosophical musings on the risks of AI have been part of both art and science for a very long time, but for 4 decades now, a very active movement has been at work to build, step by step, the software foundation that would one day protect humanity from the danger of AI: the Free Software Foundation (FSF). A lot that can be learned from the FSF in the current debate.
Back in 1985, a tech visionaire, Richard Stallman, founded the FSF. One of his objectives was to build a complete “free software stack”, the GNU project, from the ground up, operating system included. This vision got backed by a legal construct that was later formalized by the FSF: the GPL and LGPL licenses — some of the most widely used Open Source licenses, and currently used by well-known projects, including the Linux kernel.
The GPL, unlike a lot of the more permissive “open source” licenses (BSD, Apache, etc.), has a so-called “viral” behavior: if you modify, bind, or extend GPL-licensed source code with your own source code, then your own code becomes subject to the GPL and must also be GPL-licensed. This virality of the GPL makes sure that the “GPL empire” keeps growing (a “Lesser GPL” — LGPL — was later introduced to allow software vendors to simply use such LGPL code without having to GPL their own source code, as long as they provided the mean for any user to rebind to a different version of the LGPL library to their proprietary code — hence keeping users in power of what’s actually running). To many businesses, the virality requirements imposed by the GPL were too strong: this led to the emergence of more liberal open source licenses, which ultimately gave birth to the Open Source movement through the Open Source Initiative.
The reason such a virality is backed into the GPL is no accident! Richard Stallman, a big SciFi fan, had already seen, 4 decades ago, the threat to freedom that would come as “Augmented Humanity” develops (hybridization of humans with technology, the embedding of “software” in humans’ brains, autonomous AI agents, etc.). So, how does this relate to “Free” Software? Imagine if leading AI models were proprietary (code and weights), it means you’d have no way of controlling the behavior of their AI agent or, more importantly, in the context of brain implants, you’d have no way of controlling what software runs into your brain: what actual code is running? What is it programmed to do? What dataset has it been trained against? Once running in your brain, code could be triggered to make you do things against your will: this could lead to the enslavement of humanity (by profit, by a rogue power, through a bug, or through hacking — it doesn’t really matter).
Stallman’s vision, making sure that a parallel complete, and high-quality software stack gets developed under a license à la (L)GPL — assures you will always i) get access to its source code and that ii) you are both legally and technically able to modify it, recompile/rebuild it and re-upload it — into your brain: you get to have freedom. Hence the notion of FREE software, not just open, but actually free — and not just as in free beer, but — truly — as in freedom. The definition of Open Source by the OSI doesn’t require those rights to exist, the (L)GPL, on the other hand, does.
Current Impact on AI
The current debate on whether AI should be proprietary or Open Source misses the dimension that Richard Stallman has been touting for 4 decades. The debate currently revolves primarily around the ability of organizations to leverage, extend, and fine-tune Open Source models as they see fit, and this because the very ability to train a foundation model would otherwise be limited to a few organizations with very-deep pockets, essentially creating a de facto oligopoly for a function that will increasingly play a central role in our societies.
By focusing this discussion on “Open Source”, rather than on “Free Software”, we miss an essential part of the discussion: what should the freedom of individuals be? Even if core AI models were to be “Open Source”, this wouldn’t prevent an organization from taking a model (and associated source code — we’ve entered the Software 2.0 era), customizing and embedding it into their products (including brain implants), with no means for the actual end-user to have any way to understand what it does, alter it or replace it with another version. While Open Source enables vendors to have flexibility, it does not offer any protection to end-users! Nothing is said about the ability of a final user to understand what code/weight are running, what datasets it has been trained on, what code is leveraging that model, and, in case they don’t like the impact it is having on them, to be in a position to modify it or replace it with a different one — both legally and technically.
As such, if one truly cares about the impact of AI on society, it is essential to understand the rationale and philosophy that led to the FSF and the (L)GPL and shift the discussion from “Proprietary vs. Open Source” to “Free vs. Proprietary” models. Also, not just code must be taken into account, but additional elements such as the sources of training data, weights, etc. all need to be included. Different use cases can be satisfied by different licensing options, but, ultimately, if we care about the impact of AI on society, we need to care about the freedom of citizens, not just about the solutions offered to intermediary vendors of AI solutions.
Freedom!
Onward,
Sacha