Towards a "Free AI Manifesto"
3 comments
Your heart's in the right place, but if making Free Software was as easy as drafting a manifesto then Linux would have photoshop alternatives a hundred times over.
People will continue developing closed-source AI because AI is expensive to train. Someone's gotta foot the bill, and it's not going to be the Open Source community doing that for quite some time.
People will continue developing closed-source AI because AI is expensive to train. Someone's gotta foot the bill, and it's not going to be the Open Source community doing that for quite some time.
> People will continue developing closed-source AI because AI is expensive to train. Someone's gotta foot the bill
But if it's local, we don't need to rely on the compute of third parties, we use our own compute. On-device AI will replace all these companies doing stuff in a remote black box we have no insight into.
But if it's local, we don't need to rely on the compute of third parties, we use our own compute. On-device AI will replace all these companies doing stuff in a remote black box we have no insight into.
Your own compute probably sucks. It's possibly some variant of CoreML, supporting a select few ops up to whichever featureset Apple considers supported. Or maybe it's some version of Microsoft's backend, either ONNX with a custom executor or DirectML with a possibility of native CUDA drivers. Maybe you're a Linux user, maybe you're on an iPad. Potentially you are an unnamed variant of ARMv7 supporting NEON virtualized through a KVM. Do you have a GPU? What's it's BLAS coverage? Does it work upstream? Does it support old code?
This is why Nvidia ships the AI of today. In an ideal world you'd be right, and our big-tech benefactors would work together for the common good. In the real world, I wouldn't bet a single dime on FAANG's cooperation. You and I, our common wellbeing, is always a secondary consideration to them.
This is why Nvidia ships the AI of today. In an ideal world you'd be right, and our big-tech benefactors would work together for the common good. In the real world, I wouldn't bet a single dime on FAANG's cooperation. You and I, our common wellbeing, is always a secondary consideration to them.
Free AI manifesto
Artificial intelligence is fast emerging as a powerful technology in our lives. Now is the time to ask ourselves, do we want it to be controlled by corporations or by the people that use it? The risks of central control are undeniable — surveillance, dependency, manipulation. It is less clear how to achieve a future of open and free AI, but the ideas and experiences of the open source and cypherpunk movements show us the way.
At the core, the answer to the problem seems obvious: AI models must be owned by the people that use them, and nobody else (Mustafa Suleyman’s insight). This is how privacy and the alignment of interests between models and users can be guaranteed by the foundational design of the system, creating a strong bulwark against the overreach of centralized powers.
It is hard to imagine this to work for the massive, compute hungry frontier models. Where it could work really well however is in the one application where AI will most likely create the biggest value in the near future: The interface layer of direct interaction between AI and the user. If dialogue AI is good enough to consistently extract the right user intent from a spoken or written interaction and translate it into instructions for all kinds of software systems out there, a great power is put into the hands of the user through potentially quite simple systems. Still, if the history of open source tells us anything, it is that these solutions need to be and stay very good to achieve widespread use. Having a strong data flywheel and access to enough compute to keep up with corporate models are paramount for this.
We propose a solution that is at its heart about combining privacy-preserving, federated learning with a protocol that lets everyone freely exchange and monetize their model insights. By processing data locally and sharing only learnings – never the raw data – individual privacy is safeguarded. This method allows participants to contribute to and benefit from AI without sacrificing their digital sovereignty. What emerges is a barter economy of knowledge - a system where individuals enrich the collective intelligence through sharing. This network also allows for a privacy-preserving monetization of personal data, potentially supporting a sustainable model for funding computational resources. It's a tangible manifestation of the principle that when we share what we learn, the whole ecosystem grows stronger.
To bring this vision life, we draw on the collective spirit and experiences of the open source and cypherpunk movements, advocating for a community-driven approach to AI development. This path isn't new; it's paved with the successes of those who have wrestled freedom and autonomy from the powers of central control throughout time.
This manifesto is a call to action for the creation of an AI ecosystem that is truly by the people, for the people. It is an invitation to engage in dialogue, to collaborate, and to take collective action. We envision a future where AI amplifies human potential, guided by the core values of privacy, autonomy, and mutual benefit. By working together, we can construct a system where technology serves humanity, ensuring that the digital age is marked not by subjugation, but by empowerment.