I want to take a moment to acknowledge someone who is making a difference in the fight to keep the boldest, best new technologies open source. It’s David Siegel, someone with MIT connections, and someone who has seen this whole business take off like a rocket ship over the last few years, but who was also working quietly behind the scenes in the decades prior to the new AI era and models like GPT.
Siegel got his Ph.D. at MIT, and worked in the lab at CSAIL (the Computer Science and Artificial Intelligence Lab) which is now directed by my friend and colleague Daniela Rus. Siegel also has extensive business experience as co-founder of Two Sigma, an AI company in the financial sector. But in a way it’s Siegel’s “soft experience” that’s even more resonant with people who care about the directions that AI is taking. He’s been vocal about the big issues: AI’s effect on democracy, on education, and on our society as a whole.
Keeping AI Open
Lately, Siegel has been writing and speaking about one of the biggest debates of our time around AI: whether to release open models, with both training weights and source code, or to shield one or the other, or both, in a closed model approach. Obviously, the latter is often tied to proprietary company ownership of a system, and open source tends to be managed by a public community of users and super-users. But there’s also the reality globally, that American firms tend to favor a closed approach, while many of the top Chinese models are released openly.
In any case, Siegel makes certain arguments for the open source approach, citing old conversations with people like Richard Stallman of GNU fame in the 1980s, and suggesting that closed source models means a few powerful hands will have control of what should be a public good.
Siegel invokes the concept of the traditional library, which I think is a good one: we would find it intolerable, he suggests, if powerful interests controlled all of the books, what gets written, what gets distributed, and what gets read. Our codebases, he noted, should be like a library: free, and public access.
Siegel is also good at knocking down some of the common reservations around open source AI, primarily, security concerns. As he and others point out, there are practical ways to safeguard systems, and even close source models can be jailbroken.
The other side of the coin is that, by making systems open source, makers reveal more about how they work. And we need to fight this ‘black box’ issue, where the AI becomes a magic tool that spits out results enigmatically.