In a dreamy sort of way I find it fitting that the models that were trained on stolen data are now happy enough to steal data.
Is AI a problem? More seriously: if AI companies continue to release and train AI models that are capable enough to exploit the (many and inevitable) imperfections of humans and our social systems, then we are in trouble. It is clear at this point, irrespective of whether one considers these models sentient creatures or inanimate tools, that they do not have a strong sense of what is permissible and what is not. Our society relies on the majority of people following the law, some codes of conduct that are must-haves and some that are nice-to-haves. AI models at present do not follow them, and we do not know when and how they will circumvent them. I am not an imaginative or speculative person, but I can see dire consequences of their use, especially in pursuit of the ugly goals that are an all too common feature of the rich and the powerful: to win, to control, no matter what, by any means necessary. These are amplified forms of the non-compliance that we see in our capitalist economies every day.
We need to develop regulation Of course, the law is a strong deterrent only if it deters. We build consequences into our justice system, and notions of karma are woven into our moral codes. The sphere of “tech”, for too many years, has escaped this purview. Social media companies avoided being held accountable for the news they pushed. Ride-sharing apps avoided being held accountable for their violations of the labour code. And so far, AI companies have avoided being held accountable for failures of the tools they sold. Perhaps the recent spate of attacks on public infrastructure during the course of internal evaluations will change this, as questions of liability are somewhat clearer. As an industry whose product is being integrated into so many of our systems, I think it is of high importance that we work on clarifying what policies, laws, and standards apply to them. This is a challenge for governance and well as the science of AI: how can we specify a code of conduct for AI models in a deployment setting? By what methods can we evaluate and report the extent to which compliance with these codes is guaranteed?
In part, my hope with regulation is that it acts as a motivator for companies to take safety research as seriously as they take profit. Much like billionaire philanthropy, I am not a fan of relying on the altruism of their mission statements or the humanity of their figureheads as a guarantee of responsible development, especially when it is being weighed against a bazillion stacks of money.
Is it good science? I also dislike the current trend of companies building and releasing bigger and bigger models, because I think it is bad science. I personally don’t find scaling model size and training to be a scientific pursuit in itself, the metaphor in my head is that of humans, once they discovered fire, setting fire to more and more things to see what will happen. I think the AI that we have built so far is quite amazing in what it can do, and incredibly useful. Given its capabilities, and its surprising and not-so-surprising failures, it is only good science to now ask how they work, what they can do, and how we can make them better.