10:54
Someone could just develop a respiratory virus, infect themselves, get on an airplane, or sneeze, and potentially unleash a new pandemic. Governments have worried that thanks to advances in synthetic biology, it is easier than ever to develop biological weapons. Now, AI is making it even easier. Arthur, you're predicting the end of humanity here. How might that happen? AI has become exceptionally good at doing biology. Leading language models have surpassed human expert virologists on all sorts of exceptionally difficult
questions, things like bioinformatics and troubleshooting complex experiments. The worry is that these same capabilities could enable novices, or perhaps people with some biological background, to access a level of capability that previously only existed in the hands of a very small number of governments. So, what role exactly is AI going to play? The essential role that many are worried about is that they will provide what's called uplift. They will essentially give people who don't know how to do science a boost. One expert who used AI in a study published last year to help him to help guide him through the trickiest parts of assembling polio virus, described the model as being like an infinitely patient tutor who had read every
scientific paper ever published. And so, it was able to guide him with the idea being that someone who's never laid hands on a pipette could have an AI assistant at their side. And when they get something wrong, they can ask, do it again, and in theory, eventually get it right. The evidence shows both that it's unbelievably helpful to some people, but then there are other studies showing that it isn't. It isn't. So, who exactly are we worried about? I mean, could I become a bioterrorist? Well, I'm glad you asked. In fact, because uh the industry has looked into this and recent study found that actually novices, I'm going to assume that you
don't have experience with sophisticated virological tasks, but novices don't actually get that much help from AI in the lab. On the other hand, someone with, say, a PhD in molecular biology, and that might be the person who's actually going to get the most boost out of uh an AI accomplice. If you have access to a tool that has read every scientific paper every published ever published, that is somewhat akin to having access to an infinite number of experts, PhD-level experts, who can approach the question from different angles, who can troubleshoot, who can brainstorm in ways that might replicate or mimic the effect of having very large teams of experts, which is, of course, the principal bottleneck for
bioterrorism. So, how soon could some kind of deadly virus be produced with the help of AI? Governments have looked into this. Uh last year, there was a major study which found that in terms of actually using AI to develop a totally new pathogen, something that doesn't have any characteristics that we know how to defend against, that would rely on access to some data sets that don't exist yet. What would be possible in expert hands now is some capability in terms of actually modifying existing viruses to perhaps exhibit certain capabilities, uh certain characteristics that they do not currently have. In
theory, someone could just develop a respiratory virus that works if they get lucky and infect themselves and then get on an airplane or sneeze in public places as they put it. And that could potentially do the trick, but of course until someone tries that and we certainly hope that no one does, it's a little hard to tell. But you could do this by mistake then. Biology compared to other ways of causing mass panic and harm is very hard. And the fact remains that there are all sorts of other ways of creating panic and creating harm that are perhaps easier for now. But of course with the way that these capabilities, these AI capabilities are evolving, that balance could shift and it could shift very quickly, which
is why there is so much interest in actually into in how to make models less likely to provide help for that purpose. So that's what I wanted to come on to. It's not easy yet to destroy humanity. We should take some comfort from that. But what can be done to keep it that way? Whose job is it to regulate and can you regulate for this kind of thing? It probably starts with the technology itself. The best option would be to have models that do not freely give out sensitive information about how to recover and propagate viruses, for example. That begins with models that will refuse dangerous requests. And in fact, models
are getting a lot better at refusing requests related to sensitive biological information. Uh but those refusal mechanisms can be bypassed. In fact, uh they can be very easily bypassed by a motivated actor. So, then you have to look at ways of actually changing the capabilities of the model. Maybe you leave some of that sensitive information out of their training data sets so that they don't know about these capabilities. Other approaches might come down to something that is simpler, which is just to restrict who gets access to some of these models. And governments, can they regulate? Governments can get ahead of the release of models by looking at their capabilities before they've been
released to the public. And on the basis of that, they could, in theory, establish guardrails or limitations on what the models can be used for and where they can be used and by whom. So, it's not over yet, but it's pretty terrifying stuff. Yeah, it's it's scary. Arthur, thank you very much. Thanks very
