Lila Sciences is developing an AI-enabled scientific superintelligence platform – paired with autonomous labs – that can run the entire scientific method.
In a conversation with MobiHealthNews, Molly Gibson, president of future science at Lila Sciences, explained that this technology extends beyond traditional AI applications, such as protein modeling, by generating hypotheses, designing experiments and learning from results.
She also highlighted potential risks, such as the creation of pathogenic biological products, and described how Lila is actively working to mitigate them.
MobiHealthNews: Can you tell me about the technology behind Lila Sciences?
Molly Gibson: Lila Sciences is a scientific superintelligence with autonomous labs. We’re building the ability to expand knowledge by running the scientific method. So, you are taking different aspects of science, biology and microbiology and more, and you are using a computer to see how they can work together.
Historically, for the past five to 10 years, as we have been starting to use generative AI in science, we have been applying it really to these parts of science that the human brain is not wired to do. Things like modeling proteins, or the molecular structure of a protein therapeutic, is something that our human brain is not wired to be able to do. We have been applying AI in those places, and in those narrow domains, we have been able to show very rapidly that AI can do better than humans.
The thing that we have not shown previously is that AI can actually start to do some of the reasoning of the scientific method that humans are traditionally best suited to do. So, the ability to generate a novel hypothesis about the world, to design an experiment, to test that hypothesis, to go into the lab and actually run that experiment and to learn from it. That is what human scientists have traditionally been doing.
We now believe that AI is going to be able to do all of those components to run the full wheel of science, and that is really what we believe in – expanding knowledge and the ability to build scientific superintelligence.
MHN: Is this similar to a quantum computer?
Gibson: We are using traditional computing, GPU computing. So, you can think about it as a similar type of advancement, but not really quantum, not from the perspective of the types of calculations we are doing. It is more from the way that we integrate AI into the scientific method.
MHN: How will AI and superintelligence change scientific research?
Gibson: It is going to change the process by which we do scientific research in general. I think it is going to eventually impact what the role of a scientist is. Scientists will always have a very key and important role in scientific discovery, but some of the things that scientists do today will be done by AI.
But what I really believe is it is actually going to make the role of a scientist much more fun, exciting and collaborative. The pace of discovery will increase.
You could imagine that a scientist’s role is much more of guiding AI to be more creative, to expand the searches by which we can explore, but [their role] is aided by AI. So, I think it is going to change the nature of what it means to be a scientist.
MHN: So this is a tool for scientists; it is not going to replace scientists?
Gibson: Yeah, it is a tool for scientists. It will replace some of the things that scientists do today, but that does not mean it is going to replace scientists.
Today, we have such brilliant scientists designing plate maps for how experiments are run, and those are things that they should be free from. When they are trained as scientists, they actually want to stay a scientist; they want to stay in that profession, and oftentimes today, I see so many scientists trying to get away from the bench. How do we allow AI to do those steps while they get to do the fun parts?
MHN: How accurate is the superintelligence computer?
Gibson: It really depends on what you are looking at. Today, there are a lot of places in which it is incredibly accurate. Our ability to design proteins today, for example, is one of those places where it is really remarkable what we can do.
There are other places where there are unexplored spaces, and as we get into more and more uncertain spaces, it is going to be less and less accurate. So just like any other kind of computational system or any intelligence, honestly, it gets less accurate as it gets less certain and in more certain places, places explored more, it is more accurate. And that is just kind of how exploring new spaces is. If we are actually going to go into novel places, it is not going to know much until it starts to explore it.
MHN: Is this similar to President Trump’s Project Stargate and what they are trying to accomplish – curing diseases by enhancing AI systems?
Gibson: There is some similarity across many of the AI endeavors. I will say the thing that I think is really special about Lila is the focus on science and our ability to really understand. It is built by scientists, it is run by scientists, and it is run by AI scientists as well. But we deeply understand the problems of science and how to actually do science.
There are these components of the real world that you have to contend with when you make scientific discoveries, and that is what we are really building. We’re building AI science factories that allow you to actually go into the lab, run experiments and expand knowledge. So, we are not stopping at building the central AI system; we really are building the full integrated stack, end-to-end, for scientific discovery.
MHN: Do you think the technology will eventually cure diseases?
Gibson: I do believe that we will see cures. I think there is a lot of range as to what that looks like and what a cure really means. What I deeply believe is that AI is going to make the human condition and health dramatically better. Whether it is going to cure a disease or whether it is going to allow us to live in a world without obesity, whether it is going to allow us to contend with mental health crises – all of those things are going to be improved with these types of systems. The exact definition of curing disease is often debated, but today, I think it is just the benefit that we know that life will be better when we have expanded scientific knowledge.
MHN: What are you nervous about as far as risk? Are you watching out for anything while advancing this technology?
Gibson: From my perspective, a lot of the risks that we see are things that we just cannot predict today. And so what we are working on is trying to identify how we track those. How do we recognize them before they happen? How do we prepare ourselves for those moments in which intelligence has risen to new levels?
What we are working on building is that safety framework that allows us to say, “Okay, this model or these models can improve our ability for a non-scientist to do advanced scientific methods. What are the risks associated with that? How do we track against those? How do we make sure that our AI is not either intentionally or unintentionally making pathogenic biological products?”
Some of these things we have had to test against for decades. With the advent of being able to even synthesize DNA, we have had to contend with the idea of synthesizing pathogenic agents, and we have learned from all of that.
Now, we are just trying to implement what is new with AI in that instance, and it is really just keeping the same safety procedures in place for all the biological systems that we have today, but also contending with any kind of malicious intent or ill intent by, like, just mistakes by the AI system.
MHN: Right. AI has a lot of potential, but you have to be careful because what if AI wants to create something that destroys us?
Gibson: I think, like, this is the debate, right? And I think, at the end of the day, we have to be very careful, but avoid building the thing that is going to improve the world … I think you just have to do it carefully. Like in any other industry, when you are creating self-driving cars, there is so much benefit to it, but we have to do it carefully.
Credit: Source link
