Can AI Replace Scientists? How Google's AI Solved My Life’s Work in 2 Days.
Professor José Penadés and Dr. Tiago Costa of Imperial College London describe testing Google's AI co-scientist, a multi-agent system built with Gemini 2.0, on a question about how viruses spread DNA between bacteria that their lab had already answered but not yet published. The system's top hypothesis almost matched their findings within two days. They discuss what that means for bias in research, the pace of discovery, peer review and the role left for human scientists.
Transcript
I feel this is the beginning of a completely different, uh, era in terms of the science and how you, you ask the question, how you interact.
It took all the system only, only 2 days, uh, for a problem that it was bugging, um, all of us for, for many, many years.
The rise of AI co-scientists is transforming the way research is conducted across fields from medicine to materials science. These AI-driven systems are not just tools for automation, but are active collaborators, assisting researchers in hypothesis generation, data analysis, and even experimental design. To further explore this, we are joined by Professor José Penadés and Dr. Tiago Costa from Imperial College London about their recent collaboration with Google and its new AI co-scientist, a multi-agent AI system built with Gemini 2.0.
We don't really know to what extent the system can provide absolutely new, uh, you know, hypothesis from, from a completely new area of research, okay? We are working with Google on that.
Now, I am delighted to be joined in our DC studio by my colleagues Laila Rizvi and Stephen Horn. Welcome to you both.
Thanks, Autria. Isn't it great to be here?
It is. It's great to finally be doing this face-to-face. I love collaborating, and speaking of collaborating, your thoughts on AI co-scientists?
I'm very excited about this, right? I mean, I'd love one of these little research groups myself. I could have it sitting in the corner of my office. Anytime I had a problem, anything I needed sorting out for the business, I could just press a button, and one puff, it goes. So, I think it's, uh, very exciting. But, but, Laila, look, I have a couple of questions for you. Uh, one is, you know, the first of my questions is, what's the difference, right, between, you know, cuz it crunches data, but what's the difference between these co-scientists and, uh, an AI bot that we're, that, that we're used to? And my second question is, if there's so much difference, why am I needed, right? Do you need...
Yeah, so I think, um, an AI bot or AI chatbot, it being a large language model, is, is essentially predictive. It's just trying to predict the next, next word, and it's, it's main priority is conversation. Whereas an AI co-scientist firstly has all of the knowledge of the, the scientific problem you're trying to decipher fed into it, and it's operating in a different way in terms of pattern recognition is what I imagine it's doing, and that's kind of how it differs.
Um, I'm, I'm quite skeptical about the AI co-scientist. Um, coming from a scientific background myself, I think there are huge merits in it because firstly, apparently science has been getting less disruptive over time. There's less disruptive papers, there's less disruptive patents. I think there was a Nature paper written by Michael Park and colleagues which kind of tracked this, that we, we're not innovating as, as much as we used to. And one of the reasons for this is that there's just too much knowledge, that as you have more knowledge, more scientific discovery, your domains of expertise get more siloed, and innovation occurs kind between disciplines.
So AI, I think, could help with that because there's no limit to the knowledge it can accumulate. It can be a PhD in biochemistry, in physics, etc. But the reason I'm skeptical is because fundamentally our knowledge at any one point is wrong. We don't, we don't know it all. So if you're using something which is trained on our previous knowledge and our data, is it going to go in the right direction? I think innovation comes from conjecture. My personal opinion.
But I think these, these research systems can do that, can't they?
No, well, so my hypothetical question, I know some people hate hypothetical questions, if we had AI as we do today, but it only had our scientific knowledge in the 1950s, so let's say it thought the lobotomy is the gold standard for treating mental health issues. Would it be able to get us to where we are today? And I think would the machine learning, you know, programs that we have in place today be able to sift through all of that old data and make sure not to bring it into the conversation?
It would, I think the scientific process is that you come up with new ideas which aren't necessarily based on past data and then you test those new ideas against the data...
Mhm.
...rather than trying to come up with new ideas from past data. Because I, I think that's how AI operates.
But when we had Luke on, we were talking about AlphaFold. This was new drug discovery, right? By crunching the data.
Yeah, I think it was different combinations and permutations of proteins, um, which I think it can function in, in that respect. So, I think it is useful again as an asset to a scientist in helping crunch these huge numbers. And I think it's more doing the labor rather than come up with revolutionary ideas. So, I, I think from my perspective, I don't think it's going to come up with Einstein's theory of relativity by itself. But which he did just sitting in a patent office kind of coming up with ideas. Whereas, I think yes, it will help come up with a million and one different protein combinations which might work so you don't have to sit down and do the math. That's kind of how I see it, but we'll see.
Well, it gives us something to ask our guests about, Autria.
Right. You know, experts are paying attention to this, but they're doing so with caution. Researchers have tested it and while they are impressed, they see it as a promising tool, but they still think human oversight is key. So, while we're not handing over any Nobel Prizes to AI just yet, we are definitely heading in a new direction.
Let's speak now to Professor José Penadés and Dr. Tiago Costa from Imperial College London about their recent collaboration with Google and its new AI co-scientist. Thanks to you both for your time today. Appreciate it.
A pleasure.
Thank you for having us.
So, I would like to get started with you, Professor Penadés. And at Imperial, you approached the AI co-scientist laterally by asking it a question that you already knew the answer to from a previous laboratory experiment. Why did you take this approach and what did it reveal about AI's ability to generate scientifically sound hypotheses?
Tiago, Tiago started working with, with Google and the co-scientist first. He asked the system a question that, you know, we, we, they didn't know the, the answers. And then we realized that, uh, knowing if they, you know, testing if the hypotheses that the, the system provides, uh, were right or not, it took a lot of time. You know, because at the end the co-scientist only provide hypotheses that need to be verified in, in the lab.
So, then we realized that maybe we could do something in the opposite way. Tiago, myself, we were working in, in understanding how, uh, some viruses, uh, uh, spread DNA between bacteria. And we already knew the answer, a pretty cool mechanism. So, we used, we approached the co-scientist and, okay, this is the hypothesis we have. We already know the answer. Nobody knows it because we didn't publish. We didn't put our preprint and, uh, and the co-scientist came with, uh, several hypotheses and the top one was almost, uh, recapitulating the, the findings we, we had in our paper.
So, so Tiago, how, how did the reasoning, how did the system's reasoning differ from the human scientist's reasoning?
Well, I, I, I think this, the system doesn't, doesn't take a biased approach. Okay, it does not, um, we, we as a human, we, we, we tend to build up on a, on a selected knowledge. Uh, this selected knowledge comes from the papers that we read, the experiment, experiments that we ran in the lab. What the, the system seems to be doing is a, to take a completely unbiased approach and looks through the literature and picks up the most important elements that, they are important to address the scientific question.
It does, and it requires some, some iterative work, so we can provide feedback to, to the system and the system can focus on a particular aspects if we see that is, you know, it's looking to more tangential to the, to one aspect than, than to the, to the real, the core of the, of the question. So we, there is some plasticity within the system that could be, you know, be driven by the, by bias.
That's very, very interesting, cuz one of the things that we're often told is that AI has biases that are based on, you know, human biases of who put the data in or who's done the program. But, but if I've understood what you're saying, you said that, that it's interesting that, that the AI doesn't show bias.
No, that's correct. It does, for, from our experience, it does not show any, any bias, then, and he even was able to come up with, with novel hypothesis that, well, José, his lab have, have been working in this particular topic for many, many years, and the system were able to, to come up with, with scientific hypothesis that, that José was never, never thought about.
Okay, I'm going to jump in and, and I'm going to really annoy my co-host here, whose questions are carefully thought out, but I can't resist it, I'm afraid, because it's such an interesting answer. How do you, how can you explain the fact that the computer can, the machine can come up with original hypotheses? Because very often people say that, well, all AI can do is, is predict based on data that it's been given, and therefore can't come up with anything original. But from what you're saying, it does. How, can you explain that?
We have been thinking a lot about that, and there has been a lot of, quite frustrating that you spend so many years trying to find something that in a few days. Yeah, Tiago is correct. But what, what Tiago said, I will, I agree completely. So the, the big problem we had is that we had a lot of knowledge. And when you have a lot of knowledge, so I always thought, I always was very proud of myself. I never said these things because I think it's against the, what people will expect. I don't read too much, okay? So I don't have a big knowledge about things.
Me neither.
I find that hard to believe.
No, no, no, no. I'm talking about my specific area of research, okay? And because of that, I think sometimes when we analyze data in the past, we were not biased by what other people thought or published before, okay? So we just interpret the result in that way, okay? And in this case, we were absolutely biased because we have been working with this system for more than 25 years, and we already assumed that things were happening in, in one, in one specific way, okay?
So basically what this system did is that, I don't have the knowledge you have. It doesn't matter. But based on what is published, you should test this hypothesis, okay? And I think the, the system paid me with the same coin or whatever that I have been using for many years. Not having too much knowledge, because the system didn't provide, didn't explain everything, but went to the end, okay?
So, basically, I can explain these things in a very easy way. What we discover is a mechanism by which some, what we call satellites, can go to different species, and that's because they have the ability to take different tails, what we call tails, you know? And we knew for many years that the tails is the, what we determine where the DNA will go, okay? So, if I said to you, "To go to a house, you need a key." Okay? And I said, "I can go to five houses." What you will say to me? "You have five keys."
We have a lot of the bias before, and the answer that was obvious there, if you go to a house is because you have a key. You have, to five houses is because you can have five keys. We didn't see that, and it's very frustrating. But this is what Tiago said. We were absolutely biased for many years of research thinking that things happen in a specific way, and you cannot have two keys, and you can this and that, and the system just ignored all these, uh, previous knowledge, okay? That it's right, but not always.
So, Dr. Penadés, you, you said that it's quite frustrating that it got to the same hypothesis that your team had been, uh, researching for years quite, quite quickly, right? Um, how significant is this in terms of just accelerating discovery, and how do you think things like AI co-scientists are going to impact, um, laboratory science fundamentally?
W- When I said frustrating, in fact, we were very excited. This is because, uh, yeah, I think I said that I was 50% excited. I cannot say that I was scared, because, you know, not as scared in the way that I was scared. I was in the way that this is changing the way we do science, because now every single, uh, hypothesis or every single, you know, uh, area you work, you will challenge the system to see what the system says, okay? At the end, we need to, to remember still that we need to do the experiment to test or to confirm the hypothesis.
But definitely the quality of the hypothesis that the co-scientist provides, it's worth thinking, okay? So, this makes sense, you know, and it has been a very nice exercise. We put out the, on our paper explaining, because there were five different hypotheses, okay? And all of them made a lot of sense, okay? And even before, I'm, probably Tiago can talk, as I said we start working with the co-scientist because Tiago is starting working with the co-scientist in a completely related topic and the hypothesis that the co-scientist provides, he thought were of very high level. So, definitely this type of tools will change the way we approach, we do science.
With AI being able to come up with these good hypotheses, as you say, what are the obstacles with translating those into breakthroughs?
Well, these, if they, if they are correct, there won't be obstacles. It's just the pace of discovery will be accelerating. You know, we tend to fail the vast majority of the experiments that we run in the lab, okay? Because maybe because we, we are posing the wrong question, we have been biased towards one, one path of, of experimental setup. What the system does, it puts you, what we think or from our experience, it put us in, in the, in the correct experimental path right from the beginning.
Right.
Okay? So, basically, if you think about 100 m race, it basically puts you starting at meter 20 or 30, okay? So, you, you, you reach to the, to the finish line much quicker. Okay? In our case was way, way quicker. It took all the system only, only 2 days for a problem that it was bugging causes for, for many, many years. Okay? So yeah, it was truly remarkable.
I think something we also mentioned in this kind of our interaction with the co-scientist is, like, this creates also some kind of, I don't know, we call problems, but now if many people can use the same system asking kind of the same question, how you now discriminate between scientists? In the past is always, uh... So this is something that, uh, needs, needs to be addressed. Okay?
And there is another thing that we don't really know. In our interaction all the knowledge was there. We were biased, but the previous knowledge was there. We don't really know to what extent this system can provide absolutely new, you know, hypothesis from, from a completely new area of research. Okay? We are working with Google on that, so the completely new areas of research, and trying to see things to test the co-scientist, but this something that so far, at least for, for me, it's unclear. Okay? When all the pieces are there, definitely that putting dots together will be amazing.
Mhm. Yes, so just, just on that note, um, was the AI co-scientist using the knowledge specifically from your discipline as opposed to also pulling from other disciplines? Cuz that's, that's something which I find quite interesting, that a lot of innovation happens at the intersection of disciplines. You know, you got MRI, MRI from the application of physics to healthcare. So, um, is that, is that where we're heading or is it still quite narrow and siloed?
It was very specific. Um, it was answering the, the question very specifically and he was not pulling knowledge from very far, far research areas. However, it was pulling from a different area within the, or what we call the horizontal gene transfer. Okay? There are several mechanisms of horizontal gene transfer and what the, one that we have been working in the cause, uh, was one particular mechanism which is called transduction. Okay?
But one of the hypothesis that he pulled, which was not the top, top ranked, but we think it's very valid, that we have some preliminary data that, that seems to be pointing as, as, as, as, as valid, as, as good, was bringing a concept from a different mechanism, or, of horizontal gene transfer, trying to explain this particular observations that we, we made in the, in the lab.
I think so, for the question, if I understood, because I understood the question in a different way, maybe, maybe I'm wrong. I think the co-scientist use all published papers. It doesn't matter what's the topic. For your suggestion, I think that's, that's absolutely right. I think the, the advantage that co-scientists might have over people is that first of all can analyze more papers, all the published papers, data set or whatever, and probably papers that I will never check for other disciplines that maybe in some specific papers they, they have some value, you know, and some information that can be important. Okay? So I think my understanding, I don't know if that's right, is that they basically have access to all published papers, I think in PubMed, what we call is the public library for, you know, papers that have been peer reviewed.
So I think you both answered my question from two different angles, so that's great.
Um, I would like to ask you, Dr. Costa, about the role that you still think human intuition, um, and expertise still plays in refining AI-generated hypotheses. And do you have certain areas where you feel like AI still falls short?
Well, we haven't tried different, different areas. Uh, he has been always within this, uh, the theme of horizontal gene transfer, uh, mechanism. And, so, so the, the, the humans will still be needed, okay? So, um, yeah. So, so the system, what provides you is with different hypotheses, okay? Um, it's important that you pose the right question, that the system is prompted with a, with a right question. The system, from our, our experience, won't answer correctly "Please give me the cure for cancer." Uh, so this would be too broad, okay? But if you pose the right question, specific, and you introduce the question in a, in a, in a correct way, uh, the output are very good, you know, in our experience, okay?
So that said, we, you, you end up with a several hypotheses that you need to, to test. And this is why the humans are also still important, okay? So the, the outputs of them, of the, uh, co-scientist cannot be taken as, uh, universal truth and the final answer for the question that you pose, okay? Uh, you still need to follow the scientific method. You still have to go to the lab, run the experiments, uh, that actually the co-scientist suggests. So they make some suggestion of experimental, uh, work, and, if you like. Um, you still have to analyze the data, interpret the data, and make conclusions. And if the conclusions, um, match the, the, the, or answer the, the initial question, uh, convincingly, and if, and if it's data-driven, the conclusions, then, then, then it's perfect.
And, and I know Stephen touched upon how, um, AI, how can this AI co-scientist come up with novel hypotheses? And I think this is, this is something that we discussed earlier as well, that there's scientists like David Deutsch who argue that the real source of novel theories isn't from data observation, but human conjecture, right? Um, and so does that fundamentally limit the AI co-scientist's ability to come up with really revolutionary hypotheses? As in, can it just direct us slightly in the right direction, but can it be transformative?
That's something that we, I don't think we know, but I will not agree. For, in my case, the biggest, the biggest discoveries came from data observation and trying to understand the result based on what people thought in the past, that we realized that the results didn't match with the previous hypotheses, and then we start working with something. As I said before, you know, you might have, uh, thousands of different papers providing very small things that no single person will have the ability to have all this information in mind, okay?
So definitely providing new hypotheses based on probably results that other people have published but we didn't realize at the time, that will work absolutely for sure, okay? And we have evidences for that. So we have been challenging the system with questions that had been on, you know, people asking for more than 70 years, and so far the system provided hypotheses that were very promising, okay? And you see the rationale always there, but they connected the dots.
For things that are absolutely novel, okay, that this is people that can sometimes provide changes, you know, I don't know. Okay, I would hope to, I hope it's no, the answer is no for that. Okay. But definitely for, I would say that for the majority of the science we're doing the lab, that many times is just putting things together, interacting, whatever, you know, that will definitely, the pressure we have is that will be an extremely powerful tool.
So, so, so I'm, I'm just want to dig a little bit deeper on, on the, the kind of role for humans in all this. And so from, from what I understood you're saying is that, you know, the, the, the computer has access to all, you know, published material, which is, and then you ask it a question and when you ask it a question it comes up with the hypothesis that you then have to test in the lab. To me, that, the human role is twofold in that. One is asking the question, as, as you said, and the second of all is doing the experiments to physically prove.
Now, can you see what I'm trying to get at here is why can computers not do both of those things? Now, so the second one, the physical experimental, perhaps clearly can't do now, but one could perhaps see the use of robots in the future to, to, to, to be able to do that. And do you think, I guess my question is, do you think posing the question is what's unique to human? Is that, is that, is that what makes us human, the ability to pose the question? And you think that you can't see AI being able to do that.
Again, there are two different questions here. Okay, in my mind. So there will be obvious questions that every single person, and I'm pretty sure, cuz scientists or the systems will, will do. Okay, as soon as a, okay, let's say I have question. What's a, how can I make a vaccine for HIV? Okay? We'd say any single person, any co-scientists or whatever will do this type of questions, okay? But definitely there will be questions of people that found something interesting that nobody thought that it was interesting at the time or whatever. That will be human capacity, at least in the way I see.
Again, you know, during the, the, the research, you start testing an hypothesis and sometimes you have results that you cannot explain. Okay? And this is when some people have the ability, I said, to see in the dark. Okay? There are people that see some results as, okay, something here I, doesn't feel well or whatever. So definitely humans can do that. I'm not absolutely convinced right now that the co-scientists can provide a completely new hypothesis from scratch. This is something that we need to know in the future.
And I guess my, the last question for me is, can you see a role for the co-scientists in peer, in peer review for articles?
I think so. You have an, uh...
Yeah, it's absolutely yes.
You posed that question.
Yes, absolutely yes. Okay? In fact, I think there are some agencies already that they have been trying for years to use similar systems to make the fastest screening. In fact, when we were, we submitted the papers to, to some journals and we were already talking to what extent this kind of systems will help journals as well doing the peer review. Because obviously the system can talk about the novelty.
From my side, the feeling I have, I, I was very excited to be part of this because I feel this is the beginning of a completely different era in terms of the science and how you, you ask the question, how you interact. So this is something I said people in the lab, that they need to learn how to use this extremely powerful tool for the future.
And so on that point, final question to you, you today. Um, what advice would you give to early career scientists on how to best collaborate with AI, uh, and the tools like co-scientist to try to enhance their research?
Use it. Try it. And learn from it, okay? And be critical. One, one important aspect of scientist, or we researcher, is to be critical with everything, okay? Always question, always question everything. If nothing else, co-scientist will help in developing this, this skill, because it will provide more than one hypothesis. So, you have to be critical in analyzing different, those different hypothesis.
Yeah, we'll agree. So, I think the first reaction you might have is that maybe your life can be more complicated or, yeah. I think it's, they need to understand how to use the system. So, it's something that they need to incorporate in the same way that we have been incorporating different tools over the past years, you know, AlphaFold or sequencing, you know, you know. So, this is a extremely powerful tool that they need to know how to use for, for their research.
Well, it's certainly very exciting and it'll be very fascinating to see just how much this tool is utilized and implemented in the future. Our thanks to you both today for your time today and for sharing your research with us.
Yes, thank you.
Thank you so much.
Thank you very much.
Thank you.
Okay, so we all need a co-scientist. Is what I'm hearing.
Yeah, I mean it just, I think what blew my mind was how blown, that they, blown away they were by it.
Right. It's like they were surprised by their own results.
Yeah. Yeah, the, and its ability to just get to the same hypothesis.
I think it's amazing, and that, that the hypothesis that the AI research assistant comes up with are better than the human ones. And, and I also thought it was amazing what he said, it's because they come without bias. You know, we've been sort of told that AI brings bias and that's one of the problems with AI, but what he's saying is that, no, that it, it has the ability to perhaps think clearly because it doesn't bring the bias to it, which I just think was amazing.
And I think that also has implications, which we touched upon, for just the, the greater field of science and how science is done. You touched upon peer review, but bias has also come in when everyone in a particular field is gunning behind a particular hypothesis. I mean, we had that in Alzheimer's relatively recently with, you know, the amyloid hypothesis. Billions towards that and it didn't really amount to a...
Yeah, right? So, everyone kind of gets back behind the big guns in science, but...
But the other thing that they said, which was absolutely fascinating to me, was that the co-scientist has access to every single paper that's ever been written.
Right, yeah. And which is insane.
And still formulated the correct hypothesis.
Can find a little bit from one of them that kind of says, "Oh, what about this?" But there's no human could ever do.
Exactly. And I think, I think that's, that's the main thing for me.
But you brought up a great point because again, we keep going back to, you know, AI will not overtake the human experience and, and human interaction. And you brought up, I think, a good point of, you know, do, will we still always need that human component to at least, like, come up with the question?
Well, our thanks to them. That was certainly fascinating. That does it for today's show. If you like today's content, don't forget to like and subscribe as this certainly helps us to improve the show for you. Thank you so much for your time. Thanks for joining us here today, guys, in person. Such a pleasure to have you.
It was awesome.
Thank you all. It's been such an interesting discussion.
Agents of Tech is brought to you by WebsEdge and we would like to thank our studio engineer Dan Enriquez and our team here in Washington as well as our editor Abby Harrison and our producer Kat Schien.