Googles New AI Co Scientist Just Changed EVERYTHING (AI co-scientist Explained)

So today we're going to be taking a look at something that has actually taken the industry by storm, and that is of course Google's co scientist. I think this is a story that is remarkable in terms of what it's able to do and it gives us an indication as to where we are headed when we look at AI in terms of its ability to be able to help us achieve scientific discoveries and of course achieve efficiency in researching ways to effectively treat people. So essentially the co scientist is a multi agent AI system developed by Google Research and this is designed to help scientists generate new hypotheses and research proposals. So basically what this does is it acts as a virtual collaborator which is capable of understanding complex topics and suggesting new research directions. Now this thing is pretty crazy because there have been some real world implications which I'll get into later. But first we are going to dive into exactly how this works. So you can see here in this short demo, and I'm going to get into a little bit more of the details in a moment. But you see right here that we have the scientists, so we have the scientist who is of course the person that would specify the research goal. And then of course you would actually specify the research goal to the AI co scientist, which is made up of entirely sub agents. So you've got a supervised agent, a generation agent, all of these agents, and then of course you essentially have all those ideas and those ideas compete against one another and eventually you have the best one and then that one of course is proposed. So how this works in terms of what each agent does, the generation agent is going to be the one that creates new hyper hypotheses, the reflection agent. This is basically the agent that is going to be the one that evaluates and refines the ideas. This is where you can see it has full review with web search. So it's able to search the web, it's able to do a simulation review, a tournament review and deep verification. So all of these things are really good for validating the idea. Then of course you've got the evolution agent, which is basically something that improves the idea over time and gets inspiration from other ideas. And then of course you also have the meta review agent. This is where it ensures accuracy and scientific value. And then of course we have a ranking agent. And this is I think one of the most important thing because this prioritizes the best research directions. So it says right here that the research hypotheses, comparison and ranking with scientific debate in tournaments and basically it just continuously ranks the ideas and of course with the reflection, with the evolution, it basically over time and this thing actually utilizes test time compute. So over time you get more and more ideas that are just continually better. And essentially it just starts with of course the scientists basically just saying look, this is where I'd want to research of course with constraints and other attributes. So this is basically how the entire architecture is. It shows the entire multi AI agent framework with of course tool use, search, additional tools. It also has memory and right here, like it says, it continuously generates reviews, debates and improves research hypotheses over time. So this is something where it doesn't just, you know, put out one idea internally, it's reviewing. I don't actually have the number of ideas that it's actually reviewing, but it's reviewing a number of different ideas, evaluating those, reflecting on those and ensuring it can then rank those and ensure it gets better over time, all whilst browsing the web and ensuring that it can validate everything through scientific database to validate its findings. And then this is where we have the further explanation on how the AI scientist passes the assigned goal into a research plan configuration managed by a supervisor agent. The supervisor agent assigns a specialized agent to the worker queue and allocates resources. And this design enables the system to flexibly, scale, compute and iteratively improve scientific reasoning toward the specified research goal. So you can see we got the supervisor agent here that actually assigns these agents to the workers. And then of course this entire system is able to generate the research overview with detailed hypotheses. Now one of the crazy things about this is that this research agent, scientist agent workflow is something actually utilizes test time compute. So you can see right here they actually talk about how the AI co scientist leverages test time compute scaling to iteratively reason evolve and improve outputs. And this is something that's really crazy because we know that the test time compute paradigm was something that only recently occurred. And of course now we're already seeing that it has applications in many different fields. You can see right here it says key reasoning steps include self play based scientific debate for novel hypotheses generation ranking, tournaments for hypotheses comparison and an evolution process for quality improve improvement. Says that the system's agentic nature facilitates recursive self critique, including tool use for feedback to refine hypotheses and proposals. So it seems that over time this is something that continually brings better ideas. And I could probably bet that over time with even more compute and even a more efficient algorithm, this system outputs better hypotheses you can also see that this system's self improvement relies on the ELO auto evaluation metric derived from its tournaments. And basically in this context, the tournaments refer to a competitive ranking process used by the AI co scientists to evaluate the hypotheses. The system basically generates multiple ideas, then systematically compares the ideas using the predefined criteria. And basically they just compare all of those ideas against each other and they look at the strengths, the weaknesses, the supporting evidence, and the best ideas are then ranked and then of course iteratively refined using a self improvement loop. And that of course ensures only the highest quality ideas are presented by the scientists. And you can see right here it says we analyze the concordance between ELO ratings and of course the GPQA benchmark on its diamond set of challenging questions. And we found that higher ELO rating positively correlate with a high probability of correct answers. Right here you can see the concordance of the ELO rating with the AI co scientist performance on the GPQA diamond set questions. You can see Gemini 2.0 is a bit stagnant, but the AI CO. So you can see right here that seven domain experts curated 15 open research goals and the best guest solutions in their field of expertise. Using the automated ELO metric, we observed that the AI co scientist outperformed other state of the art agentic and reasoning models for these complex problems. So essentially this chart right here basically just shows how the AI co scientist performs compared to the other AI models and human experts when coming up with new research ideas. And basically scientists gave 15 different problems to both AI and human experts and then they used an ELO rating system, which is basically just the way you rank up performance like in chess, to see who came up with the best solutions in. This graph basically shows that the AI co scientist in Blue Triangles got better over time, meaning that it improved the more it worked on the problem. Of course. Test time, compute. Now you can see right here where humans are. You can see where the human expert is. You can also see where Gemini 2.0 flash is. You can also see where Gemini 2 is. We can also see that the AI co scientist overall clearly surpasses them in all metrics. So this is really good because of course this means that AI can keep improving the longer it works on a problem and it can help scientists find better solutions faster. So if you actually want to talk about the real world, use cases, because of course there are fancy demos and people are always like, okay, this is cool, but how does that actually help us? Basically they did something on How AI is helping to find new uses for existing drugs to treat acute myeloid leukemia, which is a type of blood cancer. And normally creating new medicines takes a long time and costs a lot of money. But drug repurposing means scientists find new ways to use medicines that we already have. And Google's AI co scientist was used to predict which existing drugs might help fight this type of cancer. And basically, after they tested the AI suggestions in a lab, scientists found that one of the drugs called Cura 6 was able to, you know, as you can see right here, reduce cancer cell growth at certain doses. And this graph shows how the drug concentration increases and the number of cancer cells decreases, meaning it could be an effective treatment. And this is because it means that AI can basically speed up cancer research, helping doctors find better treatments faster and with fewer side effects. So this is one of the real world use cases. You know, one of the examples that they showed, you've also got another one here, which is basically where they did a study on liver fibrosis. So this is basically liver fibrosis is basically where you have the liver getting too much scarring, which makes it harder for it to work properly. And scientists of course need new ways to stop this. But figuring out what to target is pretty tricky. So Google's AI co scientists suggested some possible treatments and scientists tested them in lab grown mini livers called these organoids. And essentially this graph right here shows the fibrosis inducer. And this is of course what made scarring worse. The second bar, the fibrosis inhibitor, this helped reduce scarring. And of course the last four bars, which are the AI suggested drugs, these also helped reduce scarring, meaning that the AI might have actually found new treatments that actually work. So you can see right here, the first two ones actually did manage to make this go down a decent amount, which is of course really good. So this one is basically how AI helped rediscover a key mechanism in antimicrobial resistance, which is basically how bacteria evolved to resist antibiotics. So normally scientists would spend years on this. So scientists basically spend years doing experiments to understand how bacteria share resistance genes. And one key discovery was of cfpicis, which is basically small DNA segments that help bacteria to transfer resistance genes. And here's exactly what happened. So in the 2013-2025, humans slowly figured out how this actually worked through experiments. But what's crazy is that in 2024, researchers asked the Google scientists the same question. And in just two days, the AI independently figured out that the CFPICs interacted or interacts with the phages to spread the resistance. So the crazy thing about this is that AI just matched humans years of research in literally just two days independently, which is absolutely insane because it managed to basically figure out what took humans around seven years. And it managed to do that, in fact, not even seven years, in fact a lot longer than that. And it managed to do that in just two days. So AI was able to do this very independently and it basically sped up the discovery process from years to date. So you can imagine how crazy that is in terms of the development of drugs. Listen, when these individuals like Dario Amade was talking about the fact that, you know, scientific research is going to be sped up by tenfold, this is the kind of thing that they were talking about. Even though some people were stating that they were just being hyperbolic with their claims. But clearly this is one of the first times where you can actually see that literally the scientist was able to, you know, do this within 48 hours, which is absolutely insane. So one of the things that I did want to look at was of course, you know, was it just in the training data? But, you know, looking at the image and you know, the text carefully, there's a few points that do indicate that this was a controlled test to verify that the I wasn't just recalling the training data, basically meaning that the AI didn't just remember this and then pretend. And basically they talk about how there was a third validation test where the researchers specifically chose a topic that had not yet been revealed in the public domain. More specifically, they talk about how this was testing the AI co scientists with an identical research question about work that only existed in original novel laboratory experiments performed prior to the use of the AI co scientist systems. So this is something that is actually real, is actually true, and it's pretty crazy. So we could actually see know all of the kinds of research that we really do need speed up by tenfold. And I mean, think about it like this guys. This was just, you know, development of the AI co scientist version one. What happens when we have, you know, version 10 and we have a million of them running 24, seven guys running, you know, just so many different experiments. What happens then? Think about it guys. That's absolutely insane. So that being said, hopefully you guys did enjoy this video and I really did want to make a video on this because I think this is a monumental piece of information that most people just didn't realize. And let me know what you guys think about this because I do hope that all of the chronic illnesses that people do suffer from, all of the horrible, debilitating disease that exists in the world can just eventually disappear.

And I think there's really potential to conquer many of the afflictions we still face.

Right.

A lot of the easy ones, like diseases that were addressed by sanitation or vaccination or antibiotics, we solved. But things like cancer and Alzheimer's disease, much more complicated. And so I'm wondering if AI is really what we need to understand that complexity and to surmount those diseases, frankly, much faster than, you know, than I think most of us are imagining. We're getting used to a world where those diseases are very hard to address and progress is very slow. I don't think it needs to be that way. I think if we get it right, these incurable diseases could actually. We could actually overcome. And, you know, we'll. We'll look back at them, you know, the way we look back at bubonic plague or mumps.

Thank you guys for watching the video, and I'll see you in the next one.