Google's AI 'Co-Scientist' Solved a 10-Year Superbug Problem in Two Days (livescience.com) 46
Google collaborated with Imperial College London and its "Fleming Initiative" partnership with Imperial NHS, giving their scientists "access to a powerful new AI designed" built with Gemini 2.0 "to make research faster and more efficient," according to an announcement from the school. And the results were surprising...
"José Penadés and his colleagues at Imperial College London spent 10 years figuring out how some superbugs gain resistance to antibiotics," writes LiveScience. "But when the team gave Google's 'co-scientist'' — an AI tool designed to collaborate with researchers — this question in a short prompt, the AI's response produced the same answer as their then-unpublished findings in just two days." Astonished, Penadés emailed Google to check if they had access to his research. The company responded that it didn't. The researchers published their findings [about working with Google's AI] Feb. 19 on the preprint server bioRxiv...
"What our findings show is that AI has the potential to synthesise all the available evidence and direct us to the most important questions and experimental designs," co-author Tiago Dias da Costa, a lecturer in bacterial pathogenesis at Imperial College London, said in a statement. "If the system works as well as we hope it could, this could be game-changing; ruling out 'dead ends' and effectively enabling us to progress at an extraordinary pace...."
After two days, the AI returned suggestions, one being what they knew to be the correct answer. "This effectively meant that the algorithm was able to look at the available evidence, analyse the possibilities, ask questions, design experiments and propose the very same hypothesis that we arrived at through years of painstaking scientific research, but in a fraction of the time," Penadés, a professor of microbiology at Imperial College London, said in the statement. The researchers noted that using the AI from the start wouldn't have removed the need to conduct experiments but that it would have helped them come up with the hypothesis much sooner, thus saving them years of work.
Despite these promising findings and others, the use of AI in science remains controversial. A growing body of AI-assisted research, for example, has been shown to be irreproducible or even outright fraudulent.
Google has also published the first test results of its AI 'co-scientist' system, according to Imperial's announcement, which adds that academics from a handful of top-universities "asked a question to help them make progress in their field of biomedical research... Google's AI co-scientist system does not aim to completely automate the scientific process with AI. Instead, it is purpose-built for collaboration to help experts who can converse with the tool in simple natural language, and provide feedback in a variety of ways, including directly supplying their own hypotheses to be tested experimentally by the scientists."
Google describes their system as "intended to uncover new, original knowledge and to formulate demonstrably novel research hypotheses and proposals, building upon prior evidence and tailored to specific research objectives...
"We look forward to responsible exploration of the potential of the AI co-scientist as an assistive tool for scientists," Google adds, saying the project "illustrates how collaborative and human-centred AI systems might be able to augment human ingenuity and accelerate scientific discovery.
"José Penadés and his colleagues at Imperial College London spent 10 years figuring out how some superbugs gain resistance to antibiotics," writes LiveScience. "But when the team gave Google's 'co-scientist'' — an AI tool designed to collaborate with researchers — this question in a short prompt, the AI's response produced the same answer as their then-unpublished findings in just two days." Astonished, Penadés emailed Google to check if they had access to his research. The company responded that it didn't. The researchers published their findings [about working with Google's AI] Feb. 19 on the preprint server bioRxiv...
"What our findings show is that AI has the potential to synthesise all the available evidence and direct us to the most important questions and experimental designs," co-author Tiago Dias da Costa, a lecturer in bacterial pathogenesis at Imperial College London, said in a statement. "If the system works as well as we hope it could, this could be game-changing; ruling out 'dead ends' and effectively enabling us to progress at an extraordinary pace...."
After two days, the AI returned suggestions, one being what they knew to be the correct answer. "This effectively meant that the algorithm was able to look at the available evidence, analyse the possibilities, ask questions, design experiments and propose the very same hypothesis that we arrived at through years of painstaking scientific research, but in a fraction of the time," Penadés, a professor of microbiology at Imperial College London, said in the statement. The researchers noted that using the AI from the start wouldn't have removed the need to conduct experiments but that it would have helped them come up with the hypothesis much sooner, thus saving them years of work.
Despite these promising findings and others, the use of AI in science remains controversial. A growing body of AI-assisted research, for example, has been shown to be irreproducible or even outright fraudulent.
Google has also published the first test results of its AI 'co-scientist' system, according to Imperial's announcement, which adds that academics from a handful of top-universities "asked a question to help them make progress in their field of biomedical research... Google's AI co-scientist system does not aim to completely automate the scientific process with AI. Instead, it is purpose-built for collaboration to help experts who can converse with the tool in simple natural language, and provide feedback in a variety of ways, including directly supplying their own hypotheses to be tested experimentally by the scientists."
Google describes their system as "intended to uncover new, original knowledge and to formulate demonstrably novel research hypotheses and proposals, building upon prior evidence and tailored to specific research objectives...
"We look forward to responsible exploration of the potential of the AI co-scientist as an assistive tool for scientists," Google adds, saying the project "illustrates how collaborative and human-centred AI systems might be able to augment human ingenuity and accelerate scientific discovery.
yet another AI lie (Score:5, Informative)
https://www.newscientist.com/a... [newscientist.com]
> “We were shocked,” says Penadés. “I sent an email to Google saying, you have access to my computer. Is that right? Because otherwise I can’t believe what I’m reading here.”
> However, the team did publish a paper in 2023 – which was fed to the system – about how this family of mobile genetic elements “steals bacteriophage tails to spread in nature”. At the time, the researchers thought the elements were limited to acquiring tails from phages infecting the same cell. Only later did they discover the elements can pick up tails floating around outside cells, too.
Re:yet another AI lie (Score:4, Insightful)
Talk about nonsense. The AI looked at the words you asked and found the most probabilistic next word, and then the next most likely word, rinse and repeat.
Designing experiments and analyzing results, absolutely not.
Re: (Score:2)
Designing experiments and analyzing results, absolutely not.
Indeed. LLMs are a) mostly useless and b) make people a ton of money because others misunderstand what they can do and project hopes and desires onto them.
The result: Constant lying to keep the hype going a bit longer.
Re: (Score:2)
Re: (Score:2)
Exactly.
Re: (Score:2)
As a retired software engineer with 40 years experience in everything from 6502 assembly language to Ada to C to Java ... AI increases my new java minecraft modding productivity by a minimum of 300%. It's not so good at maintenance yet. It's also much more effective at providing solutions to problems and issues than reddit (which is now locked away anyway) and the pile of shit that was substack.
It's really useful at exactly the kind of work it did in this case.
And it's already "smart" enough to replace
Re: (Score:2)
AI gets hyped by lies, because the actual reality is not that impressive...
Re: (Score:3)
You know what else can do that? Interns. I am not kidding. Interns don't know that things are "impossible", and they will find solutions that other established workers have ignored or don't want to try.
Google using all our data without our consent (Score:3, Insightful)
That unpublished research was probably stored on a Google Drive and thus Gemini was just replying what they expected to hear from it.
Re: Google using all our data without our consent (Score:2)
What if the AI's context is wider and so it picks up semantic relationships that humans forget about?
2 Days? (Score:5, Insightful)
Re:2 Days? (Score:4, Insightful)
When does it know it's done? And if it takes 10 years to validate the results, how can we ever rely on it? This is crap.
Re: 2 Days? (Score:4, Funny)
Re: (Score:2)
Indeed. This is crap. But one of the few things LLMs actually can do well is "better crap".
Re: (Score:2)
Re: (Score:2)
I also wondered if there was some "confirmation bias" here. The scientists kept asking and asking, and eventually the AI gave them the answer they were expecting, so they congratulate the AI on doing 10 years of research in 2 days. Maybe?
We can't really conclude much from this rather puffy piece, other than "AI is looking at some hard problems", and I suspect it's capable of doing some good in those areas. However, what it absolutely not doing is "going further" than the data it was trained with.
Re: (Score:2)
I also wondered if there was some "confirmation bias" here. The scientists kept asking and asking, and eventually the AI gave them the answer they were expecting, so they congratulate the AI on doing 10 years of research in 2 days. Maybe?
Hey, it worked for Akinator...
How great a leap was it (Score:5, Insightful)
I cannot judge since we only have a puff piece, not actual data, and tldr;
However by definition the LLM does not operate in a vacuum, and if they had actually developed something that understands concepts and scientific method with real reasoning ability I think it would be a much huger announcement and probably be classified as a strategic weapon akin to a nuclear arsenal at this point, plus needed for stock value.
Instead we need to know what data had actually been fed into the system and what intellectual leap was needed to get to the final result. Only then would you know if you just need naive language manipulation to get there or is it a "PhD level scientist" as they (not sure which company) love to say. In this case someone mentioned their 2023 paper was in fact uploaded and it seems like naive language manipulation and ignorance of the reason why only intracellular tail pickup had been considered, was sufficient to generate the findings. And as someone else posted, does this mean 2 solid days of compute and what computing resources were used. Maybe $100,000 of compute would be worth it in some cases, but doubtful that this was one of them. Juggling concepts into a list of things to think about might be sufficient to trigger brainstorming among organic scientists too without requiring days of compute. That said I'd expect a co-scientist to appear at some time but I doubt they are there yet, especially when their scores are not amazingly different from other naive LLMs and the financial incentive to make hyperbolic press releases is so high. The scientists were honestly surprised but probably the people at Google were not.
Re:How great a leap was it (Score:5, Insightful)
Excellent post, and here is some more of their BS to support what you said.
Look at this link :
https://research.google/blog/a... [research.google]
"Accelerating scientific breakthroughs with an AI co-scientist"
As you said, a puff piece tooting their own AI-is-great horn. It makes vague claims, pseudo-illustrated with grade school graphics, then promises more. The microbial resistance piece is in there, but the one that caught my eye as "Advancing target discovery for liver fibrosis", as fibrosis is a big subject in my personal domain of professional activities.
Maybe they did good work but are just illiterate and cannot write clear sentences. Maybe they are so stupid from "everything AI" disease that they used ChatGPT or the like to write this incomprehensible silliness. Dissecting two of their paragraphs:
We probed the AI co-scientist system's ability to propose, rank, and generate hypotheses and experimental protocols for target discovery hypotheses, focusing on liver fibrosis.
Fair enough - sounds good.
The AI co-scientist demonstrated its potential by identifying epigenetic targets grounded in preclinical evidence with significant anti-fibrotic activity in human hepatic organoids (3D, multicellular tissue cultures derived from human cells and designed to mimic the structure and function of the human liver).
You mean - you reviewed the literature and found evidence of chemical targets that block fibrosis in ex vivo hepatocyte cultures? Why can't you just say so. But - caught ya'. Liver organoids do not usually have the various cells that make fibrosis. If you imply that the culture used stem cells and that you forced them to re-differentiate into non-liver non-ectodermal stromal-mesodermal (fibroblast) cells, well, that's not quite how it works in living livers, but I can be convinced, but don't ask me to accept that at face value. But, I am sure you will add some references.
These findings will be detailed in an upcoming report led by collaborators at Stanford University.
Shucks, I guess you aren't going to give any further useful information.
Then, they show a graphic, six items that rise above or fall below baseline of influencing fibroblast activity. It shows that an inducer induces, an inhibitor inhibits, and then there are four "Suggestions". Are these real chemicals culled from the literature, or AlphaFold style predictions of compounds that might work, or just the generic idea that if they could or did find something, they could paste it into the graph? The graph has this caption :
Comparison of treatments derived from AI co-scientist–suggested liver fibrosis targets versus a fibrosis inducer (negative control) and an inhibitor (positive control).
Yep, that's the way biology works. I see you (the AI) trained on freshman textbooks, but these scientists running these "experiments" already read them, so what's the point?
All treatments suggested by AI co-scientist show promising activity (p-values for all suggested drugs are less than 0.01), including candidates that possibly reverse a disease phenotype.
What p-values? You didn't run an experiment. Are these the p-values from studies you found in your training set, or are you hallucinating p-values because your AI "analysis" shows that such graphs are accompanied by p-values, or are you just making crazy predictions based on who knows what?
Results are detailed in an upcoming report from our Stanford University collaborators.
Oh - there you go. Forgive my impertinence questioning you methods and reasoning. You did finally direct me to the source study. When is that coming out? Where?
Here's a prompt for AI Co-Scientist :
Analyze the following. Google 'scientist" tries to apply AI to study flatus. What is the predicted concentration of methane, hydrogen sulfide, and methyl mercaptan in the air in his head?
Re: (Score:2)
You mean - you reviewed the literature and found evidence of chemical targets that block fibrosis in ex vivo hepatocyte cultures? Why can't you just say so. But - caught ya'. Liver organoids do not usually have the various cells that make fibrosis. If you imply that the culture used stem cells and that you forced them to re-differentiate into non-liver non-ectodermal stromal-mesodermal (fibroblast) cells, well, that's not quite how it works in living livers, but I can be convinced, but don't ask me to accept that at face value. But, I am sure you will add some references.
Ouch. Thanks for the analysis, I wasn't able to look that deeply; I was only able to recognize that it was quite vague.
Dupe (Score:5, Informative)
Uh isn't this a dupe from weeks ago, and that the claims are wildly overstated?
Re:Dupe (Score:4, Informative)
Re: Dupe (Score:4, Funny)
Re: (Score:2)
I have barely been spending time on slashdot the past few years, yet i also immediately recognised this as dupe, hence my vote is for the former.
Re: (Score:2)
Clearly worse that you wasting so much time on Slashdot and your nonconstructive complaint makes that use of your time even worse.
There's a fundamental problem because many important stories don't conform to the one-day cycle time of Slashdot. I think the fairly obvious solution would be to allow stories to move down the front page at different speeds, basically assessed by sustained interest against competition with newer stories, but... Slashdot has no resources to fix anything.
Me? I visit once a day, whi
Wait. Wait. Wait. (Score:3)
Reeal story here is the study was so derivative and generic that an LLM managed to produce something that was almost indistinguishable.... That's a poor showing all round and someone needs a performance review asap.
Re: (Score:2)
That is a pretty good angle. If an LLM can do it, it cannot have been hard and no great insights or advances were involved.
I am waiting for Science intentionally done so badly that an LLM can do better, just to "prove" LLMs "can do Science"...
Re: Wait. Wait. Wait. (Score:2)
Are you like those who walked away from Galileo's canonball drop experiment still convinced the heavy ball hit the ground first?
Survivorship bias (Score:3, Insightful)
I dunno.... meh (Score:2)
Let's say I play a game of 20 questions and it takes me 16 questions to win. I take the results up to question 15 and give them to an AI, and also feed it all the backstory on the general subject.
Should I be impressed if it makes a good guess for its own question 16? My direction was probably derivable from my question set. And the fact that I picked one logical refinement and the AI chose one slightly different doesn't seem like much of stretch.
Let's be real (Score:3)
I welcome any (real) progress in AI helping science, but I wish the incident proving AI supremacy happened under less suspicious circumstances. This story looks like a sensationalist clickbait, and the gaping holes in it are obvious to spot.
Re: (Score:1)
Can we stop using the term "AI" for everything? (Score:1)
It seems like anything related to "big data" is somehow "artificial intelligence". What most things that I'm seeing out of "AI" is really just a matter of computers having the ability to search through more data at a time to find that gem in the rough or needle in a haystack. Don't get me wrong, there's value in improving our ability to process large sets of data, but it is not necessarily "artificial intelligence".
How is "artificial intelligence" defined? How is "AI" distinct from what we were doing wit
Re: Can we stop using the term "AI" for everything (Score:2)
Isn't the natural language interface the real AI?
Re: (Score:2)
Don't get me wrong, there's value in improving our ability to process large sets of data, but it is not necessarily "artificial intelligence".
I completely agree. And so would most respectable AI researchers. Just filtering stuff syntactically and statistically is not AI. It can be quite useful nonetheless.
The term "artificial intelligence" is being so overused that it appears to have lost meaning.
Indeed. There are several factors at work. One is that quite a few AI researchers are not respectable, but instead work on hopes and dreams and misdirection and lies. It is not accident that quite a few proper AI researchers pretend to be in other fields by calling it "cognitive systems", "planning algorithms", "computer vision", etc., just so t
Re: Can we stop using the term "AI" for everything (Score:2)
Remember when you could not get a computer to understand context-sensitive grammars at all?
Too many confounding factors. (Score:2)
I don't know anything about the subject matter, but when I read the headline "superbug" I thought this was a coding problem with a bug that had been unresolved for 10 years.
Anyone who has worked on a tricky bug will have known that it can take a very, very long time to really understand what is going on but often, once you have that understanding, the solution can be simple, sometimes just a one line change (it can also sometimes be hard if the bug turns out to be a "feature" of the framework)
I would think
Repost from 2/21/25 (Score:2)
Re: (Score:2)
Guilty until proven innocent (Score:2)
This could simply be just another hallucination and/or conjecture. Until the result is verified by science, any AI-generates theory is no better than what I say in my sleep.
Still a needle in a haystack? (Score:2)
Re: (Score:2)
Exactly this. That the article omitted any estimation of the number of suggestions is telling.
Imagine the possibilities! (Score:2)
Imagine if an AI was trained with the USPTO patent database, and was used to create new inventions likely to be useful in the future, and patent the most valuable ones.
Because the patent system is now a "first to file" system, even if you are working on a new problem and find a novel solution, it is likely that your company could owe patent royalties to the AI company (Google), even though the AI company (Google) may not be working in your industry at all. From drugs to software patents to microprocesso