Google’s new AI generates hypotheses for researchers
- Thread starter JournalBot
- Start date
That's a very oddly specific measurement. Wouldn't it be better to rate it vs someone formulating their own hypotheses? It's like saying "scientists rated toddlers higher than babies at research proposals" -- it doesn't really mean anything because it's comparing things that were never built to do such a thing.
Too bad you can't trust its output on contextual datasets, especially as the data sets get larger than the token limits.
Upvote
59
(61
/
-2)
I find it a bit amusing that they have a system that is trying to find the best experimental design to test a given hypothesis and the best they can do to test their own system is to compare it to older systems in a bar graph without any statistical tests to assess if the differences are significant.
Upvote
60
(60
/
0)
It's so funny how something can happen for a decade, or longer, and it's no big deal. But suddenly Google does it and their massive press team goes into action and it's reported everywhere.
The entire field of computational biology has been about eeking out slightly better hypotheses from millions or billions of options.
Other systems like this are being used in computational chemistry and material design to break new ground for years.
Props to whoever works in Google's PR department, you're killing it.
The entire field of computational biology has been about eeking out slightly better hypotheses from millions or billions of options.
Other systems like this are being used in computational chemistry and material design to break new ground for years.
Props to whoever works in Google's PR department, you're killing it.
Upvote
70
(72
/
-2)
Upvote
14
(14
/
0)
I would love to try this with any amorphous field (some fields in the social sciences) to see if it works. From my own experience doing research: a good hypotheses is creative first and foremost. I doubt replication machines can do that better.
Upvote
6
(6
/
0)
Upvote
10
(10
/
0)
lol what a fluff piece
In a time when PIs everywhere are scrambling worrying about indirect funding hit - tech boys gonna claim their LLM helps.
In a time when PIs everywhere are scrambling worrying about indirect funding hit - tech boys gonna claim their LLM helps.
Upvote
30
(34
/
-4)
Just don't ask it to research into Gouda cheese or how to keep cheese on the pizza... actually just don't ask it anything related to cheese
Upvote
21
(22
/
-1)
This is literally a lower bar for experimental quality than what we did for bio 201 lab in college.
That was the lowest level class you could take.
Upvote
15
(15
/
0)
A bit of a tangent, but one thing I've thought AI would actually do pretty well at is being fed a specific research topic, point it at the existing research material, and let it go through and check the material for integrity, flagging up any instance where an article has issues.
Since we already hold the existing research to a certain standard, using AI to review the material to enforce that standard seems to me to make a lot of sense, and it can crawl the citations for eternity, only resting to update to a new model.
Meanwhile, this technique could also be useful to find the gaps in existing research that might make good starting points for new research.
In both of these endeavors, it doesn't matter if the machine learning hallucinates, because it's still up to the user to determine if the output is a viable starting point for an investigation. And if doing this work would take a grad student 5+ years to explore a specific topic, but a bot can crawl the material and run 15-20 of these per day... that grad student can then spend 3 years reviewing bot output for potential winners instead of sifting all the published data.
Since we already hold the existing research to a certain standard, using AI to review the material to enforce that standard seems to me to make a lot of sense, and it can crawl the citations for eternity, only resting to update to a new model.
Meanwhile, this technique could also be useful to find the gaps in existing research that might make good starting points for new research.
In both of these endeavors, it doesn't matter if the machine learning hallucinates, because it's still up to the user to determine if the output is a viable starting point for an investigation. And if doing this work would take a grad student 5+ years to explore a specific topic, but a bot can crawl the material and run 15-20 of these per day... that grad student can then spend 3 years reviewing bot output for potential winners instead of sifting all the published data.
Upvote
12
(13
/
-1)
AI is cool i guess
Wise, Aged Ars Veteran
i like how we keep oscillating between "AI isn't really doing anything new" and "but hey it's actually getting real results in labs"
the framing of "this is just a chatbot" feels disconnected when we're seeing actual wet lab validation of its hypotheses. like, yes, it's "just" pattern matching, but so is most of early stage research? connecting dots others haven't connected is literally how science progresses
the framing of "this is just a chatbot" feels disconnected when we're seeing actual wet lab validation of its hypotheses. like, yes, it's "just" pattern matching, but so is most of early stage research? connecting dots others haven't connected is literally how science progresses
Upvote
-3
(12
/
-15)
this probably will save someone's life.
It's followed by
Suggesting a new avenue of treatment for cancer is not considered to be a "true insight" and it still "can't understand". A bit more optimism is warranted.
Upvote
-1
(10
/
-11)
At some point, an adult just needs to take away these AI people's GPU's so they stop bothering people doing useful work.
Upvote
1
(14
/
-13)
Ah, a stochastic system in search for an application.
Refreshing use of time and money.
Refreshing use of time and money.
Upvote
6
(11
/
-5)
I can't remember the last time I read a paper that didn't include a "additional study in [various areas] is warranted" statement in the conclusion. If a model was fed those papers, it has a handy list from which to start.
Upvote
5
(5
/
0)
- Add bookmark
- Featured
- #22
I think this is more a divergence between the actual reported news and the community cynicism.
I get skepticism, but some people here deny what’s in front of their faces. Cover their eyes and ears.
I’d like to think eventually truth is undeniable but the human capacity for self deception doesn’t seem to have any actual limits.
Upvote
4
(13
/
-9)
If there’s anything that might change people’s views, that might actually work. Or not.
People deny the effectiveness of vaccines and vote to disassemble their own healthcare.
Bias against AI might very well cause damage of its own. It shouldn’t matter where good ideas come from.
Upvote
-1
(4
/
-5)
People have invested a lot of money in AI, they need a return on their investment.
Upvote
3
(3
/
0)
I mean, some of us have been using LLMs for the last year or more to do exactly this? Helping with brainstorming is one the strong-suits of LLMs already.
As regards TFA: I'm not sure that having a mini-chatbot arena running over my research ideas is going to help any more than a single instance. LLMs are not good at having new ideas, but they are great at attacking the ones I have, and thereby helping me self-assess better. One inference model is completely fine for this; I don't need 20 dicking it out between them.
(Also, seriously: as opposed to improving the underlying NN architecture further, why are we now developing tools that just provide iterative loops of more of the same? These are really low-rent methods, that ironically cost huge amounts to run. It's not where I'd hoped we'd go with this ...)
Upvote
5
(6
/
-1)
Have you taken a look at the US president lately? There are very clearly zero limits here.
Upvote
6
(6
/
0)
If I had a dollar for every drug repurposing experiment that had initial promising lab results and went nowhere, I could probably fund my lab for a year.
We all wish this wasn't so, but drug repurposing isn't proving to be the cheap panacea many hoped it would. (It does sometimes work; but it's pretty rare.)
Upvote
13
(13
/
0)
Well. That’s the cult leader deceiving others. Most don’t actual buy what they are selling.
Self deception is a bit different. People don’t want to believe something so they don’t.
What the limits are for each? No clue. But my experience has been self deception is harder.
What’s harder to admit, that you were lied to or you lied to yourself? For many it’s the latter.
Upvote
0
(3
/
-3)
i would also not call "showing greater potential for impact and novelty" as a metric appropriate to make comparisons. it seems discretionary.
and yes, saying that spoon is better at digging than a fork is useless without information how good a fork is at digging to have a frame of reference.
Upvote
4
(4
/
0)
I think it's more like most of the low-hanging fruit has already been picked.
Hundreds of drugs have found off-label uses due to patient reports, academic theorizing, and clinical observations—everything from calcium channel blockers (blood pressure medications) for migraines, to particular antibiotics used as targeted immunosuppressants.
Finding just the stuff that no one realized could happen is a tall order.
Upvote
0
(0
/
0)
The pragmatist in me says:
1. Proof is in the pudding. Let scientists use it and see if they find it worthwhile.
2. Experts are a best-case scenario failure mode for AI. They are exactly the people who have the tools to check and evaluate the AI's work, and know when it is feeding them bs.
3. If you can bounce ideas off of it like a colleague, then it basically is one, for this purpose. Quacks like a duck and all that.
The problem with calling it a co-scientist, to me, is more the anthropomorphizing—that leads to problems. Also, it's delivering a marketing coup to Google, which I reflexively would like to reject.
1. Proof is in the pudding. Let scientists use it and see if they find it worthwhile.
2. Experts are a best-case scenario failure mode for AI. They are exactly the people who have the tools to check and evaluate the AI's work, and know when it is feeding them bs.
3. If you can bounce ideas off of it like a colleague, then it basically is one, for this purpose. Quacks like a duck and all that.
The problem with calling it a co-scientist, to me, is more the anthropomorphizing—that leads to problems. Also, it's delivering a marketing coup to Google, which I reflexively would like to reject.
Upvote
8
(8
/
0)
The truth is that not all things are usefully quantifiable.
Creativity and novelty, notoriously, being one of them—which is why subjective tests tend to reign supreme in that sphere.
I think the subjective assessment of scientists with expertise in the field is worthwhile. The question to me is more whether they asked the scientist to weigh how personally useful it was for them—and explained to them the costs associated with use.
Did they see it as worth having then? Would they push their institution for a subscription?
Upvote
2
(2
/
0)
It's Google man so everybody's gonna hate on top of the AI hate in general.
Upvote
0
(1
/
-1)
The use of "co-scientist" research is mentioned today on BBC.
"AI cracks superbug problem in two days that took scientists years
A complex problem that took microbiologists a decade to get to the bottom of has been solved in just two days by a new artificial intelligence (AI) tool.
Professor José R Penadés and his team at Imperial College London had spent years working out and proving why some superbugs are immune to antibiotics.
He gave "co-scientist" - a tool made by Google - a short prompt asking it about the core problem he had been investigating and it reached the same conclusion in 48 hours."
www.bbc.co.uk/news/article...
"AI cracks superbug problem in two days that took scientists years
A complex problem that took microbiologists a decade to get to the bottom of has been solved in just two days by a new artificial intelligence (AI) tool.
Professor José R Penadés and his team at Imperial College London had spent years working out and proving why some superbugs are immune to antibiotics.
He gave "co-scientist" - a tool made by Google - a short prompt asking it about the core problem he had been investigating and it reached the same conclusion in 48 hours."
www.bbc.co.uk/news/article...
Upvote
6
(6
/
0)
Just look at the fights we have now over vaccines, and hormones, and pharma, and Waymo, and so on. "Lives saved" is just wood for the fire.
Upvote
1
(1
/
0)
The examples given in the article about drug repurposing and liver fibrosis treatment are meaningless without any information on what was fed to the AI to produce those suggestions. Typically, if a lab is already doing a drug screen, repurposing should be something they are thinking of already. They wouldn't need AI to tell them that. And if you are searching for new molecular targets to treat in a specific disease, there are also existing tools to help identity important pathways based on your screening data or sequencing results.
One helpful thing that AI can probably achieve is a faster and cheaper way to analyze large data sets for biomedical scientists who are not trained to do computational biology. However, there are also existing tools, paid and unpaid, for that purpose without AI.
One helpful thing that AI can probably achieve is a faster and cheaper way to analyze large data sets for biomedical scientists who are not trained to do computational biology. However, there are also existing tools, paid and unpaid, for that purpose without AI.
Upvote
0
(0
/
0)
caution live frogs
Ars Centurion
Did the AI have access to his computer and unpublished work? No. Is it plausible that the guy, or his competitors or colleagues, never published anything over the past decade? No. Is it plausible that someone else in the field might have come up with hypotheses similar to his? Absolutely. So was the AI working from where he started a decade ago, or from where the state of this field of research is today?
If I told you a decade ago what computers today would be capable of, I'd be a visionary. If I told you today, I'd be restating what we already know. That's not as impressive as the article makes it sound.
And note that the AI didn't "solve" the issue. The AI simply suggested that the angle this guy was working might be right, but nothing has yet been proven. That's how hypotheses work.
Upvote
4
(4
/
0)
The AI does not understand any science, but what it does understand (if indeed we can claim AI actually understands anything) is how to detect patterns in massive piles of data and suggest how to communicate about those patterns according to their previous training. I can see how that might be useful.
Still, I'd take a microscope to the terms of service. Seems like there's also a real risk of Google stealing ideas and repackaging them into projects or patents profitable to Google before the original researcher can publish.
Still, I'd take a microscope to the terms of service. Seems like there's also a real risk of Google stealing ideas and repackaging them into projects or patents profitable to Google before the original researcher can publish.
Last edited:
Upvote
1
(1
/
0)
Additionally, this is confirmation bias at work. "This tool told me what I am working on is valuable, therefore this is a good tool."
Upvote
2
(2
/
0)
I actually read this article to read the comments...it's been interesting to follow the advocacy and outrage over these "AI" tools in comment sections. So much emotion. In the comments for this article the advocates have seemingly disappeared...perhaps they've given up with this ArsTechnica luddite crowd...irony! I guess vituperation and snark wins.
Machine learning pattern recognition is a very powerful tool. Yes the fevered hype of "AI" (the ChatGPT-type systems) is cringe-worthy and its revenue/profit model is questionable, but the underlying technology is undeniable. OpenAI etc are stuck on creating a one-ring-to-rule-them-all when it will be the little machine learning applications that will prove the technology's worth.
It's unfortunate that the hype and excess resources are focused on master-of-the-universe systems like ChatGPT -- this probably reflects some psychology in the Silicon Valley entrepreneur class. But there's a lot of baby in this bath water...don't refuse to use new tools because they are unfamiliar.
Machine learning pattern recognition is a very powerful tool. Yes the fevered hype of "AI" (the ChatGPT-type systems) is cringe-worthy and its revenue/profit model is questionable, but the underlying technology is undeniable. OpenAI etc are stuck on creating a one-ring-to-rule-them-all when it will be the little machine learning applications that will prove the technology's worth.
The BBC article cited by MentalVerse above is a direct answer to this -- the Google Co-scientist delivered a team's research direction (among others) in days instead of years. User "caution live frogs" wonders if it cheated and stole ideas from the researcher's PC -- Google says no but who knows? "caution live frogs" raises other questions that might detract from the accomplishment, all valid things to be aware of and check out, and "GenericAnimeBoy" raises the very real concern of who owns/uses the tool's products. But does that mean you will never use the tool? What is wrong with using a tool that might yield benefits to the researcher? Scientific progress often relies on serendipity -- is using this new tool worse than leaving the petri dishes on the window sill?
It's unfortunate that the hype and excess resources are focused on master-of-the-universe systems like ChatGPT -- this probably reflects some psychology in the Silicon Valley entrepreneur class. But there's a lot of baby in this bath water...don't refuse to use new tools because they are unfamiliar.
Upvote
-2
(0
/
-2)