3 ms·

Scientists begin building AI for scientific discovery using tech behind ChatGPT

40 points by gardenfelder 3y ago

gardenfelder 3y ago
"An international team of scientists, including from the University of Cambridge, have launched a new research collaboration that will leverage the same technology behind ChatGPT to build an AI-powered tool for scientific discovery."
- varelse 3y ago
  [dead]
- mistrial9 3y ago
  the BLOOM team is drinking their wine while you toil, Cambridge? https://en.wikipedia.org/wiki/BLOOM_(language_model) https://en.wikipedia.org/wiki/BLOOM_(language_model)
  - pk-protect-ai 3y ago
    They can't probably join BLOOM collaboration, while BLOOM has already released a model and probably will not receive anymore funding. BLOOM model is the only model that spooked me by answering "I do not know. I do not know. I do not know..." on some of my questions.
gardenfelder 3y ago
Github https://github.com/PolymathicAI https://github.com/PolymathicAI
fiforpg 3y ago
I am trying to think of a reason for why there are many articles announcing the start of a grand initiative, but few that would announce one coming to fruition. Can't quite put my finger on it.
- luxuryballs 3y ago
  They need content, but what do you mean they wouldn’t announce one coming to fruition? I bet they would if it was a slow news day, or just depending on what cooler things are available.
- sureglymop 3y ago
  This is business 101. Sell the idea before you have the product. Then try to create the product with the funds raised (or not).
- teaearlgraycold 3y ago
  As your sass implies, it could be this may all be hype. But also if it wasn't it would take a while for exciting results to arrive.
malux85 3y ago
I’m doing this for computational chemistry https://atomictessellator.com https://atomictessellator.com
xyproto 3y ago
Are LLMs capable of any form of innovation or discovery that is not a simple mix of existing concepts? (It could be argued that all innovation or discovery is a simple mix of existing concepts, but I don't believe this to be true).
- unsupp0rted 3y ago
  Whether they are or they aren't, we can get very far with discovery that is a simple mix of existing concepts
  - xyproto 3y ago
    Absolutely! And good science is sometimes just about organizing existing information too, which I believe LLMs can excel at.
- tracerbulletx 3y ago
  The word "simple" is really loaded here. Is the universe a simple mix of existing physics? Is a 70B parameter matrix of numbers containing the relative relationships between a significant chunk of published human knowledge actually that simple?
  - xyproto 3y ago
    Sure, but I have yet to see LLMs actually innovate and not just pretend to be innovative.
romafirst3 3y ago
I’m fairly skeptical about AI hype, but there are probably plenty of applications of the tech for use in scientific discovery. Just not the “Albert Einstein bot” or whatever that gives you fully formed scientific theories. More like tools to search large spaces for theories/theorems that seem likely that can then be independently verified. As I understand it we’ve been using AI in space exploration/astronomy for a long time to search data generated from telescopes I can imagine there are analogous uses for chatGPT style solutions.

40 points by gardenfelder 3y ago

"An international team of scientists, including from the University of Cambridge, have launched a new research collaboration that will leverage the same technology behind ChatGPT to build an AI-powered tool for scientific discovery."

varelse 3y ago
[dead]
mistrial9 3y ago
the BLOOM team is drinking their wine while you toil, Cambridge? https://en.wikipedia.org/wiki/BLOOM_(language_model) https://en.wikipedia.org/wiki/BLOOM_(language_model)
- pk-protect-ai 3y ago
  They can't probably join BLOOM collaboration, while BLOOM has already released a model and probably will not receive anymore funding. BLOOM model is the only model that spooked me by answering "I do not know. I do not know. I do not know..." on some of my questions.

gardenfelder 3y ago

Github https://github.com/PolymathicAI https://github.com/PolymathicAI

fiforpg 3y ago

I am trying to think of a reason for why there are many articles announcing the start of a grand initiative, but few that would announce one coming to fruition. Can't quite put my finger on it.

luxuryballs 3y ago
They need content, but what do you mean they wouldn’t announce one coming to fruition? I bet they would if it was a slow news day, or just depending on what cooler things are available.
sureglymop 3y ago
This is business 101. Sell the idea before you have the product. Then try to create the product with the funds raised (or not).
teaearlgraycold 3y ago
As your sass implies, it could be this may all be hype. But also if it wasn't it would take a while for exciting results to arrive.

malux85 3y ago

I’m doing this for computational chemistry https://atomictessellator.com https://atomictessellator.com

xyproto 3y ago

Are LLMs capable of any form of innovation or discovery that is not a simple mix of existing concepts? (It could be argued that all innovation or discovery is a simple mix of existing concepts, but I don't believe this to be true).

unsupp0rted 3y ago
Whether they are or they aren't, we can get very far with discovery that is a simple mix of existing concepts
- xyproto 3y ago
  Absolutely! And good science is sometimes just about organizing existing information too, which I believe LLMs can excel at.
tracerbulletx 3y ago
The word "simple" is really loaded here. Is the universe a simple mix of existing physics? Is a 70B parameter matrix of numbers containing the relative relationships between a significant chunk of published human knowledge actually that simple?
- xyproto 3y ago
  Sure, but I have yet to see LLMs actually innovate and not just pretend to be innovative.

romafirst3 3y ago

I’m fairly skeptical about AI hype, but there are probably plenty of applications of the tech for use in scientific discovery. Just not the “Albert Einstein bot” or whatever that gives you fully formed scientific theories. More like tools to search large spaces for theories/theorems that seem likely that can then be independently verified. As I understand it we’ve been using AI in space exploration/astronomy for a long time to search data generated from telescopes I can imagine there are analogous uses for chatGPT style solutions.