Post New

John Hopfield and Geoff Hinton won the 2024 Nobel Prize in physics for their AI related work.
https://www.nytimes.com/2024/10/08/science/nobel-p...

David Baker, Demis Hassabis and John Jumper won the 2024 Nobel prize in chemistry for their work on using AI to predict protein folding.
https://www.nytimes.com/2024/10/09/science/nobel-p...

I've been beating up on AI a bit in a couple of posts.
Not because I dislike it, but because I think that LLM's in particular can be better, and that in their current state it's dangerous to be pushing them to the public.
As a scientist, I feel that truly valuable contributions of AI are not generating pictures of kitties in spacesuits from a text prompt. But I guess it depends on your definition of "valuable", there's probably lots of money to be made from the kitties in spacesuits kind of stuff.
OTOH, solving the protein folding problem, and similar problems with the help of AI, will lead to many advances in medicine and biotechnology (and other fields).
This will create huge financial value.

I was at a conference today and one of the segments was on AI (shocking). There were a few speakers, including a guy from Microsoft and a start up providing AI products to big pharmaceutical companies and another from an educational AI business. Bit of a dot com vibe. Everyone is an AI user and expert now.

It reconfirmed my previous belief that AI is an important productivity tool that will obviously replace a lot of tedious jobs but that it is not real intelligent. It might give the right answer 90% of the time if you ask an easy enough question, or set it a basic high volume task. But it needs human intelligence to direct it and train it. Therefore, if the data source is from the internet, the results will be questionable. However, in a controlled organisation environment using quality data and with intelligent direction from an experienced and skilled human, it can be incredibly powerful.

I can see how it can crunch through large amounts of data and accelerate advancement in all kinds of fields. But it’s just a tool to assist intelligent humans. It’s going to be hard to trust AI with important decisions. I don’t see any evidence that AI can help with wisdom and judgement.

I really wanted to ask these AI experts about national security and AI. If AI enables making all kinds of things (including advanced weapons) exponentially easier, then rouge states will become more dangerous sooner. Hopefully Western military can maintain their military superiority. Maybe I’ll ask AI how we’re got.

I struggle to understand how Hopfield and Hinton made significant (any?) contributions to Physics. Did their work contribute to our understanding of matter and/or energy? Was their work published in Physics journals? If someone here can enlighten me I'd appreciate it.

https://www.nobelprize.org/prizes/physics/

I wouldn't think using techniques from physics would be sufficient to receive the award. If anything, I'd think they'd be better candidates for a Fields Medal, as their work is more mathematical in nature (algorithms).

John

@EVBigMacMeal
I agree.

Argh!!! Ahhhh!!! Expletive deleted !!!
As I was writing this post, I decided to upload one of my own published papers to NotebookLM (https://notebooklm.google/). This AI not only summarizes the document that you provide it, it also creates an audio podcast of two apparently quite intelligent people discussing the document. They do a "deep dive" as they call it, and you as listener get to learn along with them.

It's a great 14 minute podcast! The voices are very realistic and the interplay between the two participants is extremely engaging! It's something you'd want to listen to, heck I wanted to keep listening to it and I wrote the paper! You're learning so much, and it's fun!

But it's 14 minutes of BS.
I ought to know, I wrote the damn paper.

Admittedly, the paper was on a sort of buzzy topic so I can sort of see how an AI might go off on a riff. But the paper was narrowly technical, in principle it might have a wide application, but not after much more work, and not the applications the AI podcast people came up with. The podcast was like overhearing two bad graduate students sharing their knowledge of a technical subject, feeding off each other's "insights" with enthusiaasm, but getting it all wrong, and about your work. "NO, NO, NO!!! For the love of all that is holy, please stop! Just stop it!"

It was oddly very offensive.
Perhaps because it wasn't two humans, who in principle could be corrected, but it was a faceless machine mindlessly spewing a very convincing but very polluted version of my work.

This level of misinformation will become very ubiquitous, hence very dangerous.
And that's not including the deliberate misinformation that's created by state and non-state actors.
Apart from immediate impact, it will also create a vicious feedback loop for future training of LLMs.

I really dislike huge corporations spending huge resources (Microsoft wants the Three Mile Island nuclear reactor for gods sake) to push faulty LLM's to the public who just aren't ready for it. There's lot of dollars in the illicit drug too, but it's harmful.

Here's a conjecture, FWIW:

Chemistry Nobel:
The case for a chemistry Nobel for Baker (a biochemist), Hassabis and Jumper (both AI at DeepMind) was pretty clear. A long standing and important problem in computational biology is predicting how proteins fold given their amino acid sequence, and this was solved using AI as a tool and physical intuition. Baker had long been active and eminent in structure prediction, and his group went beyond, to design new sequences for new proteins to accomplish new tasks. The impact of the totality of this work, heavily involving AI, will probably be (literally) life-changing.

Physics Nobel:
Hopfield is a physicist, and while Hinton has some background in physics he is more of a polymath (his career associations have been as a computer scientist).
Hopfield's original neural net work relates to Ising models and spin glasses, stuff physicists understand. This lured a whole generation of physicists into neural networks, including me (I then got into computational biology, so I'm not expert on the technical details of the new "AI"s). Geoff Hinton doesn't have a classical physics background, e.g. a PhD in physics, but he thinks in a very physical way. For example, one of his earlier and very interesting work was on something called the Boltzman Machine, which is a stochastic neural net with a temperature. He later invented "deep learning" and other ways to make neural nets able to solve real world problems.

Conceivably the Nobel committee, having decided that the Chemistry prize would be awarded for solving protein folding and also to design new proteins using AI as tools, then felt the need to honor key people who made AI available. Some may consider the physics Nobel for Hopfield and Hinton a stretch, but there's a case to be made for it. Also the old divisions of science into physics, chemistry, etc are getting broken, both with cross-disciplinary work, and from basically creating entirely new fields.

BWDIK - I'm an old retired dog on the internet.

BWDIK - I'm an old retired dog on the internet

Perfect.

Permission to steal this for use as a signature line?

--sutton
just an ORDOIT these days

As a scientist, I feel that truly valuable contributions of AI are not generating pictures of kitties in spacesuits from a text prompt. But I guess it depends on your definition of "valuable", there's probably lots of money to be made from the kitties in spacesuits kind of stuff.
OTOH, solving the protein folding problem, and similar problems with the help of AI, will lead to many advances in medicine and biotechnology (and other fields).
This will create huge financial value.

Don't get distracted by the kittens - there is massive investment being made in productivity tools based on AI tech - with incredible ROI attached to them.


https://www.cnbc.com/2024/10/10/microsoft-announce...



tecmo
...

@tecmo
I agree that there is great potential for AI in areas other than hard science.

It allows doctors to consensually record their visits with patients, and AI automatically transforms them into clinical notes and summaries.
Health-care organizations will test and validate them before the company rolls them out more broadly.

I sure hope they validate the hell out of it before they go summarizing my medical records!
The summary of one of my scientific papers didn't go so well.

Regulation:
A possible approach is that Microsoft (and other providers of AI) formalize a validation protocol, the health care organizations and their providers then follow the defined protocol, Microsoft reviews and signs off on the validation, and only then is the AI application allowed by Microsoft for enterprise-wide use.
Similar to how doctors sign their names to clinical notes, those medical personnel involved in validation would need to sign their name that the AI summaries agreed with their notes.
Everyone needs to be on the hook: Microsoft, the health care organization, and the individual providers involved in validation.
Validation might need to be done annually, because Microsoft would undoubtably be rolling out new, updated, AI versions that'd need validation.

LLMs are currently error-prone.
Stupidity by isolated people can be fixed (if necessary, isolated people can be fired).
But systemized stupidity is insidious, if allowed to take root it will be almost impossible to fix.
Everyone needs to be on the hook in a validation process for AI's to actually achieve their potential in enhancing productivity (IMO).

Don't get distracted by the kittens -

That too would make a great signature block.

Jim

from the internet:

lasers were once a huge scientific breakthrough; now we use them to play with cats.

computers were once a huge scientific breakthrough; now we use them to look at cats.

conclusion: science was made for cats.

best,

mike

In term of significance, AI is like the invention of Integrated Circuit, it started with tens or hundreds of improvement in scale, and now it's in the scale of trillions. So it is certainly justified for a Nobel prize. Naming it as physics is indeed odd, probably it's the closest to math. Hinton is the number one pioneer in AI. However, almost every method or algorithm existed before his work. And I'm not sure how relevant is Boltzmann machine to today's AI.

It was oddly very offensive.
Perhaps because it wasn't two humans, who in principle could be corrected, but it was a faceless machine mindlessly spewing a very convincing but very polluted version of my work.

You should try putting your paper into a chatbot and then asking it to summarize it and then challenging it on where you think it gets things wrong or misinterprets something or misses something. It will keep modifying its summary.

AI is not really that different than actual intelligence in that regard.

R:

Some may consider the physics Nobel for Hopfield and Hinton a stretch, but there's a case to be made for it.

I sort of have to wonder what would be the value of a Nobel Prize if it was given for things which are NOT a stretch?

Post New

US Policy FAQ

---

Contact Shrewd'm
Contact the developer of these message boards.