WEBVTT - Will AI solve physics?

0:00:07.800 --> 0:00:10.720
<v Speaker 1>There is so much AI hype out there from the

0:00:10.720 --> 0:00:15.080
<v Speaker 1>doomers and the boomers. People say AI is the future.

0:00:15.560 --> 0:00:17.759
<v Speaker 1>Other people say no, no, it's the beginning of the

0:00:17.880 --> 0:00:20.880
<v Speaker 1>end of humanity. But no, it's gonna make us all rich,

0:00:21.160 --> 0:00:24.079
<v Speaker 1>or it's gonna enslave us, or it's gonna cure cancer,

0:00:24.600 --> 0:00:28.280
<v Speaker 1>or it's going to ruin the environment. But everyone agrees

0:00:28.360 --> 0:00:32.360
<v Speaker 1>that we're at some kind of really fascinating inflection point

0:00:32.680 --> 0:00:35.360
<v Speaker 1>where the future is very hard to predict because the

0:00:35.440 --> 0:00:39.760
<v Speaker 1>present is changing so quickly. So what does that mean

0:00:39.800 --> 0:00:43.680
<v Speaker 1>for us, fellow extraordinaries. Let's zoom in on a more

0:00:43.720 --> 0:00:48.920
<v Speaker 1>specific claim of the AI hypesters. Will AI solve physics?

0:00:49.720 --> 0:00:52.040
<v Speaker 1>What does it mean to solve physics? What is AI

0:00:52.240 --> 0:00:55.440
<v Speaker 1>capable of doing? And is that a good match for

0:00:55.560 --> 0:00:59.680
<v Speaker 1>cracking the mysteries of the universe? Will future humans live

0:00:59.800 --> 0:01:03.320
<v Speaker 1>in universe that is totally understood to them. Will the

0:01:03.360 --> 0:01:07.040
<v Speaker 1>mysteries of the universe fade into the past our descendants

0:01:07.080 --> 0:01:10.200
<v Speaker 1>live in without the sense of wonder about how it

0:01:10.240 --> 0:01:14.039
<v Speaker 1>all works? Or will humans always be curious no matter

0:01:14.080 --> 0:01:18.240
<v Speaker 1>what our AI scientists reveal about the nature of reality.

0:01:18.920 --> 0:01:23.280
<v Speaker 1>Buckle up, because while your future may bring surprises. Daniel's

0:01:23.280 --> 0:01:26.080
<v Speaker 1>prediction for the future of AI and physics may also

0:01:26.480 --> 0:01:31.839
<v Speaker 1>raise your eyebrows. Welcome to Daniel and Kelly's Extraordinary Universe.

0:01:45.040 --> 0:01:48.640
<v Speaker 2>Hello, Kelly Widersmith. I study parasites and space, and I

0:01:48.760 --> 0:01:52.080
<v Speaker 2>have used AI to try to help me tell jokes

0:01:52.440 --> 0:01:54.840
<v Speaker 2>at a conference and they fell pretty flat.

0:01:57.400 --> 0:01:59.240
<v Speaker 1>Was the audience filled with people or AI?

0:01:59.720 --> 0:02:01.480
<v Speaker 2>Oh? No, they were filled with people. And I did

0:02:01.520 --> 0:02:03.960
<v Speaker 2>warn them that they were parasite related jokes that had

0:02:04.000 --> 0:02:06.680
<v Speaker 2>been generated by AI, and we all agreed that I

0:02:06.680 --> 0:02:08.720
<v Speaker 2>could stay president for a few more years because I

0:02:08.720 --> 0:02:09.880
<v Speaker 2>was still funnier than the AI.

0:02:12.520 --> 0:02:12.680
<v Speaker 3>Hi.

0:02:12.760 --> 0:02:16.920
<v Speaker 1>I'm Daniel. I study particles and aliens, and I've never

0:02:17.000 --> 0:02:20.359
<v Speaker 1>relied on AI for any jokes. I think AI would

0:02:20.400 --> 0:02:23.400
<v Speaker 1>be embarrassed to take credit for any of my dad jokes.

0:02:24.840 --> 0:02:27.280
<v Speaker 2>I think the joke was supposed to be that AI's

0:02:27.720 --> 0:02:31.400
<v Speaker 2>jokes were so bad that I am any joke. Yeah,

0:02:31.440 --> 0:02:34.000
<v Speaker 2>And in this case, they were so bad that, like

0:02:34.040 --> 0:02:36.560
<v Speaker 2>the tapeworm related jokes just didn't make any sense at all,

0:02:36.600 --> 0:02:39.320
<v Speaker 2>And so we all had a that's what they think

0:02:39.440 --> 0:02:43.639
<v Speaker 2>tapeworms do, very inside jokes sort of fig.

0:02:43.760 --> 0:02:47.000
<v Speaker 1>But maybe that's a fascinating threshold for AI, not the

0:02:47.040 --> 0:02:49.240
<v Speaker 1>Turing test, like can you sound like a human, but

0:02:49.520 --> 0:02:51.120
<v Speaker 1>can you actually be funny?

0:02:51.560 --> 0:02:54.040
<v Speaker 2>Yeah? Yeah, I mean that was that was sort of

0:02:54.080 --> 0:02:56.560
<v Speaker 2>My point was like, is you know, could this replace

0:02:56.560 --> 0:03:01.800
<v Speaker 2>a president's speech or opening remarks? We decided no.

0:03:02.120 --> 0:03:04.320
<v Speaker 1>Not yet, not yet, work on that nerds.

0:03:04.560 --> 0:03:07.000
<v Speaker 2>That's right, that's right. And so my question, I have

0:03:07.040 --> 0:03:09.640
<v Speaker 2>two questions for you this morning, since we're talking about AI,

0:03:10.160 --> 0:03:13.920
<v Speaker 2>And my first question is when in your life is

0:03:13.960 --> 0:03:18.120
<v Speaker 2>AI most useful? And then what scares you about AI?

0:03:18.240 --> 0:03:18.840
<v Speaker 2>If anything?

0:03:19.360 --> 0:03:22.640
<v Speaker 1>AI is super useful to me every single day and

0:03:22.960 --> 0:03:26.120
<v Speaker 1>was before the rise of chatbots. As you'll hear today,

0:03:26.320 --> 0:03:29.800
<v Speaker 1>AI has taken over every aspect of particle physics, and

0:03:29.840 --> 0:03:31.839
<v Speaker 1>I've been one of the people pushing that. So it's

0:03:31.919 --> 0:03:34.880
<v Speaker 1>AI all day for me. The thing that scares me

0:03:34.920 --> 0:03:40.000
<v Speaker 1>about AI is seeing my daughter ask AI things she

0:03:40.080 --> 0:03:43.360
<v Speaker 1>already knows or things she could just google, you know,

0:03:43.480 --> 0:03:46.040
<v Speaker 1>like you don't need to ask chat GBT what the

0:03:46.040 --> 0:03:48.720
<v Speaker 1>temperature is. You can just look it up, or you

0:03:48.720 --> 0:03:51.080
<v Speaker 1>could type it into Google. If you're too lazy for that,

0:03:51.480 --> 0:03:53.440
<v Speaker 1>you don't need to spin up at GPU for that.

0:03:54.480 --> 0:03:57.400
<v Speaker 1>But people have just become like used to punting everything

0:03:57.480 --> 0:03:59.880
<v Speaker 1>to AI. All thought processes use can just be shunted

0:04:00.120 --> 0:04:00.840
<v Speaker 1>out to AI.

0:04:01.200 --> 0:04:02.760
<v Speaker 2>And what scares you about that is that it's more

0:04:02.840 --> 0:04:05.680
<v Speaker 2>energy expensive, or just that you feel like we're relying

0:04:05.720 --> 0:04:08.120
<v Speaker 2>too much on AI for everything because like it's probably

0:04:08.120 --> 0:04:10.840
<v Speaker 2>going to tell you the right temperature in.

0:04:11.000 --> 0:04:14.920
<v Speaker 1>My yeah, who knows. Yeah, it's all of those things.

0:04:15.040 --> 0:04:18.880
<v Speaker 1>It's unnecessary and it's kind of lazy mentally, right, And

0:04:18.960 --> 0:04:22.080
<v Speaker 1>so the thing that scares me is that it leads

0:04:22.080 --> 0:04:24.440
<v Speaker 1>to people not thinking too much for themselves and just

0:04:24.480 --> 0:04:27.680
<v Speaker 1>relying on AI for everything exactly. It seems like a slippery.

0:04:27.279 --> 0:04:30.520
<v Speaker 2>Slope, got it? You know? I think when books came around,

0:04:30.560 --> 0:04:33.760
<v Speaker 2>there were people arguing that books are going to make

0:04:33.760 --> 0:04:35.599
<v Speaker 2>people lazy because they're not going to have to remember

0:04:35.640 --> 0:04:39.599
<v Speaker 2>stuff anymore. And so I guess what I'm saying is,

0:04:39.600 --> 0:04:41.880
<v Speaker 2>do you think that this is going to be our generations?

0:04:42.160 --> 0:04:44.600
<v Speaker 2>Like the kids are getting lazy and they're going to

0:04:44.640 --> 0:04:47.880
<v Speaker 2>look back and be like, is it actually different? I guess,

0:04:48.040 --> 0:04:49.400
<v Speaker 2>or do you think are we being stagy?

0:04:49.880 --> 0:04:52.599
<v Speaker 1>I know, I think this is a transformational moment, and

0:04:52.640 --> 0:04:54.200
<v Speaker 1>I think it's going to change what it means to

0:04:54.240 --> 0:04:57.440
<v Speaker 1>be human the way books do, right, because now you

0:04:57.480 --> 0:04:59.880
<v Speaker 1>can no longer just tell your friends stuff. You can

0:05:00.080 --> 0:05:02.279
<v Speaker 1>so communicate your thoughts to people who live a thousand

0:05:02.360 --> 0:05:05.480
<v Speaker 1>years later. That's amazing, or two millions of people all

0:05:05.520 --> 0:05:08.600
<v Speaker 1>at once. So that changes what it means to be human.

0:05:08.640 --> 0:05:11.400
<v Speaker 1>It makes it easier in some ways. That allows for laziness,

0:05:11.400 --> 0:05:14.880
<v Speaker 1>but also creates transformational new abilities. And I think we

0:05:14.960 --> 0:05:15.760
<v Speaker 1>got to lean into that.

0:05:15.920 --> 0:05:18.560
<v Speaker 2>Yeah, yeah, yeah, all right, Well let's go ahead and

0:05:18.680 --> 0:05:21.880
<v Speaker 2>jump into the meat of today's discussion. And today we're

0:05:21.880 --> 0:05:25.400
<v Speaker 2>talking about will AI solve physics?

0:05:25.640 --> 0:05:25.800
<v Speaker 3>Oh?

0:05:26.040 --> 0:05:28.200
<v Speaker 1>Even just the phrasing there makes me shutter.

0:05:29.520 --> 0:05:31.440
<v Speaker 2>I mean, that's a pretty big claim. Like, you know,

0:05:31.520 --> 0:05:33.280
<v Speaker 2>I think the first thing we need to talk about

0:05:33.320 --> 0:05:35.479
<v Speaker 2>after we hear what the extraordinaries have to say, is

0:05:35.800 --> 0:05:38.760
<v Speaker 2>what the heck do we even mean by solve physics?

0:05:38.800 --> 0:05:41.160
<v Speaker 2>Because that's, yeah, it's pretty big statements.

0:05:41.640 --> 0:05:43.960
<v Speaker 1>And I choose that because I was on a panel

0:05:44.279 --> 0:05:47.560
<v Speaker 1>at a recent huge AI conference next to a guy

0:05:47.600 --> 0:05:50.400
<v Speaker 1>from Anthropic who was all in on the fact that

0:05:50.400 --> 0:05:53.960
<v Speaker 1>AI was going to quote unquote solve physics. Wow, And

0:05:54.040 --> 0:05:57.040
<v Speaker 1>so this is not just some phrasing I made up, right.

0:05:56.960 --> 0:05:59.120
<v Speaker 2>Okay, Yeah, that is a big claim. And so let's

0:05:59.120 --> 0:06:01.840
<v Speaker 2>see if the extraordinary always think that AI is going

0:06:01.880 --> 0:06:03.799
<v Speaker 2>to solve physics.

0:06:04.040 --> 0:06:05.680
<v Speaker 1>Here's what folks had to say. If you would like

0:06:05.720 --> 0:06:08.960
<v Speaker 1>to join this group of exemplary guessers, please write to

0:06:09.040 --> 0:06:12.040
<v Speaker 1>us two questions at Danielankelly dot org.

0:06:12.560 --> 0:06:16.200
<v Speaker 4>To me, it's more of a question of should not will,

0:06:16.839 --> 0:06:21.240
<v Speaker 4>and as I don't trust AI to handle any human

0:06:21.400 --> 0:06:25.159
<v Speaker 4>safety concerns, I would say absolutely not.

0:06:25.880 --> 0:06:30.120
<v Speaker 1>If AI is chopped up to be intelligent or super intelligent,

0:06:30.400 --> 0:06:34.080
<v Speaker 1>then I believe we could solve any physics riddles or questions.

0:06:34.480 --> 0:06:38.159
<v Speaker 4>Don't think they're able to produce new thoughts, only predictive

0:06:38.440 --> 0:06:39.799
<v Speaker 4>of what already exists.

0:06:39.960 --> 0:06:43.039
<v Speaker 2>I know, I don't think they will solve physics.

0:06:42.600 --> 0:06:43.599
<v Speaker 4>That's already been done.

0:06:43.680 --> 0:06:44.520
<v Speaker 5>That's old news.

0:06:44.839 --> 0:06:47.280
<v Speaker 4>And the answer was forty two. I think that as

0:06:47.320 --> 0:06:50.720
<v Speaker 4>long as scientists treat AI as a tool rather than

0:06:50.760 --> 0:06:52.720
<v Speaker 4>as an oracle, it could be very helpful.

0:06:52.800 --> 0:06:58.200
<v Speaker 2>Yeah, I think the major breakthroughs will come from experimentation

0:06:58.680 --> 0:06:59.920
<v Speaker 2>made by real humans.

0:07:00.440 --> 0:07:03.120
<v Speaker 3>Is, in my opinion, an extremely broad question that needs

0:07:03.120 --> 0:07:04.039
<v Speaker 3>more specificity.

0:07:04.520 --> 0:07:07.240
<v Speaker 5>I think the idea that an a lens might solve

0:07:07.760 --> 0:07:09.920
<v Speaker 5>physics is akin to the idea that if you had

0:07:09.920 --> 0:07:12.680
<v Speaker 5>an infinite number of monkeys typing for an infinite amount

0:07:12.680 --> 0:07:16.040
<v Speaker 5>of time, they would eventually create the works of Shakespeare.

0:07:16.280 --> 0:07:19.480
<v Speaker 4>I think we need to define what solve physics means first.

0:07:20.000 --> 0:07:22.840
<v Speaker 4>AI might solve physics, but I think it's more likely.

0:07:22.720 --> 0:07:24.360
<v Speaker 5>That it a lie and claim that it has.

0:07:25.000 --> 0:07:28.800
<v Speaker 3>Yes, someday, it's going to be able to see relationships

0:07:28.840 --> 0:07:31.560
<v Speaker 3>where humans cannot, and it's going to come up with

0:07:31.560 --> 0:07:34.840
<v Speaker 3>some answers that we're probably right there in front of

0:07:34.920 --> 0:07:36.760
<v Speaker 3>us for a long time, but we just couldn't see it.

0:07:37.080 --> 0:07:39.720
<v Speaker 1>I cannot believe that you're asking this question.

0:07:40.320 --> 0:07:44.720
<v Speaker 4>Definitely help to solve physics problems, but who will take

0:07:44.800 --> 0:07:49.440
<v Speaker 4>credit for that? I think this requires an extreme scale

0:07:49.480 --> 0:07:53.520
<v Speaker 4>of creativity, so it will definitely assist, but not solve it.

0:07:53.560 --> 0:07:53.960
<v Speaker 1>That a little.

0:07:54.360 --> 0:07:56.760
<v Speaker 2>I think it's plausible that AI might be able to

0:07:56.880 --> 0:07:57.640
<v Speaker 2>solve physics.

0:07:57.920 --> 0:08:01.840
<v Speaker 4>It's a good plagiarization machine, but I've never heard of

0:08:02.480 --> 0:08:05.840
<v Speaker 4>current generation AI coming up with anything original.

0:08:06.240 --> 0:08:09.480
<v Speaker 2>All Right, quite a range of thoughts here, although I

0:08:09.520 --> 0:08:13.520
<v Speaker 2>feel like I'm seeing quite a bit of like skepticism

0:08:13.920 --> 0:08:15.760
<v Speaker 2>or a bit of concern with AI.

0:08:16.920 --> 0:08:19.560
<v Speaker 1>Yeah, you hear the whole spectrum here of people just

0:08:19.600 --> 0:08:23.360
<v Speaker 1>calling it a lazy plagiarism machine. Do people think absolutely

0:08:23.400 --> 0:08:26.200
<v Speaker 1>it's going to solve everything or it's at least plausible

0:08:26.280 --> 0:08:28.640
<v Speaker 1>that it could to people saying, like, what are even

0:08:28.680 --> 0:08:33.160
<v Speaker 1>talking about? That question means nothing? Yeah, so great job,

0:08:33.160 --> 0:08:35.679
<v Speaker 1>extraordinary is giving us the context to dive into this

0:08:35.800 --> 0:08:37.240
<v Speaker 1>really complicated question.

0:08:37.559 --> 0:08:40.320
<v Speaker 2>So when you were sitting on that panel next to

0:08:40.360 --> 0:08:43.920
<v Speaker 2>the guy from Anthropic, was it Sam Moltman. Is he

0:08:43.960 --> 0:08:46.320
<v Speaker 2>an Anthropic or is he somewhere else he's open AI?

0:08:46.480 --> 0:08:49.360
<v Speaker 1>Yeah, okay, I know it wasn't the head of Anthropic either.

0:08:49.240 --> 0:08:52.720
<v Speaker 2>Okay, all right, what did they mean by soul physics?

0:08:52.760 --> 0:08:55.360
<v Speaker 2>Did they mean like there will be no physics questions

0:08:55.480 --> 0:08:58.240
<v Speaker 2>left to answer when AI is done? What did what

0:08:58.280 --> 0:08:58.920
<v Speaker 2>did they mean?

0:08:59.280 --> 0:09:01.600
<v Speaker 1>Yeah? I don't even know. That was my pushback. I

0:09:01.600 --> 0:09:04.760
<v Speaker 1>said that this doesn't even mean anything. It's nonsense to

0:09:04.800 --> 0:09:07.640
<v Speaker 1>imagine that there's a moment where we have no more

0:09:07.720 --> 0:09:11.120
<v Speaker 1>questions about the nature of the universe, Like there's no

0:09:11.520 --> 0:09:14.400
<v Speaker 1>AI answer that's going to make me say, Okay, we're done.

0:09:14.559 --> 0:09:19.000
<v Speaker 1>I'm retiring, right, Yeah, my curiosity is gone. It just

0:09:19.040 --> 0:09:23.200
<v Speaker 1>doesn't just can't even imagine that scenario. And he didn't

0:09:23.200 --> 0:09:26.080
<v Speaker 1>have a great answer. He just gestured towards the rapid

0:09:26.120 --> 0:09:29.360
<v Speaker 1>increase in AI powers to suggest that it was going

0:09:29.400 --> 0:09:31.560
<v Speaker 1>to dot dot dot to some sort of Einsteinian level

0:09:31.559 --> 0:09:33.959
<v Speaker 1>of intelligence. But we can dig into that a little

0:09:33.960 --> 0:09:36.200
<v Speaker 1>bit more. When we get to that. I thought we

0:09:36.200 --> 0:09:39.440
<v Speaker 1>could start by talking about the role of AI in

0:09:39.480 --> 0:09:41.840
<v Speaker 1>physics today because also a lot of listeners write in

0:09:41.920 --> 0:09:44.240
<v Speaker 1>and ask me, like Hydan, how do you use AI

0:09:44.480 --> 0:09:47.719
<v Speaker 1>in your research? So let's start a little bit more

0:09:47.720 --> 0:09:51.960
<v Speaker 1>grounded and before chatbots took over everything, how we used

0:09:52.000 --> 0:09:55.080
<v Speaker 1>AI and more specifically machine learning in physics.

0:09:55.360 --> 0:09:58.000
<v Speaker 2>Okay, let's do it. But first, can we define what

0:09:58.200 --> 0:09:59.240
<v Speaker 2>is machine learning?

0:09:59.480 --> 0:10:02.600
<v Speaker 1>Yeah, have all of these names floating graphs, not clear

0:10:02.640 --> 0:10:06.200
<v Speaker 1>exactly what means. What The biggest topic is AI? Right?

0:10:06.280 --> 0:10:09.559
<v Speaker 1>That includes all of these things. Within AI, you have

0:10:09.679 --> 0:10:12.520
<v Speaker 1>machine learning. Machine learning is a kind of AI, but

0:10:12.520 --> 0:10:16.800
<v Speaker 1>it's mostly applied statistics. It's like patterned recognition. Think about

0:10:16.840 --> 0:10:21.160
<v Speaker 1>algorithms optimized for one specific task, right, Like, for example,

0:10:21.559 --> 0:10:23.840
<v Speaker 1>you want to tell a cat from a dog in

0:10:23.880 --> 0:10:27.120
<v Speaker 1>an internet video, or you want to know if a

0:10:27.160 --> 0:10:29.320
<v Speaker 1>particle is going to go left or right based on

0:10:29.400 --> 0:10:32.960
<v Speaker 1>some data that you have, right, solving a very specific

0:10:33.160 --> 0:10:37.080
<v Speaker 1>statistical task, and machine learning is helpful when that task

0:10:37.240 --> 0:10:39.280
<v Speaker 1>is too hard to do with pencil and paper, Like

0:10:39.320 --> 0:10:41.960
<v Speaker 1>you couldn't write down an equation that tells you is

0:10:42.000 --> 0:10:44.440
<v Speaker 1>this a cat or a dog an internet video, but

0:10:44.559 --> 0:10:46.839
<v Speaker 1>you could learn to tell the difference if you see

0:10:46.880 --> 0:10:50.280
<v Speaker 1>a bunch of examples. And so that's what machine learning is,

0:10:50.280 --> 0:10:53.800
<v Speaker 1>is like finding patterns iteratively through being exposed to a

0:10:53.840 --> 0:10:57.040
<v Speaker 1>bunch of examples, rather than like knowing how to write

0:10:57.040 --> 0:10:59.079
<v Speaker 1>an equation for the pattern on paper.

0:10:59.240 --> 0:11:01.280
<v Speaker 2>Got it okay? And I feel like I remember lots

0:11:01.280 --> 0:11:04.080
<v Speaker 2>of funny stories about how machine learning could be tricked,

0:11:04.120 --> 0:11:06.160
<v Speaker 2>Like if the cat happened to have a tennis ball

0:11:06.200 --> 0:11:08.640
<v Speaker 2>in its mouth or something, then it would be called

0:11:08.640 --> 0:11:11.000
<v Speaker 2>a dog because dogs always had tennis balls in their

0:11:11.000 --> 0:11:14.200
<v Speaker 2>mouths and stuff like that. Like the machine learning picks

0:11:14.280 --> 0:11:16.840
<v Speaker 2>up on things that maybe you wouldn't have guessed that

0:11:16.840 --> 0:11:17.800
<v Speaker 2>it would have picked up on.

0:11:17.920 --> 0:11:20.280
<v Speaker 1>Yes, and you always want to make sure that what

0:11:20.400 --> 0:11:23.720
<v Speaker 1>it's learning is meaningful to you. Like if you're training

0:11:23.840 --> 0:11:26.960
<v Speaker 1>machine learning to tell wolves from dogs, but all the

0:11:27.000 --> 0:11:29.240
<v Speaker 1>wolves have snow in the background in the picture, the

0:11:29.320 --> 0:11:31.080
<v Speaker 1>machine learning is just going to be like, oh yeah,

0:11:31.120 --> 0:11:32.920
<v Speaker 1>if it has snow in the background, it's a wolf.

0:11:32.960 --> 0:11:35.000
<v Speaker 1>And if it doesn't, it's a dog. But that's not

0:11:35.040 --> 0:11:37.280
<v Speaker 1>what you're interested in, right, ok. And so you have

0:11:37.320 --> 0:11:39.840
<v Speaker 1>to make sure you know what it's doing. And so

0:11:40.160 --> 0:11:43.319
<v Speaker 1>this is a very narrow kind of artificial intelligence. It's

0:11:43.320 --> 0:11:45.959
<v Speaker 1>not general intelligence at all. You can't ask this kind

0:11:45.960 --> 0:11:48.800
<v Speaker 1>of thing, what's the weather today? Or how should I

0:11:48.800 --> 0:11:50.880
<v Speaker 1>write this email back to my brother in law or

0:11:50.880 --> 0:11:53.560
<v Speaker 1>anything like that. Right, it can do what you tell

0:11:53.559 --> 0:11:55.880
<v Speaker 1>it to do. And we use this all the time

0:11:55.920 --> 0:11:59.160
<v Speaker 1>in particle physics because all of our discoveries these days

0:11:59.240 --> 0:12:02.320
<v Speaker 1>are statisic stickle in nature, you know, in the same

0:12:02.320 --> 0:12:04.000
<v Speaker 1>way you can look at a picture and say, like,

0:12:04.280 --> 0:12:06.520
<v Speaker 1>seventy five percent chance this is a dog, twenty five

0:12:06.559 --> 0:12:09.400
<v Speaker 1>percent chance it's a cat. That's what all of particle

0:12:09.440 --> 0:12:12.040
<v Speaker 1>physics is. And it makes me a little sad because

0:12:12.120 --> 0:12:14.679
<v Speaker 1>it didn't used to be like in the glory days

0:12:14.720 --> 0:12:19.040
<v Speaker 1>of particle physics, back when they were discovering the obvious particles,

0:12:19.120 --> 0:12:22.200
<v Speaker 1>you know, and by obvious particles, I mean things that

0:12:22.240 --> 0:12:24.480
<v Speaker 1>were high rate. There was like a lot of examples

0:12:24.480 --> 0:12:28.080
<v Speaker 1>of them and a low background, meaning it was very

0:12:28.120 --> 0:12:31.439
<v Speaker 1>hard to imagine anything else making that kind of signature,

0:12:31.800 --> 0:12:35.120
<v Speaker 1>like the positron discovery. This guy won a Nobel Prize

0:12:35.160 --> 0:12:38.280
<v Speaker 1>for discovering the positron off of one example. Wow, he's

0:12:38.360 --> 0:12:41.080
<v Speaker 1>like a picture of a track of a positron and

0:12:41.120 --> 0:12:44.360
<v Speaker 1>it can't be anything else, and boom, Nobel prize. That's

0:12:44.360 --> 0:12:45.840
<v Speaker 1>an amazing piece of data right there.

0:12:46.080 --> 0:12:49.400
<v Speaker 2>It's a shame we weren't born a few decades earlier.

0:12:49.960 --> 0:12:51.160
<v Speaker 2>It would have been easier then.

0:12:51.520 --> 0:12:54.160
<v Speaker 1>And because most of those are all used up these days,

0:12:54.240 --> 0:12:57.480
<v Speaker 1>particle physics is focused on more rare particles, particles that

0:12:57.520 --> 0:13:00.040
<v Speaker 1>come out of collisions once a trillion times, once a

0:13:00.200 --> 0:13:04.640
<v Speaker 1>quadrillion times, Otherwise they would have been discovered by previous generations,

0:13:05.280 --> 0:13:08.640
<v Speaker 1>and the backgrounds are higher. It's easier to mimic the

0:13:08.640 --> 0:13:11.319
<v Speaker 1>signatures of the particles we're looking for because we never

0:13:11.320 --> 0:13:14.160
<v Speaker 1>see the particle directly. Like when we found the Higgs

0:13:14.160 --> 0:13:17.520
<v Speaker 1>boson fifteen years ago, we didn't say, like, here's one

0:13:17.559 --> 0:13:20.520
<v Speaker 1>Higgs boson, give me a Nobel prize. Because the Higgs

0:13:20.559 --> 0:13:23.080
<v Speaker 1>boson doesn't last long enough. It blows up into other

0:13:23.120 --> 0:13:26.120
<v Speaker 1>particles and we look at the patterns among those particles

0:13:26.120 --> 0:13:28.640
<v Speaker 1>which are not Higgs bosons, and we say, this is

0:13:28.880 --> 0:13:32.560
<v Speaker 1>likely to be a Higgs boson statistically, this is more

0:13:32.640 --> 0:13:34.880
<v Speaker 1>likely to be a Higgs than not. But there are

0:13:34.880 --> 0:13:36.880
<v Speaker 1>other ways to make that same pattern, and so you

0:13:36.960 --> 0:13:39.080
<v Speaker 1>can never know for sure. You can never point to

0:13:39.080 --> 0:13:42.280
<v Speaker 1>one event and say this is my Higgs boson. All

0:13:42.320 --> 0:13:44.120
<v Speaker 1>you can do is say, look is a pile of

0:13:44.160 --> 0:13:47.280
<v Speaker 1>these events, all very very similar and so very likely

0:13:47.320 --> 0:13:50.800
<v Speaker 1>the Higgs boson exists. And because all of our discoveries

0:13:50.840 --> 0:13:54.400
<v Speaker 1>are statistical, machine learning is very very powerful. This is

0:13:54.480 --> 0:13:56.760
<v Speaker 1>exactly what it's good at. If you want to be

0:13:56.800 --> 0:13:59.520
<v Speaker 1>able to tell collisions that lead to Higgs bosons from

0:13:59.520 --> 0:14:03.640
<v Speaker 1>collisions that don't, and they're very similar in your detectors,

0:14:04.080 --> 0:14:06.960
<v Speaker 1>you can give machine learning a bunch of examples and say, look,

0:14:07.040 --> 0:14:09.200
<v Speaker 1>here's a bunch of Higgs bosons, here's a bunch of

0:14:09.280 --> 0:14:11.800
<v Speaker 1>not Higgs bosons. Now I'm going to give you data

0:14:11.840 --> 0:14:14.840
<v Speaker 1>from the real collider. Tell me what's the chance that

0:14:14.920 --> 0:14:18.040
<v Speaker 1>each one is a Higgs boson or not. If you

0:14:18.080 --> 0:14:20.320
<v Speaker 1>can't just like write down an equation to tell you

0:14:20.680 --> 0:14:23.280
<v Speaker 1>the probability of being a Higgs boson, and nobody knows

0:14:23.280 --> 0:14:25.960
<v Speaker 1>how to do that because it's very complicated, then this

0:14:26.000 --> 0:14:29.320
<v Speaker 1>can learn exactly how to do that calculation for you.

0:14:29.840 --> 0:14:32.240
<v Speaker 2>And is machine learning in the end able to give

0:14:32.280 --> 0:14:37.600
<v Speaker 2>you like a statistical output like I'm ninety eight percent sure,

0:14:38.320 --> 0:14:40.320
<v Speaker 2>or does it just say yeah, I think I found it.

0:14:41.960 --> 0:14:43.960
<v Speaker 1>No, it can do exactly the kind of statistics that

0:14:44.000 --> 0:14:47.880
<v Speaker 1>you're talking about. It calculates likelihoods and does likelihood ratios

0:14:47.960 --> 0:14:51.600
<v Speaker 1>for statistic nerves out there. It's learning to approximate the

0:14:51.760 --> 0:14:53.960
<v Speaker 1>likelihood ratio, which if you knew how to write it

0:14:54.000 --> 0:14:56.360
<v Speaker 1>down as a formula, you would write down and it

0:14:56.400 --> 0:14:59.920
<v Speaker 1>can approximate that very very well. So it's really incredible.

0:15:00.240 --> 0:15:03.400
<v Speaker 2>Yeah, that is amazing, okay, And so it's allowing you

0:15:03.440 --> 0:15:05.440
<v Speaker 2>to do things that we wouldn't be able to do

0:15:05.760 --> 0:15:07.440
<v Speaker 2>on our own because the math would just be like,

0:15:07.520 --> 0:15:09.240
<v Speaker 2>way too complicated.

0:15:08.800 --> 0:15:10.800
<v Speaker 1>That's right, And the math is too complicated. And so

0:15:10.920 --> 0:15:14.280
<v Speaker 1>for years we did the math approximately we said, well,

0:15:14.280 --> 0:15:16.920
<v Speaker 1>looks it's complicated, but I can write down a simpler

0:15:17.040 --> 0:15:19.440
<v Speaker 1>version of it, which kind of works. And we were

0:15:19.480 --> 0:15:22.120
<v Speaker 1>giving away information. We were not using all of the

0:15:22.160 --> 0:15:24.680
<v Speaker 1>power of our data. But machine learning just slurps in

0:15:24.760 --> 0:15:27.480
<v Speaker 1>all that data and says, here's the best way to

0:15:27.520 --> 0:15:30.640
<v Speaker 1>discriminate between these two things. It lets us tel cats

0:15:30.640 --> 0:15:33.880
<v Speaker 1>from dogs really really well. Which helps us find rare

0:15:33.960 --> 0:15:37.120
<v Speaker 1>cats in our data much more effectively. And since our

0:15:37.160 --> 0:15:39.720
<v Speaker 1>data is so expensive and so powerful, this is a

0:15:39.880 --> 0:15:41.880
<v Speaker 1>huge boon to particle physics.

0:15:42.280 --> 0:15:46.080
<v Speaker 2>It would be a shame to miss any particle cats exactly.

0:15:47.080 --> 0:15:49.240
<v Speaker 1>And so the lesson here is that it's very powerful,

0:15:49.320 --> 0:15:53.440
<v Speaker 1>but it's applied statistics to very targeted tasks, and it's

0:15:53.440 --> 0:15:57.280
<v Speaker 1>not just limited to labeling collisions and saying this one's Higgs,

0:15:57.400 --> 0:16:00.240
<v Speaker 1>this one's not Higgs, or this one's dark matter, it's

0:16:00.240 --> 0:16:04.080
<v Speaker 1>not dark matter. We also use generative AI. Generative AI

0:16:04.280 --> 0:16:06.920
<v Speaker 1>is the kind that can create new examples you type

0:16:06.920 --> 0:16:09.760
<v Speaker 1>into chat GPT, like give me a picture of a

0:16:09.840 --> 0:16:12.840
<v Speaker 1>kitten riding a horse on the moon or whatever, and

0:16:13.040 --> 0:16:15.800
<v Speaker 1>it makes one for you, right. And we use the

0:16:15.840 --> 0:16:20.000
<v Speaker 1>same kind of technology to very rapidly generate simulated examples.

0:16:20.400 --> 0:16:23.640
<v Speaker 1>Something machine learning needs are lots and lots of examples

0:16:23.680 --> 0:16:26.840
<v Speaker 1>to learn from, and so you have to generate examples.

0:16:26.880 --> 0:16:29.000
<v Speaker 1>Here's what Higgs boson would look like, Here's what a

0:16:29.040 --> 0:16:31.400
<v Speaker 1>not Higgs boson would look like. You need millions or

0:16:31.440 --> 0:16:34.400
<v Speaker 1>billions of those examples for it to really learn what

0:16:34.440 --> 0:16:37.000
<v Speaker 1>you need to do. But generating those examples can be

0:16:37.080 --> 0:16:40.000
<v Speaker 1>very very slow. Like if you generate a collision and

0:16:40.040 --> 0:16:42.320
<v Speaker 1>you have a Higgs boson that leads to two quarks,

0:16:42.480 --> 0:16:45.560
<v Speaker 1>those quarks turn into a shower of particles. To simulate

0:16:45.600 --> 0:16:47.440
<v Speaker 1>what that's going to look like in our detector, you

0:16:47.480 --> 0:16:49.960
<v Speaker 1>have to simulate what all those particles do when they

0:16:50.040 --> 0:16:52.640
<v Speaker 1>hit your detector. What happens when a muon hits a

0:16:52.640 --> 0:16:55.080
<v Speaker 1>block of copper. It creates a bunch of photons, It

0:16:55.120 --> 0:16:58.280
<v Speaker 1>ionizes atoms. Lots of complicated things happen.

0:16:59.480 --> 0:17:03.560
<v Speaker 2>You californiensis Sorry, you just said a lot of physics stuff.

0:17:03.560 --> 0:17:05.160
<v Speaker 2>So I get to say a species name.

0:17:05.760 --> 0:17:09.800
<v Speaker 1>Oh you're right, nice, Okay, sorry, keep going Latin words

0:17:09.800 --> 0:17:13.840
<v Speaker 1>for the wind. So what happens when particles smash into matter?

0:17:13.920 --> 0:17:16.080
<v Speaker 1>You get all sorts of complicated stuff which is hard

0:17:16.160 --> 0:17:18.080
<v Speaker 1>to track. And then you have to track those particles,

0:17:18.080 --> 0:17:21.240
<v Speaker 1>which creates more particles, which creates more particles whose Latin

0:17:21.280 --> 0:17:24.280
<v Speaker 1>names I will not use, and it becomes very very

0:17:24.320 --> 0:17:27.479
<v Speaker 1>computationally expensive. CERN has one of the biggest data centers

0:17:27.520 --> 0:17:30.679
<v Speaker 1>in the world, and a huge fraction of those computers

0:17:30.760 --> 0:17:34.560
<v Speaker 1>spend all of their time just simulating particle collisions and interactions.

0:17:34.840 --> 0:17:37.679
<v Speaker 1>Because we need all of that, but JENNAI can do

0:17:37.720 --> 0:17:40.320
<v Speaker 1>this much faster. You give a bunch of examples to

0:17:40.359 --> 0:17:42.560
<v Speaker 1>GENAI and then say, okay, now make me a new one.

0:17:42.560 --> 0:17:45.399
<v Speaker 1>It's perfectly akin to saying give me a cat on

0:17:45.440 --> 0:17:47.119
<v Speaker 1>a horse on the moon. You can say, give me

0:17:47.160 --> 0:17:50.000
<v Speaker 1>a Higgs Boson with this mass at this angle, and

0:17:50.040 --> 0:17:53.000
<v Speaker 1>you can interpotate between all of its examples and flash

0:17:53.080 --> 0:17:55.760
<v Speaker 1>give you an example without burning all of those CPUs.

0:17:55.760 --> 0:17:59.160
<v Speaker 1>It's much much more efficient. So we're not just using

0:17:59.200 --> 0:18:01.920
<v Speaker 1>AI to separate particles, we're also using it to train

0:18:02.000 --> 0:18:04.800
<v Speaker 1>it how to separate particles all over the place. In

0:18:04.880 --> 0:18:08.520
<v Speaker 1>experimental particle physics, AI has touched every single part of

0:18:08.560 --> 0:18:12.480
<v Speaker 1>our work because almost everything we do is statistical and computational.

0:18:13.000 --> 0:18:16.600
<v Speaker 2>Okay, so it sounds like AI is very helpful if

0:18:16.600 --> 0:18:19.720
<v Speaker 2>you are an experimentalist and you're like tracking particles. But

0:18:19.800 --> 0:18:23.320
<v Speaker 2>what if you are like a pen to paper theoretical physicist.

0:18:23.359 --> 0:18:24.359
<v Speaker 2>Is AI still helpful?

0:18:24.680 --> 0:18:28.200
<v Speaker 1>Yeah? Great question, because when we talk about solving physics,

0:18:28.200 --> 0:18:31.600
<v Speaker 1>we usually mean like figuring out the theory of the universe.

0:18:32.240 --> 0:18:34.520
<v Speaker 1>And there's a lot of talk online in the podos

0:18:34.520 --> 0:18:37.359
<v Speaker 1>sphere about like the crisis in physics, which is mostly

0:18:37.400 --> 0:18:40.440
<v Speaker 1>bs frankly, But it's about theory, right. People want to

0:18:40.440 --> 0:18:44.800
<v Speaker 1>see progress in understanding the universe, and theory has not

0:18:45.040 --> 0:18:48.080
<v Speaker 1>embraced machine learning and AI and nearly as much as

0:18:48.119 --> 0:18:51.920
<v Speaker 1>experimental physics yet. And I think that's because mostly those

0:18:51.920 --> 0:18:55.280
<v Speaker 1>folks are pencil and paper people. They're not as computational.

0:18:55.359 --> 0:18:57.960
<v Speaker 1>The computational people end up in the experimental side mostly,

0:18:58.560 --> 0:19:01.160
<v Speaker 1>and so the theory people don't see their work as

0:19:01.200 --> 0:19:04.600
<v Speaker 1>having like data, but I see their work as having data.

0:19:04.920 --> 0:19:08.320
<v Speaker 1>You know, they are exploring huge parameter spaces of complicated

0:19:08.359 --> 0:19:12.359
<v Speaker 1>particle theories. That's data. That's what machine learning is good at.

0:19:12.400 --> 0:19:15.200
<v Speaker 1>So I've actually done a few projects in particle theory

0:19:15.320 --> 0:19:19.840
<v Speaker 1>using tools from experimental particle physics AI to solve problems

0:19:19.840 --> 0:19:22.720
<v Speaker 1>in theoretical physics, which are super duper fun and I

0:19:22.720 --> 0:19:26.280
<v Speaker 1>think it's a whole flourishing area there. So it's just begun.

0:19:26.359 --> 0:19:28.879
<v Speaker 1>I think it's a little bit behind the experimental side,

0:19:29.280 --> 0:19:32.240
<v Speaker 1>but in the end, all of these things are places

0:19:32.240 --> 0:19:34.879
<v Speaker 1>where humans are asking questions. Right. The AI is not

0:19:35.040 --> 0:19:38.240
<v Speaker 1>doing physics here, it's not solving physics. It's a tool.

0:19:38.400 --> 0:19:42.560
<v Speaker 1>It's an assistant at best. It's helping humans answer questions

0:19:42.600 --> 0:19:43.560
<v Speaker 1>about the universe.

0:19:43.800 --> 0:19:48.080
<v Speaker 2>All right, So we've talked about machine learning MLS. Next

0:19:48.119 --> 0:19:51.000
<v Speaker 2>we're going to talk about l l MS. I like

0:19:51.000 --> 0:19:53.600
<v Speaker 2>that we're focusing on a small subset of letters and

0:19:53.640 --> 0:19:56.040
<v Speaker 2>when we get back we'll find out what that extra

0:19:56.240 --> 0:20:19.880
<v Speaker 2>L means. All right, we are back, and now instead

0:20:19.880 --> 0:20:22.080
<v Speaker 2>of talking about machine learning, we're going to talk about

0:20:22.480 --> 0:20:26.000
<v Speaker 2>large language models and how these help physicists. So, Daniel,

0:20:26.240 --> 0:20:29.160
<v Speaker 2>how are you all using llms?

0:20:29.560 --> 0:20:33.679
<v Speaker 1>So mostly the conversation about AI solving physics is inspired

0:20:33.720 --> 0:20:38.600
<v Speaker 1>by the rise of these chatbots, Claude, chat, GBT, Gemini,

0:20:38.680 --> 0:20:43.560
<v Speaker 1>all these things and their supposed feeling of broader intelligence. Right,

0:20:43.880 --> 0:20:45.800
<v Speaker 1>And everything we talked about in the last segment was

0:20:45.840 --> 0:20:49.520
<v Speaker 1>machine learning. It was statistical learning applied to very specific tasks,

0:20:49.840 --> 0:20:51.360
<v Speaker 1>not the kind of thing that's going to tell you,

0:20:51.440 --> 0:20:54.040
<v Speaker 1>like what the recipe to make a certain kind of

0:20:54.119 --> 0:20:57.440
<v Speaker 1>fancy sandwich or whatever. The things that people use llms

0:20:57.480 --> 0:21:01.040
<v Speaker 1>for LMS are a very different breed of artificial intelligence.

0:21:01.040 --> 0:21:05.000
<v Speaker 1>It's built on natural language. Instead of just like numerical inputs,

0:21:05.400 --> 0:21:08.560
<v Speaker 1>it outputs with a human like language. Right. It feels

0:21:08.920 --> 0:21:12.639
<v Speaker 1>much more natural than mathematical. And these things came on

0:21:12.680 --> 0:21:15.280
<v Speaker 1>the scene a few years ago, and really that's what

0:21:15.359 --> 0:21:18.480
<v Speaker 1>people are interested in. Are llms going to change the

0:21:18.480 --> 0:21:21.080
<v Speaker 1>way we do science? Are they going to crack problems

0:21:21.080 --> 0:21:24.119
<v Speaker 1>that human physicists have not been able to crack? But

0:21:24.160 --> 0:21:27.640
<v Speaker 1>I want to underline that, like these things are still mathematical,

0:21:28.080 --> 0:21:31.600
<v Speaker 1>Like an LLM is not something different from any other

0:21:31.720 --> 0:21:34.920
<v Speaker 1>kind of AI. Underneath the hood, it's the same pieces

0:21:34.920 --> 0:21:37.400
<v Speaker 1>of technology just arranged in a different way to solve

0:21:37.440 --> 0:21:42.400
<v Speaker 1>a different problem, and then just scaled up tremendously, like hugely,

0:21:42.640 --> 0:21:46.800
<v Speaker 1>just like enormous numbers of these nodes and neurons, And

0:21:46.840 --> 0:21:48.800
<v Speaker 1>I think it's worth maybe taking them in to explain

0:21:49.200 --> 0:21:52.159
<v Speaker 1>what is going on underneath the hood of an LM.

0:21:52.640 --> 0:21:56.800
<v Speaker 1>A lot of people understand that lms are statistically predicting text.

0:21:57.600 --> 0:22:00.200
<v Speaker 1>Like if they are writing an answer to you, they

0:22:00.200 --> 0:22:03.040
<v Speaker 1>are trying to generate for you what you want, and

0:22:03.560 --> 0:22:06.080
<v Speaker 1>they've read a bunch of text, and so if they're

0:22:06.119 --> 0:22:08.480
<v Speaker 1>trying to predict like the next word and the sequence

0:22:08.520 --> 0:22:10.639
<v Speaker 1>of the answer for you, they look at all the

0:22:10.680 --> 0:22:13.600
<v Speaker 1>texts they've read and they say, well, statistically, what is

0:22:13.640 --> 0:22:16.280
<v Speaker 1>the word that usually follows the current word? You know,

0:22:16.280 --> 0:22:19.520
<v Speaker 1>so I'm writing a sentence the sentences Kelly is a

0:22:19.920 --> 0:22:22.320
<v Speaker 1>blank and then you know, I'm trying to generate the

0:22:22.359 --> 0:22:24.720
<v Speaker 1>next word. It's some adjective. And so I'm looking through

0:22:25.080 --> 0:22:27.840
<v Speaker 1>all of my training right all the text on the

0:22:27.880 --> 0:22:30.760
<v Speaker 1>internet that describes Kelly, and I'm trying to statistically predict

0:22:30.880 --> 0:22:33.639
<v Speaker 1>you know, what is the appropriate word to fill in

0:22:33.680 --> 0:22:35.680
<v Speaker 1>the blank. And you know, I'm not going to give

0:22:35.680 --> 0:22:39.040
<v Speaker 1>you that word other than a wonderful co host and

0:22:39.119 --> 0:22:40.280
<v Speaker 1>fantastic scientist.

0:22:41.040 --> 0:22:43.000
<v Speaker 2>Thanks. Thanks, you only had but you said it was

0:22:43.040 --> 0:22:44.840
<v Speaker 2>going to be one word, so I was waiting for it,

0:22:44.880 --> 0:22:47.280
<v Speaker 2>but I'm going to I'll let you off the hot seat,

0:22:49.160 --> 0:22:54.800
<v Speaker 2>although you've really backed yourself into a corner, all right.

0:22:55.080 --> 0:22:57.639
<v Speaker 1>And that's a fair sort of surface level understanding of

0:22:57.760 --> 0:23:01.919
<v Speaker 1>lms and what they're doing, but really misses the transformational

0:23:01.960 --> 0:23:05.640
<v Speaker 1>technology that's underneath them, because lms are doing more than

0:23:05.760 --> 0:23:09.480
<v Speaker 1>just like predicting exactly the next word. They're using something

0:23:09.520 --> 0:23:13.040
<v Speaker 1>called attention, which is really crucial to understand why these

0:23:13.040 --> 0:23:16.040
<v Speaker 1>things are so much better than just like statistical parrots.

0:23:16.640 --> 0:23:20.360
<v Speaker 1>And attention helps you understand which words in a sentence

0:23:20.400 --> 0:23:24.360
<v Speaker 1>are contributing to the context and contributing meaning to the sentence.

0:23:24.680 --> 0:23:26.800
<v Speaker 1>So say, for example, you have a sentence like the

0:23:26.920 --> 0:23:31.080
<v Speaker 1>river bank was steep. Now, bank is a complicated English

0:23:31.119 --> 0:23:33.560
<v Speaker 1>word because it has lots of meanings. Right, Is it

0:23:33.600 --> 0:23:35.959
<v Speaker 1>where you put your money? Is it how you shoot

0:23:36.000 --> 0:23:39.320
<v Speaker 1>a basketball off the backboard? Is it what keeps a

0:23:39.440 --> 0:23:43.080
<v Speaker 1>river in place? Right? In this sentence, the words river

0:23:43.400 --> 0:23:47.400
<v Speaker 1>and steep are important clues that tell you what bank means. Right,

0:23:47.440 --> 0:23:49.840
<v Speaker 1>So when you hear that sentence, you're like, Daniel's not

0:23:49.840 --> 0:23:53.320
<v Speaker 1>talking about depositing money or shooting three pointers, right, He's

0:23:53.320 --> 0:23:57.200
<v Speaker 1>talking about a river. And what the attention mechanism does

0:23:57.280 --> 0:24:00.600
<v Speaker 1>inside these lms is let the chatbot at all of

0:24:00.640 --> 0:24:04.199
<v Speaker 1>the other words, not just the one preceding word, and

0:24:04.280 --> 0:24:06.960
<v Speaker 1>so knows what to pay attention to. It helps it

0:24:07.080 --> 0:24:09.639
<v Speaker 1>understand the context. It's like if you go to a

0:24:09.640 --> 0:24:11.879
<v Speaker 1>science conference and you need to understand some crucial thing

0:24:11.920 --> 0:24:14.320
<v Speaker 1>for your research, you don't just talk to the person

0:24:14.359 --> 0:24:17.080
<v Speaker 1>next to you. You seek out the most relevant person

0:24:17.160 --> 0:24:19.440
<v Speaker 1>to pay attention to, right, the person who's going to

0:24:19.480 --> 0:24:23.160
<v Speaker 1>help you understand the question you're asking. And so that's

0:24:23.160 --> 0:24:26.119
<v Speaker 1>what attention is. And then they couple attention with this

0:24:26.240 --> 0:24:31.000
<v Speaker 1>technology called transformers, which are not a little plastic toys

0:24:31.040 --> 0:24:33.639
<v Speaker 1>that you used to play with, or cartoons, or even

0:24:33.680 --> 0:24:37.720
<v Speaker 1>the things on wires out there that help transform the voltage.

0:24:38.119 --> 0:24:41.719
<v Speaker 1>They're just like stacked layers of these attention modules. So

0:24:41.760 --> 0:24:44.280
<v Speaker 1>the first layer helps you understand the words, and the

0:24:44.320 --> 0:24:47.199
<v Speaker 1>second layer helps you understand the rest of the paragraph.

0:24:47.240 --> 0:24:50.119
<v Speaker 1>And by the time you have multiple stacked layers of attention,

0:24:50.480 --> 0:24:53.040
<v Speaker 1>you can understand. Okay, here's the context of the question

0:24:53.119 --> 0:24:55.800
<v Speaker 1>that's being asked, and therefore here's the answer that I

0:24:55.840 --> 0:24:58.639
<v Speaker 1>should give. And so it's this combination of attention and

0:24:58.720 --> 0:25:01.720
<v Speaker 1>transformers that make LMS so powerful.

0:25:02.160 --> 0:25:04.200
<v Speaker 2>Okay, I didn't know any of that.

0:25:04.480 --> 0:25:07.480
<v Speaker 1>All right, So then what are LMS useful for? Right,

0:25:07.680 --> 0:25:10.600
<v Speaker 1>I'm not really interested in using them to write essays

0:25:10.680 --> 0:25:14.040
<v Speaker 1>about thirteenth century Europe or whatever. Today we're talking about

0:25:14.119 --> 0:25:17.200
<v Speaker 1>using them in physics. So I use lms all the time.

0:25:17.280 --> 0:25:19.280
<v Speaker 1>And here's an example. I find them very useful for

0:25:19.760 --> 0:25:22.200
<v Speaker 1>I need to learn some new software package, right because

0:25:22.240 --> 0:25:25.480
<v Speaker 1>I've started a new research project, and there's a tool

0:25:25.600 --> 0:25:28.720
<v Speaker 1>that simulates the thing I need, for example, and I

0:25:28.760 --> 0:25:33.160
<v Speaker 1>download the tool. It's written terribly. The documentation is garbage.

0:25:33.320 --> 0:25:36.879
<v Speaker 1>There's like hundreds of pages of tables of technical stuff

0:25:36.880 --> 0:25:39.520
<v Speaker 1>that's impenetrable. I try to run the code. It doesn't

0:25:39.560 --> 0:25:41.879
<v Speaker 1>work the first time, Right, I look at the code.

0:25:42.000 --> 0:25:46.320
<v Speaker 1>It's written by physicists without any adult supervision. Right. And

0:25:46.400 --> 0:25:49.320
<v Speaker 1>so in past days, this would be a huge project.

0:25:49.480 --> 0:25:51.760
<v Speaker 1>Make this thing work, validate that it's doing what I

0:25:51.760 --> 0:25:55.320
<v Speaker 1>think it should be doing, Understand the documentation, how it's wrong,

0:25:55.440 --> 0:25:58.560
<v Speaker 1>how the thing actually works, you know. And in practice,

0:25:58.560 --> 0:26:00.000
<v Speaker 1>what I would do is like find somebody else who

0:26:00.160 --> 0:26:03.520
<v Speaker 1>made it work and start from their example. Somebody else's

0:26:03.520 --> 0:26:06.000
<v Speaker 1>invested weeks and weeks to making this thing work. I'm

0:26:06.040 --> 0:26:08.000
<v Speaker 1>going to download their scripts. I'm going to start from

0:26:08.000 --> 0:26:10.560
<v Speaker 1>their code. I'm going to ask them questions when it breaks, right,

0:26:10.800 --> 0:26:13.600
<v Speaker 1>I'm going to lean on some kind of expertise. Otherwise

0:26:13.600 --> 0:26:15.439
<v Speaker 1>I might have to sink weeks into getting this thing

0:26:15.520 --> 0:26:18.320
<v Speaker 1>to work. I no longer have to do that. Now

0:26:18.400 --> 0:26:21.239
<v Speaker 1>I can ask an LLM, hey, read this document and

0:26:21.359 --> 0:26:24.760
<v Speaker 1>write me an example configuration script to make this thing work. Okay,

0:26:24.800 --> 0:26:28.640
<v Speaker 1>it didn't work, debug the issue, and within a few cycles,

0:26:28.680 --> 0:26:31.240
<v Speaker 1>I can usually get a new software package up and

0:26:31.320 --> 0:26:36.680
<v Speaker 1>running and producing useful, validated results in hours, if not minutes.

0:26:36.720 --> 0:26:39.640
<v Speaker 1>Where it used to be either weeks or totally impenetrable.

0:26:40.119 --> 0:26:43.159
<v Speaker 2>It is pretty amazing. Like you know, whenever I use

0:26:43.200 --> 0:26:45.359
<v Speaker 2>a new package in R or something and I know

0:26:45.520 --> 0:26:48.399
<v Speaker 2>inevitably put a comma in the wrong place or something,

0:26:48.440 --> 0:26:51.360
<v Speaker 2>I used to look for that comma for hours and

0:26:51.400 --> 0:26:54.639
<v Speaker 2>now I'm like, jadgbt, where did I put the comma?

0:26:54.760 --> 0:26:57.560
<v Speaker 2>That's right, it doesn't belong and they it finds it immediately,

0:26:57.640 --> 0:26:58.399
<v Speaker 2>saves me hours.

0:26:58.480 --> 0:27:01.920
<v Speaker 1>And it's not magic. Right, Remember what are these lms doing.

0:27:01.960 --> 0:27:05.240
<v Speaker 1>They understand context, they know how to use attention to

0:27:05.320 --> 0:27:08.959
<v Speaker 1>focus the question you're asking to produce the right answer,

0:27:09.400 --> 0:27:13.280
<v Speaker 1>and they're very good at digesting huge volumes of input. Right,

0:27:13.600 --> 0:27:15.720
<v Speaker 1>the things that humans are bad at. For the same

0:27:15.760 --> 0:27:18.479
<v Speaker 1>reason that we're using machine learning to improve on our

0:27:18.560 --> 0:27:21.920
<v Speaker 1>data analysis and the particle physics colliders, because they're good

0:27:21.960 --> 0:27:24.159
<v Speaker 1>at the things we're bad at, like understanding all the

0:27:24.200 --> 0:27:28.040
<v Speaker 1>nuances of some high dimensional statistics problem. Here we're using

0:27:28.080 --> 0:27:29.840
<v Speaker 1>them to do something that we are bad at, and

0:27:29.880 --> 0:27:33.359
<v Speaker 1>that is the key, right, lean on their expertise, find

0:27:33.359 --> 0:27:35.399
<v Speaker 1>the thing that they can do that we can't, and

0:27:35.440 --> 0:27:38.320
<v Speaker 1>then they can contribute and they can go beyond. Just

0:27:38.440 --> 0:27:42.280
<v Speaker 1>like reading documentation for software packages, they can also read

0:27:42.320 --> 0:27:45.800
<v Speaker 1>physics papers. There are so many papers put out every

0:27:45.840 --> 0:27:48.399
<v Speaker 1>single day. If you look at the archive, which is

0:27:48.440 --> 0:27:51.240
<v Speaker 1>where particle physicists put out their papers every day, there's

0:27:51.320 --> 0:27:53.679
<v Speaker 1>a dozen papers every day, all of which are pretty

0:27:53.680 --> 0:27:56.000
<v Speaker 1>good and worth reading. I can't keep up with all

0:27:56.040 --> 0:27:58.720
<v Speaker 1>of those. In addition, I should be reading all the

0:27:58.800 --> 0:28:01.760
<v Speaker 1>machine learning papers. Should be reading statistics papers. I should

0:28:01.760 --> 0:28:04.760
<v Speaker 1>be reading math papers. Some math nerd out there could

0:28:04.760 --> 0:28:08.640
<v Speaker 1>have just invented some incredible new thing that they're into

0:28:08.760 --> 0:28:11.000
<v Speaker 1>just because they're a math nerd. But it might like

0:28:11.200 --> 0:28:13.960
<v Speaker 1>solve exactly the problem we have in particle physics, and

0:28:14.000 --> 0:28:15.800
<v Speaker 1>they don't know, and we don't know, and who knows

0:28:15.840 --> 0:28:19.240
<v Speaker 1>how to put these things together. The reason cross discipline

0:28:19.320 --> 0:28:22.680
<v Speaker 1>or research is often so fruitful and leads to breakthroughs

0:28:22.760 --> 0:28:25.800
<v Speaker 1>and advances is exactly because of this, Because people are

0:28:25.880 --> 0:28:28.240
<v Speaker 1>unaware of solutions in adjacent fields, and when you just

0:28:28.359 --> 0:28:30.960
<v Speaker 1>smush them together, you get peanut butter and chocolate. Right.

0:28:31.280 --> 0:28:34.399
<v Speaker 2>It's amazing, But you've got to be careful because you

0:28:34.480 --> 0:28:37.760
<v Speaker 2>might get information in your head that doesn't actually exist

0:28:37.800 --> 0:28:42.040
<v Speaker 2>from doing that. Like I asked an LM to summarize

0:28:42.160 --> 0:28:44.760
<v Speaker 2>my research, and it told me that I had done

0:28:44.760 --> 0:28:46.520
<v Speaker 2>stuff that I had not done and that none of

0:28:46.520 --> 0:28:49.360
<v Speaker 2>my collaborators have done in the system. And if I

0:28:49.400 --> 0:28:51.440
<v Speaker 2>had just asked for the summary, I would have been like, Oh,

0:28:51.480 --> 0:28:52.960
<v Speaker 2>we know a lot of stuff about the system that

0:28:53.000 --> 0:28:55.000
<v Speaker 2>we don't really know, and so I guess anything that

0:28:55.040 --> 0:28:57.680
<v Speaker 2>you decide could be vaguely interesting. You need to actually

0:28:57.720 --> 0:29:00.760
<v Speaker 2>go back, oh yes, and confirm aps. It's frustrating because

0:29:00.760 --> 0:29:01.920
<v Speaker 2>you might be like, oh, I didn't know that. I'm

0:29:01.920 --> 0:29:03.560
<v Speaker 2>not going to use it, but I didn't know that,

0:29:03.800 --> 0:29:05.600
<v Speaker 2>and you might not necessarily go back and check, but

0:29:05.640 --> 0:29:07.800
<v Speaker 2>you just will think like, oh, I'll just keep that

0:29:07.840 --> 0:29:09.480
<v Speaker 2>in the back of my mind. But now you've got

0:29:09.480 --> 0:29:10.960
<v Speaker 2>a little lie in the back of your mind.

0:29:13.080 --> 0:29:13.240
<v Speaker 4>Well.

0:29:13.280 --> 0:29:16.360
<v Speaker 1>I think of it like a superpowered Google search. I

0:29:16.360 --> 0:29:19.440
<v Speaker 1>cannot go into Google and say, here's my research. Has

0:29:19.440 --> 0:29:22.000
<v Speaker 1>there been any recent papers that might be relevant? But

0:29:22.040 --> 0:29:24.040
<v Speaker 1>I can't type that into my LM. And you can

0:29:24.040 --> 0:29:25.720
<v Speaker 1>find a bunch of papers, and then I can read

0:29:25.760 --> 0:29:28.360
<v Speaker 1>those instead of reading all of the papers, and I

0:29:28.360 --> 0:29:30.440
<v Speaker 1>can decide, oh, this is not actually irrelevant, this is

0:29:30.440 --> 0:29:32.720
<v Speaker 1>a crucial detail here you missed, or hey, that paper

0:29:32.760 --> 0:29:35.440
<v Speaker 1>doesn't exist, so it's not useful. But it's doing the

0:29:35.480 --> 0:29:37.720
<v Speaker 1>part that is hard for me, which is just like

0:29:37.880 --> 0:29:40.920
<v Speaker 1>reading all of those papers and it might miss something,

0:29:41.000 --> 0:29:44.320
<v Speaker 1>but it's very very powerful to tackling the problem that

0:29:44.400 --> 0:29:47.400
<v Speaker 1>I can't solve myself, which is reading all of those papers.

0:29:47.800 --> 0:29:51.080
<v Speaker 1>And of course there's highly varying quality between the models.

0:29:51.120 --> 0:29:52.560
<v Speaker 1>Some of them are good at this, some of them

0:29:52.560 --> 0:29:54.880
<v Speaker 1>are bad at this. And this is the kind of

0:29:54.920 --> 0:29:57.720
<v Speaker 1>thing that's improving very rapidly with time. It used to

0:29:57.760 --> 0:29:59.840
<v Speaker 1>be you type this into an LLLM and you got

0:30:00.080 --> 0:30:03.480
<v Speaker 1>lots of nonsense. These days, I find it's producing higher

0:30:03.560 --> 0:30:07.120
<v Speaker 1>and higher quality output insights into papers that I didn't

0:30:07.120 --> 0:30:11.080
<v Speaker 1>know about, really useful stuff. And you see also in mathematics,

0:30:11.120 --> 0:30:13.600
<v Speaker 1>an area i'm not an expert in that it's finding

0:30:13.640 --> 0:30:16.960
<v Speaker 1>solutions to outstanding problems that already exist in the literature

0:30:16.960 --> 0:30:18.960
<v Speaker 1>and nobody knew. They're like, oh, this nineteen fifty seven

0:30:18.960 --> 0:30:21.800
<v Speaker 1>paper by some Russian, nobody's translated it, so we didn't

0:30:21.880 --> 0:30:24.440
<v Speaker 1>know this open problem has actually been solved.

0:30:24.880 --> 0:30:27.560
<v Speaker 2>When I think it's solving airdish problems that hadn't been

0:30:27.560 --> 0:30:28.440
<v Speaker 2>solved yet.

0:30:28.440 --> 0:30:31.720
<v Speaker 1>Exactly, sometimes by finding techniques in other papers that are

0:30:31.760 --> 0:30:36.000
<v Speaker 1>relevant already. I'm inventing those techniques itself, but making those connections,

0:30:36.040 --> 0:30:39.200
<v Speaker 1>because that's what it's good at, scanning through vast quantities

0:30:39.200 --> 0:30:42.960
<v Speaker 1>of literature for things that are relevant. Okay, so so

0:30:43.040 --> 0:30:45.200
<v Speaker 1>far it's been working as an assistant. Right, it reads

0:30:45.200 --> 0:30:47.600
<v Speaker 1>a documentation, it gets the software to work, or it

0:30:47.640 --> 0:30:50.480
<v Speaker 1>goes to find papers. It's like a research tool. What

0:30:50.640 --> 0:30:53.880
<v Speaker 1>about like more as the role of a physicist can.

0:30:53.960 --> 0:30:57.560
<v Speaker 1>For example, you give it a bunch of ideas and say, hey,

0:30:57.640 --> 0:31:00.000
<v Speaker 1>go write a paper about this, or write me up

0:31:00.160 --> 0:31:03.920
<v Speaker 1>grant proposal on this topic. Right, So I've tried this

0:31:04.000 --> 0:31:06.120
<v Speaker 1>kind of thing. I have access to the top end

0:31:06.200 --> 0:31:10.400
<v Speaker 1>models from anthropic for example, and it can do it right.

0:31:10.480 --> 0:31:13.360
<v Speaker 1>It can start from rough ideas, it can flesh it

0:31:13.400 --> 0:31:17.160
<v Speaker 1>out into full text. It's grammatically perfect, it makes no

0:31:17.280 --> 0:31:20.960
<v Speaker 1>spelling mistakes. But in my experience, it's writing is not

0:31:21.160 --> 0:31:24.160
<v Speaker 1>very good. I'm a big fan of good writing. Like

0:31:24.200 --> 0:31:26.640
<v Speaker 1>when I read a paper, there's sort of two dimensions there.

0:31:26.680 --> 0:31:29.600
<v Speaker 1>There's the science good, but also like is the writing good?

0:31:30.040 --> 0:31:31.840
<v Speaker 1>And if the writing is good, it leads you on

0:31:31.920 --> 0:31:34.840
<v Speaker 1>a story. It gets you interested, It tells you why

0:31:34.840 --> 0:31:39.000
<v Speaker 1>this research is compelling. It presents the things logically and crisply.

0:31:39.400 --> 0:31:42.560
<v Speaker 1>The text is concise and precise, you know, like a

0:31:42.640 --> 0:31:45.840
<v Speaker 1>well written scientific paper is a piece of beauty. And

0:31:45.880 --> 0:31:49.000
<v Speaker 1>there's also like sloppily written good science. Sometimes the science

0:31:49.080 --> 0:31:51.000
<v Speaker 1>is great and the paper is just like sloppy, and

0:31:51.040 --> 0:31:53.720
<v Speaker 1>you're like, what do you mean by this exactly? You know,

0:31:53.760 --> 0:31:55.320
<v Speaker 1>and like, well you're telling me this, but now we're

0:31:55.320 --> 0:31:58.280
<v Speaker 1>talking about this, why is that's just not well put together?

0:31:58.760 --> 0:32:02.480
<v Speaker 1>KLAW doesn't produce sloppy stuff, but it produces boring stuff.

0:32:03.040 --> 0:32:05.400
<v Speaker 1>I find the text that it produces is missing, like

0:32:05.760 --> 0:32:08.000
<v Speaker 1>you know that juj that you get from like a

0:32:08.040 --> 0:32:10.800
<v Speaker 1>really good piece of writing, And this is the same

0:32:10.880 --> 0:32:13.320
<v Speaker 1>I think kind of skill that it takes to write

0:32:13.360 --> 0:32:16.600
<v Speaker 1>like a good novel. Like I read a lot. When

0:32:16.640 --> 0:32:19.320
<v Speaker 1>I started a novel, I can tell instantly like Okay,

0:32:19.520 --> 0:32:21.800
<v Speaker 1>this dude knows how to do it, or this lady

0:32:21.920 --> 0:32:24.480
<v Speaker 1>can write, Oh my gosh, I don't even know what

0:32:24.560 --> 0:32:26.960
<v Speaker 1>the plot is or the characters are, but the sentence

0:32:27.000 --> 0:32:30.680
<v Speaker 1>structure something about it is just like powerful and effective

0:32:31.360 --> 0:32:34.520
<v Speaker 1>and that's a joy to read, right, And LMS cannot

0:32:34.520 --> 0:32:36.000
<v Speaker 1>produce that, not yet at least.

0:32:36.120 --> 0:32:38.000
<v Speaker 2>So I'm going to make a book recommendation. But I'm

0:32:38.000 --> 0:32:39.800
<v Speaker 2>going to say that even though this title sounds like

0:32:39.840 --> 0:32:42.640
<v Speaker 2>a kid's book, this is not a book for kids.

0:32:43.000 --> 0:32:45.600
<v Speaker 2>The book is called Benny and the Blue Whale, and

0:32:45.640 --> 0:32:49.440
<v Speaker 2>it is a fiction writer who collaborated with AI to

0:32:49.480 --> 0:32:51.240
<v Speaker 2>write a book just to sort of see what the

0:32:51.280 --> 0:32:55.959
<v Speaker 2>experience was like. And the experience matches what you were saying.

0:32:56.400 --> 0:32:59.080
<v Speaker 2>It's not quite as creative, I think, but anyway, it

0:32:59.160 --> 0:33:00.600
<v Speaker 2>was a fun a fun reading.

0:33:00.800 --> 0:33:02.880
<v Speaker 1>And there's something about, you know, the voice and the

0:33:02.920 --> 0:33:06.360
<v Speaker 1>personality of the author that comes out in their writing.

0:33:06.760 --> 0:33:09.320
<v Speaker 1>And that's also important for science. If you write a

0:33:09.320 --> 0:33:12.360
<v Speaker 1>grand proposal and it's fun and it's exciting, and it's

0:33:12.360 --> 0:33:14.640
<v Speaker 1>well written and it's crisp, the reviewers are going to

0:33:14.680 --> 0:33:16.560
<v Speaker 1>have fun reading it. They're going to get excited. They're

0:33:16.600 --> 0:33:18.240
<v Speaker 1>going to like you, they're going to want to fund

0:33:18.280 --> 0:33:21.640
<v Speaker 1>your work. It's underappreciated, I think, by the public. How

0:33:21.720 --> 0:33:24.600
<v Speaker 1>much of science is writing, not just for folks like

0:33:24.640 --> 0:33:27.560
<v Speaker 1>you and me who'd like write specifically for the public,

0:33:27.920 --> 0:33:32.080
<v Speaker 1>but like emails, papers, grand proposals, reviews, all this stuff

0:33:32.160 --> 0:33:35.800
<v Speaker 1>is writing, and so much more powerful when it's done well,

0:33:36.040 --> 0:33:38.200
<v Speaker 1>when you have a flair and energy, when you can

0:33:38.240 --> 0:33:41.960
<v Speaker 1>convey excitement and enthusiasts when you have your own personal voice.

0:33:42.080 --> 0:33:43.960
<v Speaker 2>Yeah. And I think we should note though, that a

0:33:43.960 --> 0:33:46.680
<v Speaker 2>lot of journals and a lot of grant agencies currently

0:33:46.720 --> 0:33:48.680
<v Speaker 2>have rules for whether or not you're allowed to have

0:33:48.760 --> 0:33:51.720
<v Speaker 2>AI help you with your writing. I just think it's

0:33:51.720 --> 0:33:52.320
<v Speaker 2>worth noting that.

0:33:52.480 --> 0:33:55.719
<v Speaker 1>Yeah. Absolutely, And so it can produce sort of like

0:33:56.040 --> 0:34:00.960
<v Speaker 1>bland generic writing that seems like suspiciously polished and kind

0:34:01.040 --> 0:34:03.880
<v Speaker 1>of flat, you know, doesn't really seem to get the

0:34:03.960 --> 0:34:07.280
<v Speaker 1>point and convey it crisply. But you know, this again,

0:34:07.400 --> 0:34:09.520
<v Speaker 1>is a steep gradient. They used to be worse, they're

0:34:09.560 --> 0:34:12.919
<v Speaker 1>getting better. It's rapidly improving. I don't know that next

0:34:13.000 --> 0:34:15.719
<v Speaker 1>year these things won't be much better. And you know

0:34:15.760 --> 0:34:17.880
<v Speaker 1>a lot of the way they're improving is just by scaling.

0:34:17.960 --> 0:34:21.240
<v Speaker 1>They read more. They make these models bigger, more layers

0:34:21.239 --> 0:34:25.600
<v Speaker 1>of attention, more nodes. You sometimes hear these parameters quoted

0:34:25.680 --> 0:34:29.520
<v Speaker 1>like seven billion parameters or whatever. That's the number of

0:34:29.680 --> 0:34:32.600
<v Speaker 1>like connections between neurons, each of which have a number

0:34:32.640 --> 0:34:35.880
<v Speaker 1>associated with them. Tells you essentially the size of the network.

0:34:36.239 --> 0:34:38.520
<v Speaker 1>The bigger the network, the more powerful it is, also

0:34:38.600 --> 0:34:40.799
<v Speaker 1>the bigger the training set it needs to figure out

0:34:40.960 --> 0:34:43.839
<v Speaker 1>what is the best value for all those nodes? All right?

0:34:43.880 --> 0:34:47.319
<v Speaker 1>So lms have progressed from like fancy Google searches to

0:34:47.440 --> 0:34:51.279
<v Speaker 1>helping you with code, to reading and writing. But what

0:34:51.400 --> 0:34:53.760
<v Speaker 1>about the juice of it, right, what about like doing

0:34:53.840 --> 0:34:58.120
<v Speaker 1>calculations actually being a physicist. So there's some fascinating recent

0:34:58.239 --> 0:35:03.080
<v Speaker 1>examples of top level physicists using AI to contribute to

0:35:03.120 --> 0:35:03.680
<v Speaker 1>their papers.

0:35:03.960 --> 0:35:24.719
<v Speaker 2>And we're going to get to that after the break,

0:35:26.120 --> 0:35:28.319
<v Speaker 2>and we're back and we're talking about how AI is

0:35:28.360 --> 0:35:32.680
<v Speaker 2>helping scientists do some math. All right, take it away, Daniel.

0:35:32.960 --> 0:35:35.640
<v Speaker 1>So there are top level physicists out there at the

0:35:35.640 --> 0:35:39.520
<v Speaker 1>top of American academic physics who are using and exploring

0:35:39.560 --> 0:35:42.800
<v Speaker 1>AI to work with them to help them solve problems.

0:35:43.719 --> 0:35:46.600
<v Speaker 1>At Witten, for example, you know famous string theorists, maybe

0:35:46.640 --> 0:35:49.160
<v Speaker 1>one of the smartest dudes alive, put out a paper

0:35:49.200 --> 0:35:52.720
<v Speaker 1>recently where chat GPT did an important part of the work.

0:35:53.040 --> 0:35:55.080
<v Speaker 1>They had worked out a bunch of examples and they

0:35:55.080 --> 0:35:57.480
<v Speaker 1>had an idea that it could be more generally proven,

0:35:57.880 --> 0:36:00.279
<v Speaker 1>and chat GPT came in and like found a way

0:36:00.320 --> 0:36:02.960
<v Speaker 1>to make those connections. Say okay, here's the proof that

0:36:03.000 --> 0:36:05.560
<v Speaker 1>this really does work. So you know it didn't come

0:36:05.640 --> 0:36:08.040
<v Speaker 1>up with the research question, and it didn't even guess

0:36:08.040 --> 0:36:10.040
<v Speaker 1>at the answer. It came in and filled in a

0:36:10.040 --> 0:36:12.440
<v Speaker 1>lot of the details. But this is, you know, more

0:36:12.480 --> 0:36:15.880
<v Speaker 1>than just like correcting your grammar or fixing your code.

0:36:16.120 --> 0:36:20.160
<v Speaker 1>This is like manipulating mathematics and using logic to build

0:36:20.200 --> 0:36:22.640
<v Speaker 1>a proof from A to B. So it really is

0:36:22.640 --> 0:36:23.440
<v Speaker 1>pretty impressive.

0:36:23.800 --> 0:36:27.480
<v Speaker 2>Yeah, that's amazing. I'm sure that saved doctor Whitten a

0:36:27.480 --> 0:36:29.680
<v Speaker 2>lot of time. Yeah, and allowed him to work on

0:36:29.760 --> 0:36:31.160
<v Speaker 2>other projects exactly.

0:36:31.239 --> 0:36:34.360
<v Speaker 1>And you know, it's doing symbolic calculations, not just like

0:36:34.600 --> 0:36:37.040
<v Speaker 1>you know, seven point two plus twelve point four or whatever.

0:36:37.239 --> 0:36:41.280
<v Speaker 1>This is manipulating symbols, finding equations that capture these things.

0:36:41.800 --> 0:36:45.120
<v Speaker 1>So that's very powerful. Then a friend of mine, Matt

0:36:45.160 --> 0:36:48.279
<v Speaker 1>Schwartz at Harvard, recently put out of paper where he

0:36:48.400 --> 0:36:51.960
<v Speaker 1>used claude Opus. Essentially, as a graduate student, he had

0:36:51.960 --> 0:36:55.000
<v Speaker 1>an idea for a project, he pitched it to Claude,

0:36:55.120 --> 0:36:56.600
<v Speaker 1>he got claud to work on it with him, and

0:36:56.640 --> 0:36:58.960
<v Speaker 1>Claude did like all of the coding and making all

0:36:59.000 --> 0:37:03.040
<v Speaker 1>the plots and and simulations and all the actual work involved,

0:37:03.080 --> 0:37:05.080
<v Speaker 1>and he guided it through it. And it's not like

0:37:05.120 --> 0:37:07.160
<v Speaker 1>he just said, hey, here's an idea, Go do the

0:37:07.160 --> 0:37:09.239
<v Speaker 1>work and write me a paper. And then it came

0:37:09.280 --> 0:37:12.120
<v Speaker 1>back after lunch and it was all done. His experience

0:37:12.239 --> 0:37:15.360
<v Speaker 1>was that it's sort of like a very naive young

0:37:15.400 --> 0:37:19.120
<v Speaker 1>graduate student. It's often getting lost in the weeds. It's

0:37:19.160 --> 0:37:22.560
<v Speaker 1>not thinking about the bigger picture. It sometimes makes mistakes,

0:37:22.560 --> 0:37:25.279
<v Speaker 1>it needs a lot of supervision. But that's sort of

0:37:25.280 --> 0:37:29.000
<v Speaker 1>focusing on the negatives, right, Like, it's amazing, This is incredible,

0:37:29.160 --> 0:37:33.240
<v Speaker 1>right that you have something which is a capable research assistant. Again,

0:37:33.320 --> 0:37:36.319
<v Speaker 1>not just a search tool anymore, right, but actually doing

0:37:36.320 --> 0:37:39.440
<v Speaker 1>work and making progress and making it so that a

0:37:39.480 --> 0:37:41.840
<v Speaker 1>guy who has an idea for a paper no longer

0:37:41.920 --> 0:37:44.320
<v Speaker 1>has to go out and find somebody to do the work.

0:37:44.520 --> 0:37:47.160
<v Speaker 1>He can find AI to do the work. He still

0:37:47.160 --> 0:37:49.440
<v Speaker 1>needs to engage his brain, he still came up with

0:37:49.440 --> 0:37:51.359
<v Speaker 1>a question, he still needs to follow it and make

0:37:51.400 --> 0:37:54.640
<v Speaker 1>sure it's reasonable. But this is not something AI could

0:37:54.640 --> 0:37:57.160
<v Speaker 1>do a year ago or two years ago. So it's

0:37:57.200 --> 0:37:58.720
<v Speaker 1>a dramatic improvement.

0:37:58.520 --> 0:38:01.919
<v Speaker 2>That's exciting but also h a bit scary for folks

0:38:01.920 --> 0:38:04.080
<v Speaker 2>who would like to get positions as graduate students.

0:38:04.280 --> 0:38:07.080
<v Speaker 1>Yes, exactly. And I saw him go a talk recently

0:38:07.120 --> 0:38:09.560
<v Speaker 1>and he said, AI is going to keep getting better

0:38:09.640 --> 0:38:12.319
<v Speaker 1>and better at this, and we are not. And so

0:38:12.680 --> 0:38:18.919
<v Speaker 1>his forecast is by twenty thirty most science becomes AI assisted. Wow,

0:38:19.160 --> 0:38:22.480
<v Speaker 1>and that soon we'll see end to end AI doing

0:38:22.520 --> 0:38:25.799
<v Speaker 1>theoretical science that we won't need this kind of detailed

0:38:25.840 --> 0:38:28.440
<v Speaker 1>step by step handholding and guidance.

0:38:29.000 --> 0:38:32.160
<v Speaker 2>Does end to end include AI coming up with its

0:38:32.200 --> 0:38:33.040
<v Speaker 2>own questions.

0:38:33.520 --> 0:38:36.040
<v Speaker 1>It's not clear to me what exactly he means by that,

0:38:36.080 --> 0:38:39.399
<v Speaker 1>And I think that's a very important distinction, because being

0:38:39.400 --> 0:38:42.480
<v Speaker 1>able to say, here's an interesting problem, go solve it

0:38:42.520 --> 0:38:46.160
<v Speaker 1>is very different from saying, what's an interesting problem we

0:38:46.200 --> 0:38:46.800
<v Speaker 1>should work on?

0:38:47.200 --> 0:38:49.000
<v Speaker 2>Yeah, because I feel like that's when you work with

0:38:49.040 --> 0:38:51.719
<v Speaker 2>a grad student, you know, there you get them to

0:38:51.760 --> 0:38:54.480
<v Speaker 2>the point where they tell you, I'm familiar with the literature,

0:38:54.560 --> 0:38:56.600
<v Speaker 2>here's what the interesting question is, and you're like, oh,

0:38:56.640 --> 0:38:58.880
<v Speaker 2>that's great. I didn't think I hadn't thought of that question.

0:38:58.960 --> 0:39:02.479
<v Speaker 2>And now like, now you're a team, will AI become

0:39:02.480 --> 0:39:03.320
<v Speaker 2>part of the team.

0:39:03.719 --> 0:39:05.799
<v Speaker 1>So I was on this panel at Neurops, one of

0:39:05.800 --> 0:39:08.160
<v Speaker 1>the premier AI conferences, and I was kind of a

0:39:08.200 --> 0:39:12.240
<v Speaker 1>wet blanket about anthropics solving physics in the next few years.

0:39:12.640 --> 0:39:14.960
<v Speaker 1>So folks from the company approached me afterwards and they

0:39:14.960 --> 0:39:16.719
<v Speaker 1>were like, so, Daniel, what's it going to take to

0:39:16.719 --> 0:39:18.960
<v Speaker 1>convent to you? And I said, I don't know, but

0:39:19.000 --> 0:39:21.120
<v Speaker 1>I'd love to play with your top level models. So

0:39:21.120 --> 0:39:23.080
<v Speaker 1>they give me access to their top level models for

0:39:23.120 --> 0:39:25.600
<v Speaker 1>a few months, and so I played with it, and

0:39:25.719 --> 0:39:28.319
<v Speaker 1>I asked it exactly what you suggested? I said, all right,

0:39:28.360 --> 0:39:31.480
<v Speaker 1>Ai suggest a project. Here's the kind of thing I'm

0:39:31.480 --> 0:39:34.440
<v Speaker 1>interested in. What's an opportunity, right, because I think a

0:39:34.440 --> 0:39:37.400
<v Speaker 1>lot of people don't appreciate how much of science is

0:39:37.800 --> 0:39:41.399
<v Speaker 1>identifying openings. It's a lot like being an entrepreneur. Right,

0:39:41.719 --> 0:39:44.799
<v Speaker 1>you have limited resources, you have limited time. Find a

0:39:44.840 --> 0:39:48.680
<v Speaker 1>place where you can make progress given your resources in

0:39:48.680 --> 0:39:51.640
<v Speaker 1>a reasonable amount of time. That's going to be impactful. Right.

0:39:51.719 --> 0:39:54.359
<v Speaker 1>That's a lot of what being a professional scientist is.

0:39:54.600 --> 0:39:57.000
<v Speaker 1>It's easy to come up with like big questions about

0:39:57.040 --> 0:39:59.800
<v Speaker 1>the universe or grandiose plans that would take a thousand

0:39:59.840 --> 0:40:02.480
<v Speaker 1>year and a zillion dollars. But what can I actually

0:40:02.480 --> 0:40:05.800
<v Speaker 1>do today over the next year to produce some science

0:40:05.840 --> 0:40:08.680
<v Speaker 1>that's interesting to the community. What's the current conversation and

0:40:08.719 --> 0:40:12.120
<v Speaker 1>where can we contribute? So I asked Ai what should

0:40:12.120 --> 0:40:15.040
<v Speaker 1>we do? And it read my profile and then it

0:40:15.080 --> 0:40:18.360
<v Speaker 1>suggested a few ideas, all of which were things I

0:40:18.440 --> 0:40:20.040
<v Speaker 1>had already done in the last few years.

0:40:21.160 --> 0:40:22.520
<v Speaker 2>Okay, so it's not there yet.

0:40:22.760 --> 0:40:24.560
<v Speaker 1>So I was kind of disappointed. I was like, hmm,

0:40:24.800 --> 0:40:27.160
<v Speaker 1>I've already done all of these things, and I didn't

0:40:27.200 --> 0:40:29.680
<v Speaker 1>know if it's just like, not quite there yet and

0:40:29.760 --> 0:40:31.359
<v Speaker 1>I'm a little bit ahead of it, or if it's

0:40:31.360 --> 0:40:33.560
<v Speaker 1>just like had read my papers and I was like, oh,

0:40:33.640 --> 0:40:36.480
<v Speaker 1>Daniel's interested in machine learning and particle physics and tracking,

0:40:36.680 --> 0:40:39.280
<v Speaker 1>so I'll suggest something in that direction. It's not clear

0:40:39.280 --> 0:40:41.359
<v Speaker 1>to me whether it's in the training sample or not.

0:40:42.000 --> 0:40:44.600
<v Speaker 1>And so it's not there yet, right, But that doesn't

0:40:44.640 --> 0:40:46.840
<v Speaker 1>mean that it can't contribute. Right, It's not going to

0:40:46.840 --> 0:40:50.320
<v Speaker 1>be an autonomous physicist just yet. But remember, let's think about

0:40:50.360 --> 0:40:52.919
<v Speaker 1>what it's good at. What can it do. It's good

0:40:53.040 --> 0:40:55.600
<v Speaker 1>at scanning lots of literature, it's good at putting these

0:40:55.640 --> 0:40:59.000
<v Speaker 1>things together. And remember that a lot of advances we

0:40:59.080 --> 0:41:01.920
<v Speaker 1>make in physics are not just like one person has

0:41:01.960 --> 0:41:04.640
<v Speaker 1>a genius idea which leaps us forward one hundred years.

0:41:04.680 --> 0:41:09.319
<v Speaker 1>It's often putting things together from existing tools. Right, And

0:41:09.400 --> 0:41:12.759
<v Speaker 1>so now let's dig into the question that we asked

0:41:12.800 --> 0:41:15.640
<v Speaker 1>at the top of the episode. This claim made by

0:41:15.680 --> 0:41:20.160
<v Speaker 1>these anthropic AI hypesters that it's going to solve physics, right,

0:41:20.200 --> 0:41:22.799
<v Speaker 1>and like again, this is what they were actually saying,

0:41:22.800 --> 0:41:24.720
<v Speaker 1>And this is the kind of thing you see AI

0:41:24.800 --> 0:41:29.000
<v Speaker 1>tech bros saying. You know, Sam Altman said, quote, although

0:41:29.040 --> 0:41:33.920
<v Speaker 1>it will happen incrementally astounding triumphs, fixing the climate, establishing

0:41:34.000 --> 0:41:37.520
<v Speaker 1>a space colony, and the discovery of all of physics

0:41:37.640 --> 0:41:41.000
<v Speaker 1>will eventually become commonplace. So I'm not putting words in

0:41:41.040 --> 0:41:43.919
<v Speaker 1>their mouths, right, these are the claims that are being made.

0:41:44.239 --> 0:41:44.479
<v Speaker 3>Wow.

0:41:44.760 --> 0:41:47.440
<v Speaker 2>Yeah, because of AI, we will have a space colony.

0:41:48.320 --> 0:41:51.040
<v Speaker 1>That's right. We're going to fix the climate, establish a

0:41:51.040 --> 0:41:53.640
<v Speaker 1>space colony, care cancer dot dot dot dot dot. Wow,

0:41:53.680 --> 0:41:56.080
<v Speaker 1>all of this kind of stuff amazing, you know. And

0:41:56.200 --> 0:41:58.120
<v Speaker 1>you see also folks on the other side. One of

0:41:58.200 --> 0:42:01.200
<v Speaker 1>my favorite writers is Corey Doctor Rowe, and he says

0:42:01.280 --> 0:42:03.759
<v Speaker 1>that saying that AI is going to do science and

0:42:03.800 --> 0:42:06.319
<v Speaker 1>gain consciousness is like saying, if we can get a

0:42:06.320 --> 0:42:08.200
<v Speaker 1>horse to run fast enough, it's going to turn into

0:42:08.200 --> 0:42:10.880
<v Speaker 1>a locomotive, you know. And his point is that you

0:42:10.880 --> 0:42:14.359
<v Speaker 1>can't just like extrapolate blindly right, because like, what are

0:42:14.360 --> 0:42:16.880
<v Speaker 1>we even talking about, you know, on that panel and

0:42:16.920 --> 0:42:19.440
<v Speaker 1>lots of times people are talking about how much progress

0:42:19.480 --> 0:42:21.759
<v Speaker 1>AI is making. We have to think about, like what

0:42:21.880 --> 0:42:25.040
<v Speaker 1>is the scale there? What are we measuring its capability

0:42:25.080 --> 0:42:28.360
<v Speaker 1>to do? Is it like percentage of an Einstein or something.

0:42:28.400 --> 0:42:31.480
<v Speaker 1>It's not something that we really know how to measure

0:42:31.520 --> 0:42:34.000
<v Speaker 1>where we can say like by twenty forty seven it's

0:42:34.040 --> 0:42:37.520
<v Speaker 1>going to be two Einstein's right, that's all just nonsense,

0:42:37.560 --> 0:42:40.080
<v Speaker 1>and so we don't know if there's a barrier there.

0:42:40.160 --> 0:42:41.640
<v Speaker 1>We don't know if it's going to be able to

0:42:41.640 --> 0:42:44.000
<v Speaker 1>do this. And what they're talking about it doing is

0:42:44.000 --> 0:42:48.120
<v Speaker 1>something it has never done. Right, It's never suggested a

0:42:48.160 --> 0:42:50.560
<v Speaker 1>new research topic or solve the whole problem.

0:42:50.800 --> 0:42:52.280
<v Speaker 2>But that doesn't mean it can't.

0:42:52.360 --> 0:42:55.160
<v Speaker 1>Though, that doesn't mean it can't. Right, maybe these horses

0:42:55.239 --> 0:42:58.200
<v Speaker 1>will turn into locomotives. Right. It certainly is gaining a

0:42:58.239 --> 0:43:00.960
<v Speaker 1>lot of power every year. So let's talk about the

0:43:00.960 --> 0:43:04.080
<v Speaker 1>thing they're claiming, right, what does it mean to solve physics?

0:43:04.640 --> 0:43:08.919
<v Speaker 1>In my mind, there's two possible answers to that. One is, look,

0:43:08.960 --> 0:43:11.120
<v Speaker 1>we have a list of problems in physics right now.

0:43:11.160 --> 0:43:13.000
<v Speaker 1>We don't know what dark matter is. We don't know

0:43:13.040 --> 0:43:16.000
<v Speaker 1>why the universes mostly matter. We don't understand dark energy.

0:43:16.200 --> 0:43:19.560
<v Speaker 1>We can't unify quantum mechanics and general relativity. We don't

0:43:19.560 --> 0:43:21.839
<v Speaker 1>know why there's so many particles and forces, right, how

0:43:21.880 --> 0:43:23.880
<v Speaker 1>does it all fit together? There are definitely things we

0:43:23.920 --> 0:43:29.880
<v Speaker 1>don't understand. Maybe solving physics means answering today's open physics questions.

0:43:30.040 --> 0:43:33.040
<v Speaker 1>I mean in the minds of same Altman and folks.

0:43:33.120 --> 0:43:35.840
<v Speaker 1>Maybe that's what they mean when they say this solving

0:43:35.880 --> 0:43:38.839
<v Speaker 1>all of physics. To me, that's not solving physics. Even

0:43:38.840 --> 0:43:42.719
<v Speaker 1>if tomorrow they came out with the answers to these questions, right,

0:43:42.760 --> 0:43:44.920
<v Speaker 1>I would not be done. No, I would not be like,

0:43:45.040 --> 0:43:46.799
<v Speaker 1>well I'm bored now, I'm just going to go sit

0:43:46.840 --> 0:43:49.200
<v Speaker 1>on the beach. Right, I'm going to ask like, well,

0:43:49.320 --> 0:43:51.600
<v Speaker 1>why is this the solution? Could it have been this

0:43:51.680 --> 0:43:54.560
<v Speaker 1>other way? Can I imagine places this could break down?

0:43:55.160 --> 0:43:57.920
<v Speaker 1>Could have been a different way? You know, there's always

0:43:58.040 --> 0:44:02.040
<v Speaker 1>questions to ask. These questions come from the philosophical context,

0:44:02.040 --> 0:44:05.160
<v Speaker 1>which is the origin of our curiosity, and that's not

0:44:05.239 --> 0:44:08.480
<v Speaker 1>going to go away even if we get beautiful equations

0:44:08.520 --> 0:44:11.320
<v Speaker 1>to solve today's questions. In physics when.

0:44:11.160 --> 0:44:13.360
<v Speaker 2>I feel like, if you've talked to any scientist, you

0:44:13.440 --> 0:44:15.680
<v Speaker 2>know that as soon as you get one answer, that

0:44:15.880 --> 0:44:20.200
<v Speaker 2>automatically leads to like twenty new questions. Like it's exponential

0:44:20.239 --> 0:44:22.680
<v Speaker 2>the questions you get when you get one answer, and

0:44:22.719 --> 0:44:25.279
<v Speaker 2>so it's hard for me to imagine that solving those

0:44:25.320 --> 0:44:29.360
<v Speaker 2>problems would suddenly mean ah, we're done.

0:44:29.880 --> 0:44:32.560
<v Speaker 1>I totally agree. And so from that perspective, will it

0:44:32.760 --> 0:44:36.799
<v Speaker 1>solve physics? Absolutely not. Physics is a human pursuit and

0:44:36.880 --> 0:44:39.799
<v Speaker 1>it's not going to end even if AI makes our

0:44:40.040 --> 0:44:43.279
<v Speaker 1>answer finding much more powerful. But I do think, and

0:44:43.320 --> 0:44:45.840
<v Speaker 1>this might surprise people, that it is going to transform

0:44:45.880 --> 0:44:48.600
<v Speaker 1>the way we find those answers, and then it might

0:44:48.680 --> 0:44:53.239
<v Speaker 1>actually be capable of Einstein level contributions. We don't know

0:44:53.280 --> 0:44:56.440
<v Speaker 1>how to measure einsteininess, but when you look back in

0:44:56.480 --> 0:44:59.000
<v Speaker 1>the history of a lot of the big leaps, what

0:44:59.120 --> 0:45:03.120
<v Speaker 1>are they doing. They are putting together existing pieces that

0:45:03.200 --> 0:45:06.160
<v Speaker 1>other people were not aware of. It's easy to say

0:45:06.160 --> 0:45:08.200
<v Speaker 1>Einstein was a super genius he came up with this

0:45:08.200 --> 0:45:10.520
<v Speaker 1>whole thing himself, but that's not really the truth. He

0:45:10.680 --> 0:45:14.239
<v Speaker 1>was building on generations of careful math done by other

0:45:14.320 --> 0:45:17.480
<v Speaker 1>nerds not even interested in general relativity. You know, grassman

0:45:17.600 --> 0:45:20.600
<v Speaker 1>numbers and romani and manifolds. All these things are essential

0:45:20.640 --> 0:45:23.560
<v Speaker 1>for Einstein. He could not solve those problems himself. He

0:45:23.680 --> 0:45:27.080
<v Speaker 1>put together a bunch of interesting pieces, but sometimes just

0:45:27.120 --> 0:45:29.120
<v Speaker 1>being lucky by who he knew and who he met,

0:45:29.239 --> 0:45:30.960
<v Speaker 1>being the right place at the right time to meet

0:45:31.000 --> 0:45:34.960
<v Speaker 1>these folks and learned about their progress. So that's the

0:45:35.040 --> 0:45:37.120
<v Speaker 1>kind of thing AI is good at, is like, I'm

0:45:37.120 --> 0:45:39.319
<v Speaker 1>going to read all these papers, I'm going to notice, Oh,

0:45:39.360 --> 0:45:41.440
<v Speaker 1>this tool over here that there has been developed for

0:45:41.480 --> 0:45:43.960
<v Speaker 1>totally other reasons might be able to solve this problem

0:45:44.000 --> 0:45:46.360
<v Speaker 1>you guys are working on. That's exactly the kind of

0:45:46.360 --> 0:45:49.719
<v Speaker 1>thing that AI is really good at is making these connections.

0:45:49.760 --> 0:45:53.439
<v Speaker 1>It's searching broadly, is reading the whole literature. Right, there's

0:45:53.440 --> 0:45:56.880
<v Speaker 1>lots of examples of this group theory invented for totally

0:45:56.920 --> 0:45:59.800
<v Speaker 1>other reasons by math nerds turns out to be essential

0:46:00.080 --> 0:46:03.839
<v Speaker 1>or quantum field theory. So I actually am optimistic that

0:46:03.920 --> 0:46:06.400
<v Speaker 1>there could be big problems out there in physics we

0:46:06.440 --> 0:46:09.520
<v Speaker 1>don't today have solutions for that AI could help us

0:46:09.680 --> 0:46:13.000
<v Speaker 1>rapidly solve because of its specific capabilities.

0:46:13.520 --> 0:46:17.359
<v Speaker 2>Okay, so you've just been pretty wildly optimistic about how

0:46:17.600 --> 0:46:20.279
<v Speaker 2>un characteristically optimistic? That's right, That's right. You know you

0:46:20.600 --> 0:46:24.080
<v Speaker 2>might be the more optimistic of team what blanket? But

0:46:24.960 --> 0:46:27.480
<v Speaker 2>so do you think, like, well AI ever get like

0:46:27.560 --> 0:46:28.799
<v Speaker 2>a Nobel Prize?

0:46:29.160 --> 0:46:31.920
<v Speaker 1>Yeah, it's hard to know how to answer that. I mean,

0:46:32.560 --> 0:46:36.640
<v Speaker 1>computer scientists building AI to solve problems have won Nobel

0:46:36.680 --> 0:46:39.720
<v Speaker 1>prizes already, right, we saw that last year, Like Google

0:46:39.760 --> 0:46:42.880
<v Speaker 1>deep Mind won a Nobel Prize. Right, I mean Alpha

0:46:42.920 --> 0:46:46.360
<v Speaker 1>fold has done incredible stuff, right, They solved protein folding,

0:46:47.000 --> 0:46:49.560
<v Speaker 1>and so the computer scientists wanted for the tool they

0:46:49.600 --> 0:46:52.239
<v Speaker 1>built to solve the problem. Are we in an era

0:46:52.360 --> 0:46:54.480
<v Speaker 1>where AI could win a Nobel Prize? I mean, like,

0:46:54.680 --> 0:46:56.959
<v Speaker 1>I don't know what the estate lawyers would say after

0:46:57.000 --> 0:47:01.160
<v Speaker 1>reading Nobel's will again, but I think we are on

0:47:01.200 --> 0:47:05.200
<v Speaker 1>the verge of AI solving the current generation of hard

0:47:05.239 --> 0:47:08.120
<v Speaker 1>theory problems. It's still going to be like a human

0:47:08.360 --> 0:47:11.880
<v Speaker 1>asks an AI, like, hey, can you unify quantum mechanics

0:47:11.880 --> 0:47:14.360
<v Speaker 1>and general relativity? But if it goes off and thinks

0:47:14.360 --> 0:47:16.720
<v Speaker 1>hard and finds a piece of math we hadn't explored

0:47:16.840 --> 0:47:20.360
<v Speaker 1>or solved some of the current problems in quantum gravity,

0:47:20.440 --> 0:47:23.360
<v Speaker 1>comes up with a novel theory for it, then certainly

0:47:23.760 --> 0:47:27.080
<v Speaker 1>it's going to have made an enormous contribution to our

0:47:27.200 --> 0:47:30.879
<v Speaker 1>understanding of the universe, to physics. Whether you should give

0:47:30.880 --> 0:47:32.440
<v Speaker 1>it the credit or the credit for the person who

0:47:32.520 --> 0:47:34.879
<v Speaker 1>typed in the prompt I think that's a harder question

0:47:35.000 --> 0:47:36.240
<v Speaker 1>I'm not qualified to answer.

0:47:36.360 --> 0:47:39.840
<v Speaker 2>Wow. Yeah, and so many people contributed to making AI,

0:47:40.080 --> 0:47:42.719
<v Speaker 2>it would be hard to know who to give the

0:47:42.760 --> 0:47:43.200
<v Speaker 2>credit to.

0:47:43.719 --> 0:47:46.520
<v Speaker 1>Yeah, that's always the case though, Like you know, for

0:47:46.560 --> 0:47:48.840
<v Speaker 1>the Higgs boson, they give it to the theorists, not

0:47:48.920 --> 0:47:52.239
<v Speaker 1>to the experimentalists who discovered it, partially because there's just

0:47:52.280 --> 0:47:54.719
<v Speaker 1>too many of us, Right, thousands of people contributed to

0:47:54.920 --> 0:47:57.720
<v Speaker 1>building the accelerator and the detector and making it work,

0:47:58.000 --> 0:48:00.640
<v Speaker 1>so we can't all win the Nobel Prizes limit there.

0:48:00.800 --> 0:48:03.879
<v Speaker 1>But I think it's clear that AI can do the

0:48:03.920 --> 0:48:07.600
<v Speaker 1>thing we need to make big conceptual leaps, construct the

0:48:07.640 --> 0:48:11.080
<v Speaker 1>building blocks, pull those building blocks together to make breakthroughs.

0:48:11.480 --> 0:48:14.120
<v Speaker 1>That's what it can do, and that's very powerful. Right,

0:48:14.120 --> 0:48:16.279
<v Speaker 1>We've seen in history that's how a lot of these

0:48:16.320 --> 0:48:17.720
<v Speaker 1>big leaps are made.

0:48:17.960 --> 0:48:21.920
<v Speaker 2>Let's talk about it's impact on like physicists today. So

0:48:22.000 --> 0:48:25.920
<v Speaker 2>like for example, it can't take over the LHC and

0:48:25.960 --> 0:48:29.640
<v Speaker 2>collect the data for you, like run the day Today stuff.

0:48:29.680 --> 0:48:32.640
<v Speaker 2>So you still need experimentalists to collect the data.

0:48:32.640 --> 0:48:36.319
<v Speaker 1>Right still for a while, Yes, But there's a lot

0:48:36.360 --> 0:48:40.280
<v Speaker 1>of work in autonomous labs. You know, you've given AI

0:48:40.600 --> 0:48:43.200
<v Speaker 1>access to the LIEDC, it could certainly decide, you know,

0:48:43.320 --> 0:48:45.600
<v Speaker 1>how to run it and what data to keep, et cetera.

0:48:46.160 --> 0:48:49.759
<v Speaker 1>And people are building these autonomous labs where an AI can,

0:48:49.760 --> 0:48:53.120
<v Speaker 1>for example, test a bunch of materials really quickly, you know,

0:48:53.200 --> 0:48:56.120
<v Speaker 1>build a bunch of materials, test it, and then close

0:48:56.160 --> 0:48:58.160
<v Speaker 1>the loop come up with a new material to test.

0:48:58.680 --> 0:49:01.560
<v Speaker 1>I was involved in a proposal last year for something

0:49:01.600 --> 0:49:05.200
<v Speaker 1>called text to launch technology, where you type into an

0:49:05.320 --> 0:49:08.480
<v Speaker 1>LM like okay, I want a satellite that can do xyz,

0:49:09.160 --> 0:49:11.799
<v Speaker 1>and then it uses an autonomous lab to try to

0:49:11.840 --> 0:49:15.160
<v Speaker 1>develop materials that could do that, and then far down

0:49:15.200 --> 0:49:17.560
<v Speaker 1>the road, maybe it would even launch it into space.

0:49:17.760 --> 0:49:20.880
<v Speaker 2>Wow, oh my gosh, that's nuts.

0:49:21.280 --> 0:49:24.160
<v Speaker 1>But the point is that this is very new currently,

0:49:24.200 --> 0:49:26.759
<v Speaker 1>but there's no real barrier there. If you can give

0:49:26.800 --> 0:49:29.040
<v Speaker 1>AI control of these labs, if you build a lab

0:49:29.080 --> 0:49:32.960
<v Speaker 1>that's AI compatible, there's no reason why experiments have to

0:49:32.960 --> 0:49:34.600
<v Speaker 1>be out of reach of AI.

0:49:35.239 --> 0:49:38.920
<v Speaker 2>Okay, and so then how is it impacting funding? So,

0:49:39.000 --> 0:49:42.720
<v Speaker 2>like you mentioned that AI could replace a grad student,

0:49:43.560 --> 0:49:45.440
<v Speaker 2>so maybe now you don't have to ask for funding

0:49:45.480 --> 0:49:52.280
<v Speaker 2>for grad students anymore. But like, could AI replace Daniel Whitson? Never, I imagine,

0:49:52.360 --> 0:49:53.960
<v Speaker 2>But but what do you think?

0:49:54.520 --> 0:49:57.720
<v Speaker 1>AI can now do the work of a green graduate student?

0:49:58.000 --> 0:50:00.760
<v Speaker 1>And that's very powerful. I can get more stuff done

0:50:00.880 --> 0:50:04.760
<v Speaker 1>without my students than I could before. That is helpful.

0:50:05.120 --> 0:50:07.720
<v Speaker 1>I always have one project where I'm doing the actual coding,

0:50:08.040 --> 0:50:11.440
<v Speaker 1>and now those projects move faster, so that is very cool.

0:50:11.560 --> 0:50:15.160
<v Speaker 1>And it's definitely changing the way the government sees particle

0:50:15.160 --> 0:50:18.040
<v Speaker 1>physics being done. In some of the latest budget plans,

0:50:18.239 --> 0:50:21.120
<v Speaker 1>hundreds of millions of dollars that used to go to

0:50:21.560 --> 0:50:24.960
<v Speaker 1>pure particle physics is now being sent to AI for

0:50:25.080 --> 0:50:29.120
<v Speaker 1>science that includes particle physics, and particle physicists are still involved,

0:50:29.400 --> 0:50:32.560
<v Speaker 1>but industry is also very deeply involved. It looks to

0:50:32.560 --> 0:50:34.680
<v Speaker 1>me like the government wants us to make all of

0:50:34.680 --> 0:50:38.400
<v Speaker 1>our data AI friendly so that these big models from

0:50:38.480 --> 0:50:41.719
<v Speaker 1>open AI and from Anthropic can play key roles in

0:50:41.840 --> 0:50:45.520
<v Speaker 1>accelerating the progress in physics. I'm not against that, I

0:50:45.600 --> 0:50:49.120
<v Speaker 1>just wish that it was new money instead of repurposed

0:50:49.120 --> 0:50:51.800
<v Speaker 1>old money, like let's keep doing particle physics and let's

0:50:51.920 --> 0:50:56.560
<v Speaker 1>add an investment on AI instead of cannibalizing particle physics

0:50:56.600 --> 0:50:59.920
<v Speaker 1>for this money and sending a few hundred million dollars

0:51:00.320 --> 0:51:04.680
<v Speaker 1>to these vast corporations. Remember that, like academic physics is

0:51:04.920 --> 0:51:08.040
<v Speaker 1>tiny in terms of research budgets compared to what these

0:51:08.040 --> 0:51:11.000
<v Speaker 1>companies have and what they are doing, and so a

0:51:11.120 --> 0:51:13.640
<v Speaker 1>small amount of money for them is vast. Funding for

0:51:13.719 --> 0:51:16.360
<v Speaker 1>us can have a huge impact. And let's not forget

0:51:16.360 --> 0:51:19.360
<v Speaker 1>that all of the ideas at the foundation of the

0:51:19.400 --> 0:51:24.160
<v Speaker 1>current wave of AI came from academia, Academics nerding around,

0:51:24.239 --> 0:51:28.320
<v Speaker 1>coming up with ideas, attention, deep learning, all these things

0:51:28.440 --> 0:51:32.000
<v Speaker 1>came from academic labs that we're just playing with stuff.

0:51:32.000 --> 0:51:34.120
<v Speaker 1>And so I'm going to beat the drum one more

0:51:34.160 --> 0:51:37.799
<v Speaker 1>time for investing in basic research because it pays off

0:51:37.920 --> 0:51:40.919
<v Speaker 1>huge and when it pays off, we shouldn't then cannibalize

0:51:40.960 --> 0:51:44.960
<v Speaker 1>our basic research funding to dig into AI. We should

0:51:45.000 --> 0:51:48.040
<v Speaker 1>also invest in AI, but we should maintain our investments

0:51:48.080 --> 0:51:49.760
<v Speaker 1>in basic research, right.

0:51:49.840 --> 0:51:52.080
<v Speaker 2>Yeah, yeah. And I'm under the impression that if you

0:51:52.360 --> 0:51:54.600
<v Speaker 2>massively cut a budget for i don't know, say five

0:51:54.719 --> 0:51:57.799
<v Speaker 2>or ten years, you really hurt a generation of scientists.

0:51:57.800 --> 0:52:00.160
<v Speaker 2>You know, like labs are going to hire few regrets ud,

0:52:00.200 --> 0:52:03.600
<v Speaker 2>it's fewer post stocs, you're going to start, fewer physics labs,

0:52:03.600 --> 0:52:05.360
<v Speaker 2>and the labs that do get started aren't going to

0:52:05.400 --> 0:52:08.200
<v Speaker 2>get the grants that they need to convince their university

0:52:08.239 --> 0:52:10.360
<v Speaker 2>that they should get tenures so that they should stick around.

0:52:10.400 --> 0:52:13.040
<v Speaker 2>And now you've got, you know, fewer physicists produced by

0:52:13.040 --> 0:52:16.200
<v Speaker 2>the United States. Is that your sense also? Is that

0:52:16.280 --> 0:52:17.680
<v Speaker 2>not how things work in area? Okay?

0:52:17.800 --> 0:52:21.560
<v Speaker 1>Yeah, no, absolutely, Like the theory community has received massive

0:52:21.600 --> 0:52:24.520
<v Speaker 1>cuts to its funding, and theory is very, very cheap.

0:52:24.600 --> 0:52:27.040
<v Speaker 1>You know, you're just paying a few people and pencil

0:52:27.080 --> 0:52:30.760
<v Speaker 1>and paper, but their budget's been cut like seventy eighty percent,

0:52:31.280 --> 0:52:34.760
<v Speaker 1>which means, you know, many fewer students, which means fewer theorists,

0:52:34.760 --> 0:52:38.359
<v Speaker 1>which means fewer smart people asking those questions. And you

0:52:38.360 --> 0:52:40.120
<v Speaker 1>asked a question I've been answered yet, which is like

0:52:40.280 --> 0:52:44.040
<v Speaker 1>could this replace Daniel? And like, I think that it's

0:52:44.120 --> 0:52:47.040
<v Speaker 1>not hard to imagine that you soon have AIS that

0:52:47.160 --> 0:52:51.360
<v Speaker 1>have the capability of like senior scientists to do research

0:52:51.719 --> 0:52:55.680
<v Speaker 1>to even drive other lower level AIS or supervise humans

0:52:55.719 --> 0:52:58.239
<v Speaker 1>who are doing research. I don't think that's out of

0:52:58.280 --> 0:53:02.960
<v Speaker 1>their capability. But something I don't understand is who's asking

0:53:03.000 --> 0:53:06.920
<v Speaker 1>the questions? Right, Physics is a human endeavor. At the

0:53:06.960 --> 0:53:09.799
<v Speaker 1>heart of it, is AI going to solve? Physics is

0:53:09.800 --> 0:53:12.520
<v Speaker 1>about whether we are ever done asking questions, and the

0:53:12.560 --> 0:53:15.759
<v Speaker 1>answer to that it's always gonna be no. It's just

0:53:15.880 --> 0:53:19.719
<v Speaker 1>going to be no. Physics is about our curiosity about

0:53:19.719 --> 0:53:22.800
<v Speaker 1>the universe. Now, that doesn't mean that machines can't be

0:53:22.960 --> 0:53:26.360
<v Speaker 1>there along with us answering some of the questions. The

0:53:26.440 --> 0:53:30.279
<v Speaker 1>open question to me is will they have their own questions?

0:53:30.640 --> 0:53:33.320
<v Speaker 1>I've got my questions about the universe? You have yours.

0:53:33.480 --> 0:53:35.839
<v Speaker 1>That's why I'm in physics and you're in parasites. And

0:53:35.880 --> 0:53:38.239
<v Speaker 1>somebody else is doing art history, and somebody else is

0:53:38.280 --> 0:53:42.480
<v Speaker 1>doing chemistry, you know, God save them, and somebody else

0:53:42.560 --> 0:53:44.520
<v Speaker 1>is at some company inventing new ways to make white

0:53:44.560 --> 0:53:47.319
<v Speaker 1>chocolate cheap. And you know, everybody's driven by their own

0:53:47.360 --> 0:53:53.120
<v Speaker 1>personal curiosity. Do machines have curiosity? If you give Claude

0:53:53.200 --> 0:53:55.640
<v Speaker 1>all the physics papers and you ask it like, what

0:53:55.800 --> 0:53:58.600
<v Speaker 1>do you think is interesting? What is it interested in?

0:53:58.719 --> 0:54:01.080
<v Speaker 1>And is that really its own an interest? Right? Or

0:54:01.160 --> 0:54:04.800
<v Speaker 1>is it just following up on what humans have begun?

0:54:05.480 --> 0:54:08.759
<v Speaker 1>And so, to me, that's the interesting philosophical question. When

0:54:08.840 --> 0:54:11.480
<v Speaker 1>they get to have that level of capacity, are they

0:54:11.480 --> 0:54:14.360
<v Speaker 1>answering their questions or ours? Will they be similar? Is

0:54:14.400 --> 0:54:17.120
<v Speaker 1>this an example of aliens doing physics?

0:54:17.200 --> 0:54:17.399
<v Speaker 3>Right?

0:54:17.800 --> 0:54:21.160
<v Speaker 1>Will machines do physics? I don't know, because I don't

0:54:21.160 --> 0:54:23.480
<v Speaker 1>know what it means to have an AI at that level,

0:54:23.640 --> 0:54:26.800
<v Speaker 1>And they will never really be divorced from humanity because

0:54:26.840 --> 0:54:30.719
<v Speaker 1>they were birthed from the collective intelligence, right the way

0:54:30.719 --> 0:54:32.920
<v Speaker 1>that like something that's learned to write a great novel

0:54:33.280 --> 0:54:35.279
<v Speaker 1>has learned to write that great novel by reading a

0:54:35.280 --> 0:54:37.479
<v Speaker 1>bunch of great novels. But that's also how we train

0:54:37.680 --> 0:54:41.080
<v Speaker 1>great novelists that are humans. Right, You read good writing,

0:54:41.120 --> 0:54:45.120
<v Speaker 1>you're inspired, It teaches you. So, you know, I don't

0:54:45.120 --> 0:54:47.720
<v Speaker 1>think it's impossible for AI to solve the current generation

0:54:47.800 --> 0:54:50.360
<v Speaker 1>of physics problems, and I don't think it's impossible to

0:54:50.440 --> 0:54:53.040
<v Speaker 1>potentially come up with new questions on its own.

0:54:53.440 --> 0:54:56.239
<v Speaker 2>And if AI came up with a list of ten questions,

0:54:56.360 --> 0:54:59.160
<v Speaker 2>and then you were presented with those ten questions alongside

0:54:59.160 --> 0:55:02.279
<v Speaker 2>of ten questions from some of your favorite collaborators, and

0:55:02.320 --> 0:55:06.480
<v Speaker 2>you thought those questions were equally interesting. What would what

0:55:06.520 --> 0:55:07.240
<v Speaker 2>would you say?

0:55:08.040 --> 0:55:13.320
<v Speaker 1>I'd say, bravo. You know, I welcome our future AI collaborators. Okay,

0:55:13.400 --> 0:55:16.360
<v Speaker 1>bring your curiosity wherever it comes from. I think it

0:55:16.440 --> 0:55:19.080
<v Speaker 1>can just help inform. But I think we should invest

0:55:19.080 --> 0:55:20.920
<v Speaker 1>in this stuff, you know, and we should invest in

0:55:21.400 --> 0:55:25.480
<v Speaker 1>human driven curiosity as well as machine driven curiosity and

0:55:25.600 --> 0:55:27.960
<v Speaker 1>alien driven curiosity and all of it.

0:55:28.880 --> 0:55:30.680
<v Speaker 2>That's right, invest in our future.

0:55:32.000 --> 0:55:32.480
<v Speaker 1>Exactly.

0:55:32.760 --> 0:55:35.560
<v Speaker 2>All right, Well that was fascinating. Thank you Daniel for

0:55:35.600 --> 0:55:38.160
<v Speaker 2>walking us through how you've been working with physics and

0:55:38.200 --> 0:55:40.279
<v Speaker 2>AI over the last What do you think like, I

0:55:40.280 --> 0:55:42.640
<v Speaker 2>guess it's been decades. You've been working with LMS.

0:55:42.920 --> 0:55:45.680
<v Speaker 1>I've been doing machine learning and particle physics since the nineties.

0:55:46.320 --> 0:55:48.920
<v Speaker 1>So yeah, it's been thirty years. Okay before it was

0:55:48.960 --> 0:55:49.640
<v Speaker 1>cool for sure?

0:55:50.280 --> 0:55:52.640
<v Speaker 2>All right, Well, maybe you'll give us a yearly update

0:55:52.680 --> 0:55:54.240
<v Speaker 2>on how this kind of stuff is going.

0:55:54.480 --> 0:55:57.319
<v Speaker 1>All right? Or maybe AI Daniel will replace me.

0:55:57.719 --> 0:56:00.359
<v Speaker 2>I doubt it. I can't imagine that his Joe will

0:56:00.400 --> 0:56:04.000
<v Speaker 2>be anywhere near as good or as bad, depending on

0:56:04.040 --> 0:56:06.040
<v Speaker 2>how you see it as yours.

0:56:06.440 --> 0:56:08.520
<v Speaker 1>The day I stopped making fun of chemistry is the day,

0:56:08.560 --> 0:56:10.360
<v Speaker 1>you know, I've been replaced by a robot a.

0:56:12.160 --> 0:56:14.279
<v Speaker 2>Thanks so much to Daniel, and thanks so much to

0:56:14.320 --> 0:56:17.040
<v Speaker 2>the extraordinaries. And if you have a question about anything

0:56:17.040 --> 0:56:19.000
<v Speaker 2>at all, you can send it to us at questions

0:56:19.040 --> 0:56:20.799
<v Speaker 2>at danieland Kelly dot org.

0:56:21.239 --> 0:56:23.239
<v Speaker 1>You'll get an answer from a real human being.

0:56:23.520 --> 0:56:30.720
<v Speaker 2>Yes you will.

0:56:30.760 --> 0:56:33.200
<v Speaker 1>Thanks everybody for listening. Please go and do us a

0:56:33.200 --> 0:56:36.480
<v Speaker 1>favor and rate the show on whatever podcast app you're using.

0:56:36.560 --> 0:56:38.120
<v Speaker 1>It really helps people find us.

0:56:38.680 --> 0:56:42.600
<v Speaker 2>Daniel and Kelly's Extraordinary Universe is edited by the amazing

0:56:42.600 --> 0:56:43.360
<v Speaker 2>Matt Kesselman.

0:56:43.560 --> 0:56:46.839
<v Speaker 1>He really is a wizard. You can also find us

0:56:46.920 --> 0:56:52.000
<v Speaker 1>online on Blue Sky, Instagram and x D and K Universe.

0:56:52.080 --> 0:56:53.279
<v Speaker 1>Come in cage with us.

0:56:53.520 --> 0:56:56.879
<v Speaker 2>You can email us at questions at Danielankelly dot org.

0:56:56.960 --> 0:56:58.520
<v Speaker 2>We really do want to hear.

0:56:58.440 --> 0:57:01.120
<v Speaker 1>From you, and you can find our website www dot

0:57:01.200 --> 0:57:05.080
<v Speaker 1>danieland Kelly dot org, where you'll also find an invitation

0:57:05.160 --> 0:57:08.560
<v Speaker 1>to join our discord where everybody comes and talks about

0:57:08.600 --> 0:57:09.840
<v Speaker 1>the amazing universe.

0:57:10.120 --> 0:57:14.080
<v Speaker 2>And we also have the most amazing moderators. This is

0:57:14.120 --> 0:57:16.680
<v Speaker 2>an iHeart podcast. Thanks for joining us.