Latest / Logixly Live Unsupervised Learning Podcast / OpenAI's Math Breakthrough

Transcript

OpenAI's Math Breakthrough · Logixly Live Unsupervised Learning Podcast

• 0:00
  Speaker 2: Connects directly to some of the stuff we're talking about today.
• 0:02
  Speaker 2: So welcome everybody to Logixly Live, uh, where we discuss all the big stories in AI or try to.
• 0:08
  Speaker 2: Uh, we have uh, of course, Ted and myself, Matt, and Katie's back again two weeks ago.
• 0:13
  Speaker 2: We had Ben in.
• 0:13
  Speaker 2: Yep.
• 0:14
  Speaker 2: Right?
• 0:14
  Speaker 2: Yep.
• 0:14
  Speaker 2: Yep.
• 0:15
  Speaker 2: When we got Katie back today.
• 0:16
  Speaker 2: So, topic today is on OpenAI released a little bit of a press release or or some some news, I guess, about how they've uh made some serious progress on a number of math problems.
• 0:30
  Speaker 2: Um, so I I don't know, Katie, if you wouldn't mind given the very high level on kind of what they did and why it's important, and then we'll move pretty quickly.
• 0:39
  Speaker 2: We're not gonna stick around and and go all as much as we'd like to, maybe all that deep on the exact math.
• 0:45
  Speaker 2: But I think it's interesting to talk about why it is interesting that these are the problems that are getting solved and kind of relate that to what I just talked about with Excel formulas and what we see in in sort of the enterprise in general.
• 0:59
  Speaker 1: Well sure before you answer that.
• 1:01
  Speaker 1: Yeah.
• 1:01
  Speaker 1: I know OpenAI put out a press release, you're saying.
• 1:05
  Speaker 1: Yes.
• 1:06
  Speaker 1: I think NLP Logixly put out a press release about being an open AI select partner.
• 1:11
  Speaker 1: So pretty cool stuff.
• 1:13
  Speaker 2: Very cool stuff.
• 1:14
  Speaker 2: And honestly, related to this too.
• 1:16
  Speaker 2: So the the details there are uh yeah, open AI.
• 1:19
  Speaker 2: I mean, I feel like this has been a theme these last few weeks of the dichotomy of what you see, where the dichotomy today is you have open AI making real progress on these hard math problems, and yet at the same time, open AI needs partners to help integrate and build these solutions, like connect these models to business problems, not math problems, right, but like business problems.
• 1:45
  Speaker 2: So that's the partnership.
• 1:46
  Speaker 2: We're proud to be select partners, and we expect it to you know have a great long relationship with the people.
• 1:52
  Speaker 1: That's so cool to watch.
• 1:54
  Speaker 1: We've been through so many hype cycles in the last 15 years, and the latest one is oh, open AI and throttle, you know, all the different uh frontier labs are gonna be able to solve everything, but in the end, you still need humans.
• 2:08
  Speaker 1: You gotta connect them and you gotta test and make sure that it um everything, you know, all the outputs are are accurate, and that's really cool that our business model of integrating models into business flows, all of a sudden there's there's this great demand and to be able to uh qualify as a select partner is cool.
• 2:25
  Speaker 1: You were in the middle of the whole process.
• 2:26
  Speaker 1: Tell us about how that worked.
• 2:28
  Speaker 2: 100%.
• 2:28
  Speaker 2: Yeah, all the the parts on the process that I did think were interesting, and certainly it's open AI, so they're leveraging AI behind the scenes to evaluate the the applicants.
• 2:37
  Speaker 2: But the parts on the application that I loved were um there were two different sections.
• 2:43
  Speaker 2: One to section, one section was to talk about 10 times, up to 10 times, where your organization has successfully integrated AI into and solve business problems.
• 2:56
  Speaker 2: And that was a layup.
• 2:57
  Speaker 2: I was I was wishing there were 20.
• 3:02
  Speaker: Yeah.
• 3:02
  Speaker 2: So I mean that part was kind of cool, right?
• 3:04
  Speaker 2: Because it was like it was um, you know, leveraging, you know, the 15 years and stuff, and and um a lot of great, a lot of great stuff.
• 3:11
  Speaker 2: And honestly, our our application it went through pretty quickly, I think it was because we could stand on the work we've done for so long.
• 3:17
  Speaker 2: But before we get to the math part, Ted, the second thought of all say is you know, Anthropic yesterday, so it's not just open AI.
• 3:24
  Speaker 2: So Anthropic has committed a hundred million to train 10,000 engineers to deploy claw inside businesses by the end of 2027.
• 3:33
  Speaker 2: Wow.
• 3:33
  Speaker 2: So I mean, you know, the clearly while you hear about AGI and automation and solving every math problems, we don't need mathematicians.
• 3:42
  Speaker 2: I just it feels like it's a very different story in the enterprise and maybe the world we live in.
• 3:48
  Speaker: Right.
• 3:48
  Speaker: Well, and I mean I love it from the argument.
• 3:50
  Speaker: Like, if you if I tell somebody I work in AI, I get one of two arguments.
• 3:55
  Speaker: A, it's cool, and then I don't know, me and that person get to nerd out for a little while, or they go, Yeah, you're stealing everyone's job.
• 4:01
  Speaker: And it's like, well, you know, I don't till fields either, right?
• 4:05
  Speaker: Like, you know, the industrial revolution happened, like those things happen.
• 4:09
  Speaker: So yeah, there's jobs going away, but like, look, this is a company investing a hundred million in making new jobs.
• 4:15
  Speaker: There are new things needed.
• 4:16
  Speaker: Yeah, we just happen to be doing the thing that they're trying to build, right?
• 4:21
  Speaker: Like it's just something that we've been doing for a while, but like, yeah, there's new jobs coming in.
• 4:26
  Speaker: Like, it's not AI is coming to destroy the world, right?
• 4:29
  Speaker: It is AI is coming and yeah, there's new stuff.
• 4:31
  Speaker: We're gonna have to learn new skills.
• 4:33
  Speaker 2: Okay, so then let's connect that directly to the the start.
• 4:36
  Speaker 2: Well, so uh and I I do want you to at some point explain a little bit of the the nerdy stuff, but but on the impact of uh, you know, taking jobs away and stuff, the there's a second part to the story, which is the mathematicians are mad because um and and I I don't know if that I don't want to mischaracterize this, but I I feel like there's a sentiment around um you you know, I I guess there was a coalition of mathematicians that got together with OpenAI, and this is reported from Wired, and we'll include this in the notes in the show so you can go and read about this yourself.
• 5:12
  Speaker 2: But um, but but basically OpenAI got together the mathematicians, and I guess there was this understanding that OpenAI would pace out some of this research and not drop it all.
• 5:24
  Speaker 2: And I feel like that's that that is the mathematician saying, Hey, don't take our problems.
• 5:30
  Speaker 2: We want to work on these problems, or how do you interpret that?
• 5:34
  Speaker: But like, okay, I think you do have to look at it from a curve standpoint.
• 5:37
  Speaker: And you know, three data points does not, you know, a relationship make, but it was six papers in August and then a hundred in September, and then what were 732, I think.
• 5:49
  Speaker: Uh let's see.
• 5:50
  Speaker 2: So 722 manuscript across 372 result families is what they do.
• 5:57
  Speaker: So if I'm looking at uh, you know, we're looking at a little bit of exponential growth.
• 6:00
  Speaker: We're already hitting the hockey stick on on that release, going from six to a hundred to seven hundred and twenty, right?
• 6:07
  Speaker: Like you're seeing that come out super quickly.
• 6:10
  Speaker: But I think I don't know, mathematicians, like you're gonna be mad, you know, and I think everybody wants to get grumpy about it, but I don't, you know, when you look at the problems that are getting solved, yeah, there they are sort of some of the age-old problems, and like, yeah, everybody wants to solve the age-old puzzle, like that's pretty awesome.
• 6:29
  Speaker: But there's plenty of research that comes out of new, like, here's new conjectures, and like the thing I haven't seen is hey, open AI came up with new conjectures.
• 6:38
  Speaker 2: But what what is better for, and maybe this is I'm not I don't want to take like this altruistic kind of like Nirvana, but like what is better for everybody to solve these open math problems or to set them aside and save them so they're solved by people.
• 6:55
  Speaker 2: Like, is it not kind of it's it's weird, right?
• 6:57
  Speaker 2: Like the point of the point of it, we want them solved.
• 7:01
  Speaker 2: We don't want them solved conditionally.
• 7:04
  Speaker: Okay, but there's this thing.
• 7:05
  Speaker: Okay, so uh we're gonna get we're gonna get really nerdy here.
• 7:08
  Speaker: There's two two perspectives.
• 7:10
  Speaker: We want them solved.
• 7:11
  Speaker 2: Okay, right?
• 7:11
  Speaker: But okay, so this is gonna go way back.
• 7:13
  Speaker: Like Paul Urdish is the most prolific mathematician of modern times, right?
• 7:18
  Speaker: And he What did he do?
• 7:20
  Speaker: Uh he published thousands of papers.
• 7:22
  Speaker: There is like you know how there's the bacon number?
• 7:25
  Speaker: They were great.
• 7:26
  Speaker: Okay, but so the bacon number is like uh Kevin Bacon.
• 7:30
  Speaker: Oh you know, you've got six degrees of Kevin Bacon.
• 7:32
  Speaker: Okay, mathematicians do the same thing.
• 7:35
  Speaker 2: Okay, I saw Kevin Bacon one time walking down the street in New York.
• 7:38
  Speaker: Sweet now, my bacon number's too, because like I've seen mathematics.
• 7:40
  Speaker: He doesn't know him, but he doesn't just I'm taking the credit back.
• 7:44
  Speaker: Come on.
• 7:45
  Speaker: But it what you also have is an Urdish number.
• 7:48
  Speaker: So like that's where math get really nerdy, right?
• 7:51
  Speaker: So you got mathematicians that are going, okay, Paul Urdish is number one.
• 7:55
  Speaker: Somebody who worked with him is gonna have this Urdish number, and then you have like connections of like who's written papers because he's that prolific.
• 8:02
  Speaker 3: Wow.
• 8:03
  Speaker: Um, so it's super interesting.
• 8:05
  Speaker: Like he's out there doing this work, but he had this thing that he was proposing called the book that is not it's not just hey, we solved this problem, but this is the most elegant way to solve it.
• 8:20
  Speaker: Like you have people, there was two teenage girls, I think, uh, in the past year made a new, like incredibly beautiful proof for the Pythagorean theorem.
• 8:32
  Speaker: That got solved, like ancient Greece wearing togas, right?
• 8:36
  Speaker: Like we don't need a new proof for the Pythagorean theorem, but we want beautiful proofs.
• 8:42
  Speaker: Like there is a thing in math.
• 8:44
  Speaker 2: Is simpler is simpler beautiful or is no?
• 8:46
  Speaker: Yeah, I mean, like, think about beautiful code, right?
• 8:48
  Speaker: Like you've seen code and you've gone that is really written anybody now.
• 8:53
  Speaker: No, I mean, like, okay, yeah, I I can write code that accomplishes a problem.
• 9:01
  Speaker 1: Right.
• 9:01
  Speaker: That's the well, I'm gonna I'm gonna reference back.
• 9:04
  Speaker: I'm gonna tease our our coworker Anton here because I wrote some code once and I got stuck on a thing and I was like, hey, can you uh come help me with this thing?
• 9:11
  Speaker: He said, Yeah, sure.
• 9:12
  Speaker: And he comes over and he goes, Yeah, just don't tell anybody I touch this.
• 9:16
  Speaker: Because I don't write beautiful code, right?
• 9:19
  Speaker 2: So same thing with math.
• 9:25
  Speaker: Not and uh not get blurred out.
• 9:27
  Speaker: But no, so it's you with math, there's beauty in like this is the best proof.
• 9:32
  Speaker: And like there, if you look at the problems that got solved in this 722, right?
• 9:38
  Speaker: You've got 20 of 20% of them are counterexample, and I think it was 15% or new bounds on a problem.
• 9:45
  Speaker: It's not actually solving the problem, it's just advancing the problem.
• 9:48
  Speaker: And I think there's probably 25% that are advancing the problem.
• 9:51
  Speaker: So 50%, like, yeah, this is a proof for a new thing.
• 9:55
  Speaker: 20% are a counterexample.
• 9:57
  Speaker: If I told Ted every apple is green and Ted throws a honey crisp at me, cool, he's proven that my my conjecture is wrong.
• 10:05
  Speaker: And you know, Ted could just throw apples at me all day long until I find a red one.
• 10:10
  Speaker: And you know, that is gonna take a whole lot of Ted work because he's got to go find all these apples.
• 10:15
  Speaker: Computers are really good at finding the apples, right?
• 10:17
  Speaker: Like when we're looking for counterexample proofs, like great, good job.
• 10:21
  Speaker: Come that is why why wouldn't you use a computer to do that?
• 10:24
  Speaker: So go and use computers and like, yeah, they are these age-old problems, but I've yet to see a an article of new conjecture, right?
• 10:34
  Speaker: Like the question I loved, I think it was Hugging Face's founder, posed can actual scientific breakthrough happen?
• 10:41
  Speaker: We're like, yeah, this is solving a thing that that's been out there versus you know, this is a brand new like Einstein was not great because he took what everyone else had done and added one step to it.
• 10:53
  Speaker: Einstein was brilliant because he thought of the things that had never been thought of before.
• 10:58
  Speaker: And AI is going, here is my history of all the things I know, and it's really good at knowing all of those things.
• 11:05
  Speaker: It doesn't know things that have never happened before, it doesn't have the training data for that.
• 11:10
  Speaker: So, like when you're looking at beautiful math, like I want to see us come up with new stuff.
• 11:16
  Speaker: Like, what are the new ideas?
• 11:18
  Speaker: Like, that's awesome.
• 11:19
  Speaker: Like, if it can prove problems, great, let it prove problems, especially like come on, these kind of like counterexample problems, let a machine do that.
• 11:28
  Speaker: Yeah, you know, and it it's everybody's gonna have their moment of like I don't want you to play in my sandbox.
• 11:34
  Speaker: Yeah, but but I think the mathematicians get over it because there's also validation, right?
• 11:40
  Speaker: When I look at those 50 problems or 50% of the problems, so we got what 350 that are it wrote a proof.
• 11:47
  Speaker: Even if you're a human, if I was to try and write a paper, I mean, God, it's been years and I would probably not do a great job, but if I try and write a math paper and you know, here's my proof of something and submit it to a journal, they're gonna bring in a mathematician and say, Can you check that?
• 12:05
  Speaker: Can you make sure?
• 12:06
  Speaker: So there's an editor, and you're gonna have multiple people that are gonna go and they're gonna be so nitpicky, and they're gonna come back and go, you're not quite there.
• 12:15
  Speaker: Try it again.
• 12:16
  Speaker 1: Didn't you just write a paper about computer vision that you're presenting?
• 12:19
  Speaker 1: We wrote a paper about dolphins.
• 12:21
  Speaker: Writing a writing a math proof.
• 12:23
  Speaker: Oh like it's it's been a hot minute, but like it gets hard.
• 12:29
  Speaker: Um you know, that's a skill that atrophies.
• 12:31
  Speaker: Side side note, just an article that one of our teammates shared with me yesterday that I think is maybe we should talk about this at some point.
• 12:41
  Speaker: It was about people doing um endoscopies, and they brought in AI to identify lesions.
• 12:50
  Speaker: And after three months, they did this test where the people that had had access to the AI had actually become less efficient at identifying the things.
• 13:02
  Speaker: So it took three months to sort of de-skill these endoscopies.
• 13:06
  Speaker 2: Okay, that's interesting.
• 13:07
  Speaker: Like so cool.
• 13:07
  Speaker 2: We should that is an interesting one.
• 13:09
  Speaker: We'll talk about that another day.
• 13:11
  Speaker 2: Yeah.
• 13:11
  Speaker: But yeah, like it's it's really, you know, like let's use AI for these cases that we can, and like let's just there's still fun, beautiful math, and there's still joy in solving problems that have already been solved in a little bit of a better way.
• 13:29
  Speaker 3: Yeah.
• 13:29
  Speaker: Like get off, get off your high horse mathematicians, scale.
• 13:33
  Speaker 2: So what what about the the the and again I love the dichotomy theme.
• 13:37
  Speaker 2: So what what about the piece of we are you know we are solving the or making progress at least on these really hard math problems, and then at the same time, we have you know, we talked about the partnership with open AI, open AI, also anthropic, like basically the recognition that there is such a gap between what well we can point, and I think that was actually kind of a cool thing.
• 14:02
  Speaker 2: They were um OpenAI was saying that on average across those 722 problems, it was three hours of agent compute time, was all it took to create the manuscripts versus you know the what is required in messy reality business to really drive a business outcome.
• 14:24
  Speaker 2: I mean, any any thoughts off top of your head on like on just the the differences there that maybe people don't think about?
• 14:30
  Speaker 1: Yeah, just it's it's always the human factor, and there's you know, you you're not gonna I don't think it's gonna be a it's gonna be a very, very long time until we're somehow programming emotional intelligence into into computers.
• 14:45
  Speaker 1: There's there's a saying if you know you can study math and the hard sciences, but if you want to study humans, you read Shakespeare.
• 14:52
  Speaker 1: Because all the different plots and the different um you know storylines and everything talks about human emotions and how humans interact.
• 14:59
  Speaker 1: And in the end, people do you know business with people and we have different personalities, we react to different things.
• 15:07
  Speaker 2: Right.
• 15:07
  Speaker 1: Um, and we run into that all the time because you're changing business processes.
• 15:12
  Speaker 1: And so now you've got the human element to it.
• 15:16
  Speaker 2: And then it's that change, like, like, and you gotta fact check me on this, Katie, but these math problems, the problem is the problem.
• 15:24
  Speaker 2: It has not changed in ever since the problem.
• 15:27
  Speaker 2: Okay, right versus like the dynamics, right?
• 15:31
  Speaker 2: On you know, your customer, but whatever, wherever, wherever you're operating, like business is constantly changing.
• 15:37
  Speaker 2: Your people change, your customers change, the industry changes, yeah.
• 15:41
  Speaker 2: And it's just this fluid dynamic, right?
• 15:44
  Speaker: Yeah, and I mean I I love that's what makes it fun.
• 15:46
  Speaker: That it right?
• 15:47
  Speaker: Yeah, that's the fun part.
• 15:48
  Speaker: That's exciting.
• 15:49
  Speaker 1: You meet new people, what what makes them tick?
• 15:51
  Speaker 1: You know, we do these predictive index and you do the Myers Briggs, and it's like it's it's so uh to me, it's yeah, that's that's the fun part because it's hard in the end, right?
• 16:01
  Speaker 2: Messy.
• 16:02
  Speaker: Yeah, messy.
• 16:03
  Speaker: The other thing I would say is look at the problems that you're solving with AI, right?
• 16:09
  Speaker: When I go and look at code to write a website, right?
• 16:12
  Speaker: And like I can't JavaScript to save my life, but you know, I could go and write code that would say, when I push this button, a cat's gonna show up, right?
• 16:23
  Speaker: And a picture of a cat comes when you push this button.
• 16:25
  Speaker: And a hundred years from now, as long as you got a way to interact with that code, when you push that button, that cat's gonna show up, right?
• 16:32
  Speaker: Like it's it is a you take an input, you get an output.
• 16:35
  Speaker: But what makes AI different?
• 16:37
  Speaker: And like this is like the hardest thing to explain to somebody because it sounds so nuanced and like pedantic to say it's different, but I have input and data, and then it goes through a process and you get an output, and data changes, and that means your output isn't static.
• 16:53
  Speaker: So I don't know what's gonna happen, and data will change, and you have to monitor stuff, and you can't go.
• 17:00
  Speaker: So, like, there's the human side of change, but there's the data side of change, and like that comes from truly changes in like what's happening.
• 17:09
  Speaker: Anybody who went through 2020, you know, unemployment really changed in 2020, and anything that was dependent on that, like, yeah, your model kind of you know took a little turn for a while, so you've gotta be aware of like economic conditions can change that, but also humans enter data and they're really bad at it, you know.
• 17:28
  Speaker: At the end of the day, somebody has to kick the process off.
• 17:31
  Speaker: Well, maybe not with the dot though.
• 17:33
  Speaker: The dot's gonna get your uh didn't even ask for it, which is which is awesome.
• 17:38
  Speaker: But like for some of these things, human has to start the process, and humans are going to do stuff wrong.
• 17:43
  Speaker: We are not great at not doing things wrong, and that you need to be able to account for that.
• 17:49
  Speaker 2: Yeah, yeah, yeah.
• 17:51
  Speaker 2: And it's a it's a you know, I think it's a dynamic, fluid system, it's not static.
• 17:55
  Speaker 2: These problems, math problems are hundreds of years old.
• 17:58
  Speaker 3: Right.
• 17:58
  Speaker 2: Business problems are you know probably not even the same as they were yesterday as the day before.
• 18:03
  Speaker 2: Like there's that constant nuance.
• 18:05
  Speaker: So from the math side, like this problems that are a hundred years old, yeah, you've got a hundred years worth of people attempting to solve this.
• 18:14
  Speaker: Right.
• 18:14
  Speaker: Right.
• 18:15
  Speaker: And if I'm a human trying to solve this problem, I'm gonna go down a few rabbit holes.
• 18:22
  Speaker: I can't like physically have read and remembered every single paper written on this topic.
• 18:28
  Speaker 3: Right.
• 18:29
  Speaker: But if you had access to the entire internet and all of that was compressed, yeah, and you had that knowledge, yeah, humans are gonna be faster.
• 18:36
  Speaker: Like for these problems that are hundreds of years old, and there is a hundred years of this is attempts that we've made to solve this, this is where we found it didn't work, it did work, here was a great, like, here's a change in bound.
• 18:47
  Speaker: Those AI gets to benefit from that information that's out there.
• 18:53
  Speaker: Yep, and that's awesome.
• 18:55
  Speaker: Yeah, like it if anybody thinks that you know, like from again, from the math side and get off your high horse, AI is standing on the shoulders of giants, it's able to do that because of a hundred years of work of mathematicians.
• 19:07
  Speaker: Yeah, I'm like that is incredible.
• 19:09
  Speaker: And I don't think open AI is going, yeah, it turns out the mathematicians are really dumb.
• 19:13
  Speaker: It's going, man, they were they were really close.
• 19:15
  Speaker: Yeah, they just needed this one final step.
• 19:17
  Speaker: Or like if we put this one and this one together, yeah, you got it.
• 19:21
  Speaker: Yeah, but the you know, everybody just wants to look at this as like, oh, we have to do it.
• 19:25
  Speaker: You cannot physically do that, right?
• 19:28
  Speaker: You cannot go through every possible counterexample.
• 19:30
  Speaker: Computers are really good at doing that.
• 19:32
  Speaker: Let the computer go.
• 19:33
  Speaker 1: We hire Katie, yeah, what she said.
• 19:35
  Speaker 1: Yeah.
• 19:37
  Speaker 1: She said, I never want to talk to anybody.
• 19:40
  Speaker: Leave me in a corner with leave me in a corner with my numbers.
• 19:43
  Speaker: You ask her a math question.
• 19:45
  Speaker: There we go.
• 19:46
  Speaker: Remember the first I do too.
• 19:48
  Speaker: Remember the first time you lied to me?
• 19:51
  Speaker: That was yeah.
• 19:54
  Speaker: Oh, that was the first time that that Ted Willich just really broke my heart.
• 19:59
  Speaker: Next day.
• 19:59
  Speaker: Hey, we need you on a sales call.
• 20:02
  Speaker 1: No, I'm because you're very good at explaining complex topics that people can consume, and that's the real talent.
• 20:09
  Speaker 1: Sometimes.
• 20:09
  Speaker 1: Well, you went with the apple and the the green sweater.
• 20:12
  Speaker 1: Oh, it's perfect.
• 20:14
  Speaker: I'm I'm trying to go with my logic stuff.
• 20:15
  Speaker: We got headshots done.
• 20:17
  Speaker: I wore this in one of my headshots because I was like, I'm gonna match our slides.
• 20:21
  Speaker 2: Oh, that's smart.
• 20:22
  Speaker 2: Anyway, all right, everybody.
• 20:26
  Speaker 2: Thank you.
• 20:26
  Speaker 2: We're gonna wrap it there for this episode of Logixly Live.
• 20:30
  Speaker 2: As always, hit us up on LinkedIn.
• 20:32
  Speaker 2: We've been doing that lately, um, making a few connections through LinkedIn.
• 20:35
  Speaker 2: So we love doing that.
• 20:36
  Speaker 2: Keep it up if there's something you want us to cover specifically.
• 20:39
  Speaker 2: Happy to do it on the next show or one off.
• 20:42
  Speaker 2: Um, and until next time, keep learning, even when it's unsupervised.