Latest / Logixly Live Unsupervised Learning Podcast / The AGI Claim Nobody Feels

Transcript

The AGI Claim Nobody Feels · Logixly Live Unsupervised Learning Podcast

• 0:00
  SPEAKER_00: I feel like they are failing the world, but um, but maybe I'm wrong.
• 0:05
  SPEAKER_00: Uh I I want to hear some things.
• 0:06
  SPEAKER_00: We're talking about it.
• 0:07
  SPEAKER_00: I want to hear some things.
• 0:08
  SPEAKER_00: We're talking about it.
• 0:09
  SPEAKER_00: So yeah, I feel like you cannot make this up.
• 0:11
  SPEAKER_00: So the three storylines that caught my eye this week or this last weekend, one is um Astra being less transparent.
• 0:23
  SPEAKER_00: So, so you know, everybody, and Ted and I were talking about this earlier this week, where it's like every time we've ever built a model for the c for our customer, invariably it always comes back at some point to the customer wanting to know, well, what happened?
• 0:37
  SPEAKER_00: Why you're getting, you know, this is the prediction.
• 0:40
  SPEAKER_00: I don't agree with it, or maybe I do agree with it, but I want to make sure it got there using reasoning that I understand.
• 0:46
  SPEAKER_00: And so Astra, you know, even according to OpenAI's model card, is now less transparent and less, you know, researchers, users are gonna have less ways and tools to understand how the model came to whatever output it came from.
• 1:06
  SPEAKER_00: So so that's the first item, just like less transparency.
• 1:09
  SPEAKER_00: I don't think anybody likes hearing that.
• 1:11
  SPEAKER_00: Second thing is AGI, so artificial general intelligence.
• 1:15
  SPEAKER_00: And I think you know, we can come back and put a good definition on that.
• 1:19
  SPEAKER_00: But one of the one of the topics around AGI has always been like when AGI is artificial general intelligence, you know, broadly means that you know, technology, AI can do things that humans can do in meaningful ways.
• 1:34
  SPEAKER_00: And I think that there's a few different definitions on that.
• 1:37
  SPEAKER_00: So the reason why I think that's interesting this week, Sam OpenAI is saying we've hit AGI, and it's I don't think most people in the world are really wanting to necessarily hear that because that's gonna come with the terror of where do the jobs go.
• 1:54
  SPEAKER_00: Um third quick thing, the third thread that that they're nailing on the PR front is the um the math award, the Millennium Award.
• 2:04
  SPEAKER_00: OpenAI had their algorithms solve one of the or maybe solve one of the millennium challenges.
• 2:10
  SPEAKER_00: We'll get into the millennium challenge, but but how they did it seems to be a little nefarious.
• 2:15
  SPEAKER_00: All right.
• 2:16
  SPEAKER_00: Um so where do you guys want to start?
• 2:18
  SPEAKER_00: We a lot of a lot of topics.
• 2:20
  SPEAKER_00: I feel like again, it's just a it's just another week of the AI PR machine just running right off the road.
• 2:27
  SPEAKER_02: But but yeah, but are they gonna tear down children's hospitals and build data centers?
• 2:33
  SPEAKER_02: I think that would be surprised.
• 2:35
  SPEAKER_00: It would not be surprising.
• 2:40
  SPEAKER_01: Yeah.
• 2:40
  SPEAKER_01: No, I think for me, like super okay, A, I love the math side of it, but before we get to that, yeah, the statement of AGI.
• 2:47
  SPEAKER_00: Okay.
• 2:48
  SPEAKER_01: Because I think we were all sort of everybody seemed to be on the same page for a while about what AGI was.
• 2:53
  SPEAKER_00: Okay.
• 2:54
  SPEAKER_00: And what is it to you?
• 2:56
  SPEAKER_01: It is able to learn, it is able to complete tasks that it was not trained to do.
• 3:00
  SPEAKER_01: You do not have to retrain the model in order for it to complete a new task.
• 3:04
  SPEAKER_00: Okay.
• 3:04
  SPEAKER_01: So it is not specifically trained for this, it is learning a new thing.
• 3:08
  SPEAKER_00: Okay.
• 3:09
  SPEAKER_01: And, you know, it's sort of like I get, I get where you would say maybe it's learning new things, but I don't know that it's learning so much as it's the training information is getting more and more rich.
• 3:20
  SPEAKER_01: We're giving it more and more detail.
• 3:22
  SPEAKER_01: And to come out and say, you know, I think as Jensen Wine came out and said, hey, I think that this is AGI because for us, that means you could pass a professional certification or do well on a standardized test.
• 3:34
  SPEAKER_00: So that to me is really so I'm curious.
• 3:36
  SPEAKER_00: In my mind, I've always had AGI, AGI as like, okay, something that can drive a car, summarize an email.
• 3:45
  SPEAKER_00: Right.
• 3:45
  SPEAKER_00: Like, like do things that that we do every day in a very autonomous fashion.
• 3:52
  SPEAKER_00: I feel like Jensen is very narrow, right?
• 3:54
  SPEAKER_00: Like you can take a test, you can get the right answer on this very structured kind of a form.
• 4:00
  SPEAKER_01: Right.
• 4:00
  SPEAKER_01: And that that feels like a real drop down from the idea of you have this intelligence where like, Matt, I'm pretty sure maybe you do play tennis.
• 4:08
  SPEAKER_01: I don't know.
• 4:08
  SPEAKER_01: But if you didn't play tennis, I bet yeah I could get you a tennis racket and in a couple days you could hit a ball pretty well.
• 4:14
  SPEAKER_00: Right.
• 4:14
  SPEAKER_01: Have a mind, and you can, you know, oh, when I hit the ball like this, it does well.
• 4:18
  SPEAKER_01: When I hit the ball like this, it doesn't, and you learn this new thing.
• 4:21
  SPEAKER_00: Yeah.
• 4:22
  SPEAKER_01: And so yeah, it you know, the idea of being able to drive a car and write an email and do all that stuff is because it it can do sort of any of these tasks, it's learning like a human.
• 4:30
  SPEAKER_01: Like to me, that's the AGI side of things.
• 4:33
  SPEAKER_02: Yeah.
• 4:33
  SPEAKER_01: Not can you do well on the SATs?
• 4:36
  SPEAKER_01: Yeah.
• 4:39
  SPEAKER_02: Is there a difference between those two?
• 4:42
  SPEAKER_01: I mean, yeah, I think like when you look at it from truly like I am learning a brand new thing, and you know, it is completely outside the scope versus like I'm learning a little bit each time I do something.
• 4:54
  SPEAKER_01: Like I don't I don't know that recursive learning necessarily you know would would get to that point of AGI for me, but it feels like that that sort of like okay, we started here and then we stepped down a little bit and then we stepped down a little bit, and now all of a sudden we're talking about passing a standardized test.
• 5:09
  SPEAKER_00: Yeah.
• 5:09
  SPEAKER_01: And it's yeah.
• 5:11
  SPEAKER_00: I mean, I do think recursive self-improvement, like I do think that that's inter super interesting because you know, if you can if you can crack that nut meaningfully, which as I understand it was the acquisition thinking behind Cursor and SpaceX, is like yeah, if you can crack that nut, it means that you can autonomously make progress in a direction without the human and research involvement.
• 5:36
  SPEAKER_00: So it's like, yeah, that's uh that's super valuable.
• 5:40
  SPEAKER_00: Uh I'm curious, and and maybe some people think that will lead you to AGI.
• 5:44
  SPEAKER_00: The thing that so so my my litmus test is if I'm gonna go back like two or three years ago when we're talking about if you said two years ago, Matt, AGI is here, I would say meaningfully, I think we live in a very different world.
• 5:59
  SPEAKER_00: Like I think I think there's so many things that would be so very different.
• 6:03
  SPEAKER_00: It just without me giving a really precise definition of what is or is an AGI, but if you tell me, like if if Sam Altman gets on the news today or he writes a blog post and he says AGI is here, like I I didn't feel it.
• 6:18
  SPEAKER_00: I don't feel like I'm in a meaningfully different.
• 6:21
  SPEAKER_00: I know we have crazy technology, but I don't feel like we're in a that different of a place to say, yeah, we have AGI because what changed?
• 6:30
  SPEAKER_00: Like if this model is AGI, do you guys feel like I have you played with it at all?
• 6:35
  SPEAKER_00: I've I've rewritten five emails today with like some abuse of model power right there.
• 6:41
  SPEAKER_01: I mean, I did the same thing.
• 6:42
  SPEAKER_01: Like, I'm like, okay, I don't know what you know, I'm not gonna go and say, you know, could you please solve a clay math problem for me?
• 6:48
  SPEAKER_01: But I am gonna go and say, you know, hey, can you create me a presentation?
• 6:51
  SPEAKER_01: I'm gonna give it a little bit less prompting than I typically would.
• 6:54
  SPEAKER_00: Yeah.
• 6:54
  SPEAKER_01: And I don't know, I mean it made a slightly nicer PowerPoint, but it's still not a PowerPoint that I would use.
• 6:59
  SPEAKER_01: Like it's still something that I would go, oh, I gotta tweak a whole lot of that.
• 7:02
  SPEAKER_00: And right.
• 7:03
  SPEAKER_01: It yeah, I don't the AGI doesn't feel like it would have to.
• 7:08
  SPEAKER_01: Yeah.
• 7:09
  SPEAKER_00: And like a world with more PowerPoints and Word documents, like I guess in my mind, when we're here with AGI, like I kind of hope for more than like more PowerPoints and more I the Jetsons, come on.
• 7:22
  SPEAKER_01: We were supposed to have robot maids and flying cars by this point.
• 7:25
  SPEAKER_02: So what was one of the other ones?
• 7:27
  SPEAKER_00: You said the millennium channel.
• 7:29
  SPEAKER_01: Yeah.
• 7:29
  SPEAKER_00: You want to I I I am uh I am so far away from that.
• 7:33
  SPEAKER_00: So, Katie, do you mind defining what the Millennium Prize is and just some some basic details?
• 7:38
  SPEAKER_01: We got seven super hard math problems that would be pretty influential if they were solved.
• 7:43
  SPEAKER_01: And they're currently unsolved, real difficult.
• 7:47
  SPEAKER_01: Said Clay Institute came out and said, we'll give you a million dollars if you can solve one of these to try and make progress.
• 7:54
  SPEAKER_01: And if I recall rightly, humans solved two, maybe three.
• 7:58
  SPEAKER_01: So like there has been some solved, but there's remaining problems.
• 8:02
  SPEAKER_00: And any idea how like over and two or three have been solved over the past X many years.
• 8:07
  SPEAKER_01: They came out with the problems in 25.
• 8:09
  SPEAKER_01: 2025.
• 8:09
  SPEAKER_01: They well the problems existed before, but they defined these are the seven.
• 8:13
  SPEAKER_01: Okay.
• 8:13
  SPEAKER_01: Okay.
• 8:13
  SPEAKER_01: And sorry, 2021.
• 8:14
  SPEAKER_01: So 25 years ago.
• 8:16
  SPEAKER_01: Um 2000.
• 8:17
  SPEAKER_01: 2001.
• 8:18
  SPEAKER_01: Got it.
• 8:19
  SPEAKER_01: There you go.
• 8:20
  SPEAKER_01: That's my map, man.
• 8:22
  SPEAKER_02: All right.
• 8:22
  SPEAKER_02: Well, that's my own clay high school account.
• 8:27
  SPEAKER_01: Um But yeah, so you've got you know, these problems that they've been around for quite some time, but we're saying, okay, hey, these are this is the challenge.
• 8:35
  SPEAKER_01: Can you solve these?
• 8:36
  SPEAKER_01: Like one of them is P equals NP.
• 8:38
  SPEAKER_01: So the question of like some problems are really easy for a computer to solve, and some problems are really, really hard for a computer to solve, and it can only use brute force.
• 8:46
  SPEAKER_01: And so those easy ones we call P, because they're solved with a polynomial.
• 8:50
  SPEAKER_01: And then the NPs are the ones that are really, really difficult to solve.
• 8:53
  SPEAKER_01: And if you could solve a non-polynomial problem with a polynomial time, like so really quickly solve those, like overnight encryption doesn't work.
• 9:04
  SPEAKER_01: And so the question is, does p equal NP?
• 9:07
  SPEAKER_01: Can there be a solution that's really fast for these really difficult problems, or can there not be?
• 9:12
  SPEAKER_01: I'm sort of in camp disprove.
• 9:14
  SPEAKER_01: I don't think that there is a you know solution for that.
• 9:18
  SPEAKER_01: Um you know, that's my whole master's was about can we use DNA?
• 9:21
  SPEAKER_01: And like, yeah, you could totally use DNA and you can do it really, really quickly, but you need an infinite amount of DNA, so kind of moot point there.
• 9:28
  SPEAKER_01: Right.
• 9:28
  SPEAKER_01: But you know, when you when you look at this, it's like these are really important to solve.
• 9:34
  SPEAKER_01: And then open AI is you know working with mathematicians and then all of a sudden um solving it themselves.
• 9:44
  SPEAKER_01: Right.
• 9:45
  SPEAKER_01: And maybe, maybe not.
• 9:46
  SPEAKER_01: Like the news is talking about it's possible that not not so clean.
• 9:51
  SPEAKER_00: So so I I think a lot of this will get probably reported and and maybe more details added, but but the the basic trajectory here is there's a couple of mathematicians that are working on one of these millennium problems that don't work at open AI.
• 10:07
  SPEAKER_00: That do not work at OpenAI.
• 10:09
  SPEAKER_00: They're using uh they are using Codex, they're using Claude, and they're they're working on on one of these problems.
• 10:18
  SPEAKER_00: It sounds like there was some uh back channel probably between mathematicians at OpenAI and and these two individuals, and OpenAI got wind that somebody was making progress on this this challenge.
• 10:34
  SPEAKER_00: And so they this is to me kind of wild.
• 10:37
  SPEAKER_00: So they they um uh as it was reported, they let loose 10,000 concurrent agents that were working on this one problem, spent roughly, if it was retail, roughly 30 million dollars in tokens and kind of front run their own customer.
• 10:54
  SPEAKER_00: So that that's even just that's like just the half of it.
• 10:58
  SPEAKER_00: So they they hear that this problem's getting advanced using through their tools, but or not.
• 11:04
  SPEAKER_00: Um, but then they they front run the progress and basically try to beat the researchers using the tools by investing a bunch of cash in it, which is pretty interesting.
• 11:14
  SPEAKER_00: But the part that's crazy is that um the uh Tristan Buckmaster, which is one of the researchers that was working on this problem, he uh is he says that there are drafts in the tools and sessions and codecs, and he's questioning whether they influenced OpenAI's efforts, as in the telemetry and prompts that those research researchers had used to work on this problem.
• 11:43
  SPEAKER_00: Did OpenAI leverage that progress and build on top of it to build their own solution?
• 11:50
  SPEAKER_00: That to me, and obviously we got to get clarity on exactly how all that goes down, but but man, even you know, and that's the thing, right?
• 11:58
  SPEAKER_00: Is even just that the reporting on that where you don't know though, is like it's another one of those, um, I don't know, like a like a black.
• 12:10
  SPEAKER_00: I mean, to me, it's like yeah, it's like it's like how many ways can Sam Altman say AI is not only you know open AI being nefarious, but AI in general just constantly has this cloud around it, right?
• 12:24
  SPEAKER_01: It's kind of crazy.
• 12:25
  SPEAKER_01: No, it totally is like and you know, we've talked about this a little bit, like he they came out about this model, and like it is equal or better than humans, and that's really great from a tech standpoint and really shitty messaging.
• 12:39
  SPEAKER_01: Like, that's not the way that we should be talking about this.
• 12:41
  SPEAKER_01: Like, oh hey, you know that computer, it's better than you.
• 12:44
  SPEAKER_01: It's like I kind of feel bad.
• 12:45
  SPEAKER_01: Have you ever been to a CVS and they've got a self-checkout?
• 12:48
  SPEAKER_01: And there's somebody standing at the checkout, and you're like, nah, I'll just go to the cell phone.
• 12:52
  SPEAKER_01: It's like I'd rather talk to a computer than you.
• 12:54
  SPEAKER_01: Yeah, that's sort of that's not the communication strategy that I'd be going for if I was doing the PR there.
• 13:00
  SPEAKER_00: Yeah, you know, I'm with you 100%.
• 13:02
  SPEAKER_00: It uh I mean I I can't wait to see how this resolves, but I I gotta say, like from from a you know, you mean take away the math, but like come back to your business.
• 13:13
  SPEAKER_00: Like I think it was Gates a few right a month ago.
• 13:16
  SPEAKER_00: He was like, You're you're paying these providers for intelligence tokens, and oh, by the way, you're also giving them the data on how you run your business.
• 13:25
  SPEAKER_00: Like you're paying twice for the intelligence.
• 13:28
  SPEAKER_00: Um it it it is a flavor of that.
• 13:31
  SPEAKER_00: Like, like hey, do you do you trust like how much do you got do you all trust Sam Altman or do you trust OpenAI or Anthropic?
• 13:40
  SPEAKER_00: Like at what point are they gonna say, you know what, Logics, you guys got something there?
• 13:45
  SPEAKER_00: Let let us see how we can leverage the telemetry or change our terms of use and um and see what makes you guys so special.
• 13:54
  SPEAKER_02: Yeah.
• 13:56
  SPEAKER_02: It's a great business model.
• 13:58
  SPEAKER_02: They pay you, you provide compute, but then you get a bunch of IP.
• 14:04
  SPEAKER_02: What could go wrong?
• 14:05
  SPEAKER_02: Yeah, that's just uh you know, and then you know, as it relates to our business, like our customers trust us that we'll you know, they're we're not gonna take their stuff and go and sell it somewhere else.
• 14:16
  SPEAKER_02: We just would never do that.
• 14:16
  SPEAKER_02: And that's why we've been in business 15 years.
• 14:19
  SPEAKER_02: Right.
• 14:19
  SPEAKER_02: But it's uh yeah, I don't know how long how long that'll last.
• 14:22
  SPEAKER_02: Now if you're screaming to be regulated, say, hey, I'm I'm building something very dangerous here, and oh, by the way, I also take the information, you know, it's gonna accelerate that topic.
• 14:33
  SPEAKER_02: We already have senators saying, hey, we gotta stop AI development right now.
• 14:37
  unknown: Right.
• 14:37
  SPEAKER_02: And it's like, oh, that's that that to me is scarier than anything else.
• 14:40
  SPEAKER_01: Yeah, you can't put the toothpaste back in the tube.
• 14:43
  SPEAKER_01: Like progress is gonna continue getting made in AI no matter what we do.
• 14:47
  SPEAKER_01: So I mean, do we want to be part of it or not?
• 14:49
  SPEAKER_01: Because it's happening.
• 14:51
  SPEAKER_01: Like we're not gonna stop it.
• 14:52
  SPEAKER_02: I think the answer is yes.
• 14:53
  SPEAKER_02: We want to be a part of this.
• 14:54
  SPEAKER_01: Well, no, I mean, yes, we do, but like go and talk to our senators, do you?
• 14:58
  SPEAKER_01: Like they're not they're not all clear on that decision.
• 15:00
  SPEAKER_01: Yeah.
• 15:01
  SPEAKER_00: Yeah.
• 15:02
  SPEAKER_00: I I'd say maybe a minor point too, but also the I think specifically the model that worked on the millennium problem was actually a an unreleased version beyond Astra, which again, it's this cloak and dagger, like like this this shroud of secrecy that I just feel like it's it's just not to me, it's not landing very well.
• 15:24
  SPEAKER_01: So I like that we've got the Fargo picture when you just said the most Fargo thing, like the cloak and dagger.
• 15:31
  SPEAKER_01: I love it so much.
• 15:32
  SPEAKER_01: But like the and you go, it goes back to the lack of traceability, right?
• 15:35
  SPEAKER_01: Why are you removing traceability?
• 15:36
  SPEAKER_01: And like you talked about for our customers, like they ask why does why did the model make this decision?
• 15:42
  SPEAKER_01: And sometimes it's for their own knowledge, but the other side of it is people get audited.
• 15:47
  SPEAKER_00: Yeah.
• 15:47
  SPEAKER_01: And like, you know, we've we did a lot in debt collection, and like the CFPB isn't gonna take the answer of like it just made the decision, don't worry about it.
• 15:56
  SPEAKER_01: Don't ask.
• 15:57
  SPEAKER_02: Yeah, we've all all three of us serve on boards, whether they're nonprofits or other boards.
• 16:04
  SPEAKER_02: And in the end, it it really comes down to governance.
• 16:07
  SPEAKER_02: Like who's on the board at at OpenAI?
• 16:10
  SPEAKER_02: I don't know.
• 16:11
  SPEAKER_02: I should probably know this before I get on the podcast, but but there is a strength.
• 16:15
  SPEAKER_02: Not Elon Musk.
• 16:16
  SPEAKER_02: I think it's that's part of the problem.
• 16:19
  SPEAKER_02: But it's like at some point there's governance, and so when you have you know a CEO going out there and making alleged claims and things like that, there there is a responsibility, you know, for for the board to step in and say, hey, no moss, you keep you you you're gonna have to dial it back, um, or at least be very, very, very sure of what you're saying out there.
• 16:41
  SPEAKER_02: Yeah.
• 16:42
  SPEAKER_01: So but I mean it goes back to the question of is it all press is good press, right?
• 16:46
  SPEAKER_01: I mean, we're talking about open AI.
• 16:48
  SPEAKER_01: It's so yeah, yeah.
• 16:50
  SPEAKER_01: I it's not the way I'd you know run my um giant hyperscaler type thing, but you know, it's it's their choice, that's what they're doing, but they are really sort of digging their heels in in a way that seems a little strange.
• 17:05
  SPEAKER_00: Yeah, and and I you know, I I I I just like I feel like it's like they you know, they release a new product and it's great, but but I feel like it like it it all gets drowned out to me over the stuff like this where it's like there's all these question marks and the like what is the intent and what else is going on behind the scenes and what you know what other data are you stealing and what right it it just it it there's just so many questions I feel like that come with it.
• 17:34
  SPEAKER_01: That's also like I talked the marketing video they put together for Astra is a one wall that is an entire screen, and it is somebody interacting via voice from the one chair that's also in this room with the giant screen, and it's you know bringing up pictures and it's doing all this super interactive and like, oh, by the way, check for a tennis court.
• 17:56
  SPEAKER_01: Nobody's using open AI this way.
• 17:58
  SPEAKER_01: Like, why why are you advertising this thing that like I don't I don't know that that's what you got?
• 18:03
  SPEAKER_01: Right, it's really not.
• 18:04
  SPEAKER_01: So it's kind of weird.
• 18:06
  SPEAKER_02: Yeah.
• 18:07
  SPEAKER_02: So can I ask you guys a quick question?
• 18:08
  SPEAKER_02: Yeah, I chose children's hospitals as to where they could tear down to build data centers.
• 18:14
  SPEAKER_02: What would be more offensive?
• 18:16
  SPEAKER_02: I mean, that's your guys' opinion.
• 18:17
  SPEAKER_00: I mean, that's a tough one.
• 18:18
  SPEAKER_02: Would it be the Fargo Theater?
• 18:20
  SPEAKER_00: Oh, it could be, yeah, it'd be the Fargo dome.
• 18:22
  SPEAKER_00: Yeah, you put it where the buys in place.
• 18:24
  SPEAKER_01: No, come on, you're tearing down the orphanages, Ted.
• 18:26
  SPEAKER_01: We don't just want the sick kids, we want the sick parentless children.
• 18:29
  SPEAKER_01: Like that's where we really want to tear it down.
• 18:32
  SPEAKER_00: It's almost hard though to make, yeah.
• 18:34
  SPEAKER_00: I mean, it's like, yeah, what what is what is the next thing?
• 18:39
  SPEAKER_01: How could it be worse?
• 18:40
  SPEAKER_00: And honestly, I mean I was gonna say that's great.
• 18:44
  SPEAKER_00: That's uh that's exactly it.
• 18:46
  SPEAKER_00: So uh we'll wrap it there.
• 18:48
  SPEAKER_00: As always, um anything comes up between now and white next week, go ahead and and reach out.
• 18:53
  SPEAKER_00: Happy to hop on a call or or find some other way to connect.
• 18:58
  SPEAKER_00: Um so until then, um keep learning, even when it's unsupervised.