Why The Next AI Breakthroughs Will Be In Reasoning, Not Scaling

Harj Taggar 0:00 I remember about a year ago, one of these conversations around, "Are we going to have AGI? What would that look like?" One- one of the arguments for it was that, "Well, like, at some point the AI will get good enough to just, like, design chips better than, like, humans can and then it will just, like, eliminate one of its bottlenecks for, like, getting greater intelligence." And so it feels we're on the pathway to that in a way that we just weren't before.

Garry Tan 0:20 The last episode, we were talking about, you know, what are you going to do with these two more orders of magnitude? Since then, Sam has, uh, told me that he actually wants to go to four orders of magnitude. It's the worst that these models are ever going to be right now, right this moment. Week to week, you know, there are things that you couldn't do maybe a month ago that you could do really, really well right now. So that sounds like a pretty crazy moment in history. (laughs) Welcome back to another episode of The Light Cone. I'm Gary. This is Jared, Harj, and Diana. And at Y Combinator, we've funded companies worth more than $600 billion, and we fund hundreds of companies every single year. So we're right there on the edge of seeing what is going to work, both in startups and in AI.

Garry Tan 1:15 Recently, Sam Altman wrote this pretty wild essay that predicted that AGI and ASI are coming within thousands of days. Seeing him on Monday, he actually directly estimated, you know, between 4 and 15 years. Have you guys read this essay yet and what do you think?

Jared Friedman 1:35 Yeah, I read it, and one- one of the interesting places where I think we have a unique perspective is that we were- we had a front row seat to the very beginnings of OpenAI 'cause OpenAI basically spun out of YC. And so what was cool to me reading this essay is that it's literally the same ideas that Sam was talking about in 2015 when he started OpenAI. Like, he's been talking about this, like, basically since I've known the guy. Um, and in 2015, when he said these things, he sounded kind of like a crazy person-

Garry Tan 2:06 (laughs)

Jared Friedman 2:06 ... and not that many people took him seriously. And now, 10 years later, it turns out he was right, and actually, we were much closer to AGI than anybody thought in 2015. And now it doesn't sound crazy at all, it sounds, like, totally plausible.

Garry Tan 2:20 I mean, the essay itself is pretty much the most techno-optimist thing I've read in a really long time. Some of the things that he says are coming are pretty wild. Space colonies, fixing the climate problem, um, y- intelligence on tap, being able to solve abundant energy. Uh, yeah, I think he's basically ushering in this sort of Star Trek future on the back of literally human intelligence, being able to figure out all of physics.

Jared Friedman 2:49 Yeah, Sam was always a re- like, I- I remember back when he was starting OpenAI, one of the things that really motivated him to do it was he believed that when we actually had AGI, basically it'd be better at doing science than humans were, and therefore it would accelerate the rate of all scientific progress in every scientific field. That was- that was part of the motivation from the very beginning, and I think it's really connected to o1. Even when Sam came and spoke at our- at our- at our batch a year ago, this is long before o1 was publicly released but it was being worked on in, you know- in, uh, in secrecy by OpenAI. That was the thing that he was most excited to talk about, was giving GPT more advanced reasoning capabilities. And I think this is the reason, it's because, like, the thing that's missing from its ability to actually do science and, like, accelerate technological progress is it needs- it needs to be able to, like, think through things. (laughs)

Harj Taggar 3:43 One thing that really stands about o1, um, in particular, is if you read, um, one of the papers talking about it, so Capabilities and Potentials for the Future, it talks about how it does really well in chip design. Um, and I remember about a year ago, one of- like, one of these conversations around, "Are we going to have AGI? What would that look like?" One- one of the arguments for it was that, "Well, like, at some point, the AI will get good enough to just, like, design chips better than, like, humans can and then it will just, like, eliminate one of its bottlenecks for, like, getting greater intelligence." And so it feels like that's already kind of, like, we're on the pathway to that in a way that we just weren't before.

Jared Friedman 4:17 Diana

Jared Friedman 4:18 is going to show a cool demo of her doing exactly that.

Diana Hu 4:20 (laughs) It's fun because we ran this, um, hackathon with OpenAI and Sam came over and judged the winners. And one of the participants was actually chip design. This company is called Diode Computer, I think we mentioned them earlier. What they're building is basically AI designer for circuit design. And their previous product, it could handle in the- if you think about PCB design, there's four ma- major steps. The big expensive part that you need a lot of- all of these need a lot of expertise, is like the system design, how do you really put together the architecture of it? How do you design all the components, like the resistors they need, the sensors, the specific, um, processing units? Then you needs to go do the layout with schematics, placing- then doing the routing, and routing is known to be a NP-complete problem because as you have different layers in circuit boards, there's interference. And this is why companies like NVIDIA, Intel, Apple have a gazillion electrical engineers, because this is a NP-complete problem. Up to GPT-4, which is this company have built, it actually put some constraints and was able to automate a lot of the, uh, schematic design that you as a human had to design what components it needed to go on the design, and to some extent the routing, it was simple, which is still pretty cool up to that point. So they were able to automate all that, but the thing that they demonstrated now with o1 was actually able to do the system design and component selection, which is-... crazy. So, it would be able to read all the data sheets and select the right components. So, the way the product would work, it could say, "I want to build a wearable heart rate monitor with an accelerometer and a microcontroller." Very high level. And given these constraint and looking at the database, it would be able to match the specific accelerometer and microcontroller and heart rate monitor sensor and connect it and just output the end result.

6:25 What we are trying to build today is a wearable heart rate monitor, something like you would see in a Whoop, for example. Um, the O1 is amazing, but one of the downsides is that it's a bit slow. So, we actually cached a pre-generated, like, system diagram that O1 was able to generate. It's pretty good. It has a USB-C connector, an IMU, like we requested, a heart rate sensor, um, and, uh, like, this is a microcontroller. So, I'm going to show you how you can go from this, uh, and, like, build a PCB. So, we are going to, like, build the project. Um, the output of this is code. We actually use AutoPyL, which is a, uh, electronics as code, uh, language. And you can see that it took all the blocks in the block diagram, stitched them together exactly how we want. Um, and, uh, the second step is it actually is going to generate a layout for the board. Um, and so now, like, we can directly open it and, uh, here you go. (laughs) Here's the board. It's, uh, quite i- it's quite nice. Uh, there's still, like, a couple of, like, fine-tuning steps required. For example, um, we could, like, move, uh, like, this USB Type-C connector slightly. Um, we can, like, change the shape of the board. But, but these are all the components. Um, and then, like, thanks to the system that we built, uh, we can, like, call the auto-router on this specific board and actually get a fully working, uh, printed circuit board back.

7:48 So, this is, uh, actually one of the examples on the O1 paper.

7:51 Yeah.

7:52 That it would do EDA, but actually they went a step for- forward because the example on the paper, they described the EDA as this step process, is this set of tools for circuit design. It does sort of the design of the schematic, also the simulation and bug verification. It's easier to verify stuff than to select and write it.

8:14 Hmm.

8:15 So this company actually went a step further beyond the paper.

8:18 Okay, cool.

8:18 Because the paper did mostly the last stages of, uh, verification and simulation.

Garry Tan 8:22 I guess it's an interesting example of using different models for different tasks and in different workflows. So, in order to actually pick the correct components off the bat, you know, even before you, you know, place it on a circuit board, you've got to actually have probably RAG on structured... You know, taking unstructured data like PDF documentation and turning into a structured form that then 40 mini sounds like is being used to actually extract the data and then put it into, um, like, a usable form.

Diana Hu 8:53 Into a format for O1.

Garry Tan 8:55 Yeah.

Diana Hu 8:55 I think this is actually a very common pattern that we're seeing. A lot of the interesting products built with AI, you use different kinds of models. So yes, 40 minis for PDF extraction and then O1 for the reasoning, because it's actually very hard to select the components for parts. I know, uh, Jared, you also work with a lot of hard tech companies and the whole part of selecting whatever, the servos, the motors, the sensors, it's like so... It's, it's, it takes a lot of thinking for human.

Jared Friedman 9:22 Yeah. The, the other thing I think is interesting about this example is, like, during the batch, before O1 came out, Diode had tried to do this with GPT-4o, and it just flat out didn't work. And then they basically tried the same thing, the same prompts, but fed it to O1 and boom, all of a sudden it worked. (laughs) And so there really is a sort of like step function capability unlocked.

Diana Hu 9:45 They were so excited when I talked to them and they showed me. They were, they had this big smiles, like, "Wow."

Jared Friedman 9:49 (laughs)

Diana Hu 9:49 They themselves were super impressed.

Jared Friedman 9:51 This hackathon that Diana ran, incidentally, I, I think is a really interesting concept for a hackathon. Like most, most hackathons are, like, people who are just sort of like building something that they plan to throw away. Um, and the cool thing about this hackathon is it was all actual YC-funded startups that have real businesses, that have funding, that have like a real thing with real users, and they were all building actual features for their product that they plan to release to real users. It was, it was really cool, I think, for us to see how O1 unlock capabilities for real companies, not just like toy projects.

Garry Tan 10:24 Yeah.

Diana Hu 10:25 There's another one that was similar in here in terms of reasoning for O1. I think, uh, Harsh, you work with, uh, Camfer.

Harj Taggar 10:31 Yep.

Diana Hu 10:31 So, want to tell us what Camfer does?

Harj Taggar 10:34 Uh, it's... I mean, the tagline is Devon for CAD, but basically they, um, let you create CAD designs with just natural language. Like, you just type in, like, you know, something that you want to design and it just, like, spits out, like, a CAD design for you.

Garry Tan 10:49 So can you design me five airfoils optimized for 50 miles per hour with a minimum drag to lift of 15 at a five-degree angle of attack? This is very specific.

Harj Taggar 11:00 Yep.

Diana Hu 11:01 Normally this would require, uh, actually a mechanical engineer to be running all the simulations and solving through the equations. And, well, you're seeing why it's, like, flashing as they running all the multiple simulations for them at the same time.

Garry Tan 11:15 So it's actually kind of like a copilot to, uh, SolidWorks.

Harj Taggar 11:18 Yeah. They actually built their... Like, initially they were going to build this as a plugin to SolidWorks, but they went for, like, the even harder technical approach, which was like, "No, this is just like a executable that runs on your desktop and it essentially opens up SolidWorks for you and..."

Jared Friedman 11:32 And then just starts, like, clicking around in the UI-

Harj Taggar 11:35 Yeah.

Jared Friedman 11:35 ... pretending to be a person?

Harj Taggar 11:36 Yeah.

Jared Friedman 11:36 Nice.

Diana Hu 11:37 And you saw there what was really cool earlier, they flashed the, the math trace. So O1 was actually able to write all of these equations, all these-

Jared Friedman 11:45 Whoa.

Diana Hu 11:46 ... partial differential equations and solve basically naive Stokes questions to actually solve airfoil.

Garry Tan 11:52 That is really cool. The last episode, we were talking about, you know, what are you gonna do with these two more orders of magnitude? Since then, Sam has, uh, told me that he actually wants to go to four orders of magnitude-

Jared Friedman 12:05 (laughs)

Garry Tan 12:05 ... to get to a trillion dollars (laughs) in, uh, you know, sort of spend.

Garry Tan 12:09 I mean, pretty wild. But on the other hand, like, you could see where that might go, you know. You could imagine an airfoil is still... it's, you know, very impressive and complex, but sort of what's... we're capable of doing today in 2024, you could imagine abstracting that to, like, understanding the nature of (laughs) physics, I suppose. Like, it, it would be sort of hard to see that maybe in the current version of o1, but if the scaling laws hold, it seems entirely plausible that, you know, far more difficult engineering challenges such as, um, you know, room temperature fusion. Like, these are all sort of ultimately engineering problems.

Diana Hu 12:51 Fluid mechanics, there's weather prediction, there's all these complex physical phenomenas that are very hard to solve and you need basically PhDs. And to Sam's essay, this is a glimpse into what AI and where o1 is heading with, with this chain of thought and reasoning.

Jared Friedman 13:12 Especially, like, like, Sam's essay, The, The Vibes Are sort of training intelligence and this new age of intelligence, and then the o1 paper just... The... I think this whole idea of now you can actually give, like, feedback not just on, like, the outputs and whether you got the correct answer, but, like, on all of the steps to get there, and, like, you're basically teaching a model how to think. Like, the comfort guys are mentioning it too, right? The reasoning traces, and they can probably go back and, like, fine-tune the various steps for, like, every output to make sure that the model's thinking how they want it to think. That one, that just is, again, very, like, the AGI conversations I feel, like, a year ago were all sort of in this direction of, like, what happens once you can actually start teaching the models to think better versus just, like, um, spitting out the correct answers. Uh, and then the scaling laws, this is, like, even more surface area for, like, throwing compute at the problem, right?

Garry Tan 14:03 Yes.

Jared Friedman 14:03 Like, now you can just basically put compute at the inference step and-

Garry Tan 14:09 And iteratively have something come out that, you know, you can actually spend more money and more time and have a result that iteratively gets better-

Jared Friedman 14:18 Yeah.

Garry Tan 14:18 ... similar to what you might expect from a human scientific organization.

Jared Friedman 14:22 Yep.

Garry Tan 14:24 Maybe more consistently, even. (laughs)

Jared Friedman 14:26 Diana, do you, do you want to talk about the architecture and how they actually created o1?

Diana Hu 14:30 I think a lot of it is inspired from what they've been working for many years since the beginning of, uh, OpenAI. I think one of the inspirations

Diana Hu 14:42 is a lot of the work they did with DOTA.

Jared Friedman 14:44 Yeah. Does everyone remember when, like, before OpenAI was famous for GPT, the one thing that it was, like, kind of famous for, that at least people in the tech industry knew, was DOTA, (laughs) was, like, winning video game competitions. That was their first big breakthrough.

Diana Hu 14:58 And the funny thing to think, back then, DOTA wasn't something that took the world for a storm. I mean, maybe only the research community kind of knew about it, but it wasn't anything practical. But what was impressive, it was beating a lot of the best DOTA players. So, DOTA is this complex game of resources and planning, right? And they implemented a lot of, uh, kind of reinforcement learning type of techniques in there, which I think were also inspired early days from AlphaGo and AlphaZero as well on how it solved Go.

Jared Friedman 15:29 Yeah.

Diana Hu 15:29 It wasn't just brute forcing through it, but actually having a rework function and, and trying to solve towards it.

Jared Friedman 15:36 Yeah.

Diana Hu 15:37 And even... This is why there's just so, so much talk about Q-learning, because that's sort of the fundamental algorithms behind... family of algorithms behind RL.

Jared Friedman 15:46 So, yeah. So, like, because of DOTA, they got really good at doing reinforcement learning. That's how they got it to work. They just had it play against itself, like, a million games. (laughs) Um, and then how, how does that connect to, to, to o1?

Diana Hu 15:59 So, I think this is where there's a bit of a big stepping function, because how do you then incorporate that into the family of GPT type of models? GPT is all generative based on predicting the next token and patterns, and then getting those results to check that they're correct. So, I th- I think a lot of it is you had to have a lot of data that was factually correct and fed into probably the model and the training and having a reward function that get it to reason a bit more about the output and make sure that it's correct. So, they've probably done a lot of, um, interesting techniques with that, and there's probably a lot of secret sauce on the type of, uh, data sources they've done. Maybe one of the speculations we could do is a lot of very factually correct information.

Jared Friedman 16:48 Like math problems and science problems and things like that, yeah.

Diana Hu 16:51 And that's why it outperforms so much in those.

Jared Friedman 16:53 Yes. Yeah, one of the things I think is interesting, Gary, to your point about the scaling laws, is a lot of people are really focused on the next, like, scale up of the model, like the GPT-5 series of models, which are being trained now, and people are working on them, and they are gonna come out. Um, but I think people may be underappreciating how big an unlock this other direction is, 'cause there's, there, there's two research directions being explored in parallel, right? Like, one is the straightforward scale up of the underlying LLM, and then this o1 direction is, like, a totally orthogonal research direction in which you unhobble the model by having it do reinforcement l- learning while actually trying to do things in the real world and getting better at them. The version that's come out so far, it's still only o1 mini, and if you look at the actual-

Diana Hu 17:40 o1-preview.

Jared Friedman 17:41 Sorry, o1, o1-preview, and, like, if you look at the performance tests that they released, like, the full o1 model, which is coming out any day now, is a huge step function- Huh.

Garry Tan 17:49 Yeah.

Jared Friedman 17:49 ... above even o1-preview, which is what enabled all these incredible results at the hackathon. Sam was just telling us that, like, o2 and o3 are not far behind, and so, like, I think people may be underappreciating just how big an unlock we're gonna get.

Garry Tan 18:02 Yeah. And o1 also is really opaque still. I mean, it... From a, you know, sort of business perspective, this is a new method. I think at great cost to themselves, they actually did create a new dataset to train the chain of thoughts. You know, it, it's essentially a giant dataset of, you know, given task X, can you break it down and, uh, into... you know, break it down into parts? And you know, what's funny is this sort of rhymes with what Jake...... uh, Heller figured out for case text that-

Jared Friedman 18:31 Yeah.

Garry Tan 18:31 ... if, uh, a given task that you give an LLM is hallucinating or, you know, not consistently giving the output you want, you're trying to make that particular prompt do too many things, you need to break it down into steps. And so what's funny is Jake's prescription is really two parts. You know, one is break it down into steps, and then the other part is evals. And it sounds like basically with O1, the chain of thoughts will replace the workflow. So, you might not need to break it down into steps yourself, but the evals are still really important. Um-

Jared Friedman 19:09 Mm-hmm.

Garry Tan 19:10 Even like in the aftermath of that episode with Jake Heller, it sounds like some YC alums are reaching out and saying, "That episode helped us figure out and unlock something really big." Like-

Jared Friedman 19:21 Yeah.

Garry Tan 19:21 ... a lot of people really were just raw dogging their prompts. (laughs)

Jared Friedman 19:25 Raw dogging their prompts. (laughs) Yeah.

Diana Hu 19:26 They got to, uh, you have the example of a company you worked with, Jared, that got to 100%.

Jared Friedman 19:31 Yeah, just by doing exactly what Jake recommended, which is like having a really big eval set and being very careful about testing every step of your reasoning pipeline.

Garry Tan 19:42 So, one of the theories that I have now is ultimately like if you superimpose that on what is a moat? I mean, that's one of the key-

Jared Friedman 19:50 Right.

Garry Tan 19:50 ... questions that everyone's sort of asking themselves right now, you know. Okay, like GP5's coming, two more orders, maybe four more orders of magnitude are gonna come in terms of a trillion dollars spent on more training. That's pretty wild. You know, if I'm a wrapper company or I'm trying to do vertical SaaS or I'm trying to, you know, build my own business, what do I do? My theory would be it's the evals. It's, you know, write the 10,000 test cases and the only way you get access to the test cases that are proprietary data that are not av- like commonly available is that you literally, you know, that's what a bunch of our companies in this current YC batch are doing. They're doing the hard work of doing enterprise sales, they're getting embedded and sort of going "undercover" into these, uh, you know, sometimes really boring jobs, sometimes really complex or arcane jobs. You know, it's everything from, you know, I think accounts receivable all the way over to how do you do like financial accounting or forensic accounting? Like it's just all kinds of things that are really, uh, not readily available. Um, you- you can almost argue that anything that is consumer and publicly available on the internet, that's gonna be in the base model.

Jared Friedman 21:04 Yeah.

Garry Tan 21:05 So then your moat ultimately is for all of the other things that are not already online, whether it's, you know, for case text being a lawyer, or maybe over here on science or, you know, in terms of building airfoils, like what you're trying to find is the, the data that is proprietary in some use case, some vertical, that allows you to build that 10,000 test case, um, eval, and then that's actually the value. I mean, this is just a crazy-

Jared Friedman 21:31 Yeah.

Garry Tan 21:32 ... theory, but-

Jared Friedman 21:32 Yeah.

Diana Hu 21:32 It's actually, uh...

Garry Tan 21:33 ... it's something that might-

Jared Friedman 21:34 Yeah.

Garry Tan 21:34 ... might be what happens.

Jared Friedman 21:35 Totally.

Harj Taggar 21:35 An interesting implication for startups based on everything you just said, um, is it may be worth thinking about who, like of your customers, picking the ones that will pay you a lot for that final like 10% like accuracy and perfection. I think like Canva

Harj Taggar 21:52 are actually a good example of it, where there's lots of interest in this sort of text to CAD design amongst like hobbyists or people who want to prototype things and get something up and running very, very quickly. Um, but there's also like a segment of the market where it's people who are literally designing like, you know, airplane parts where there is no room or like margin for error. And O1 makes it quite easy or easier now, right, to get to like the prototype like, you know, 80% of the way there. But I think the strongest technical teams have the option to go all the way and go after this segment of customers who want like 100% accuracy and will pay a lot for it. Um...

Garry Tan 22:29 Harsh, always go all the way.

Harj Taggar 22:30 Yeah, well, I can always go all the way. (laughs)

Garry Tan 22:31 You have to go all the way. Yeah.

Harj Taggar 22:33 But I- I think it's interesting because o- one, one of the things that gets pushed is does O1 or does AI in general actually make it commoditize a lot of the tech and make it less important to be a strong technical team? Um, and it just seems unlikely to me. It seems like actually the last time...

Jared Friedman 22:49 It seems it's the opposite.

Harj Taggar 22:51 Yeah.

Jared Friedman 22:51 Yes.

Garry Tan 22:51 (laughs)

Harj Taggar 22:51 Like all of the value is probably going to be captured by like the strongest technical teams who can build on top of whatever the base level of tech is and get the final 10%.

Jared Friedman 23:02 Hey Gary, I think it's the prompts, it's- it's the evals, and it's also like the UI layer and the integrations that go around it. 'Cause like just the prompts themselves are not a product. For a company to actually adopt Camfer, like it needs to actually integrate into their existing tools, it needs to have a well-thought-through UI and-

Harj Taggar 23:20 Yeah.

Jared Friedman 23:20 ... workflow and all the tooling to sort of-

Harj Taggar 23:22 Yeah.

Jared Friedman 23:22 ... make the prompts useful.

Harj Taggar 23:23 Yeah.

Garry Tan 23:24 Well, and then it's distribution, right? Like-

Jared Friedman 23:26 Yeah.

Garry Tan 23:26 ... how do you actually get in front of people?

Jared Friedman 23:28 Yeah.

Garry Tan 23:28 How do you establish your brand? And then, uh, a perfectly good moat is difficulty switching, actually.

Jared Friedman 23:34 Totally.

Garry Tan 23:35 And once you have all your data and it's working-

Jared Friedman 23:37 Yeah.

Garry Tan 23:37 ... and you're paying $10,000 or $100,000 ACV, sometimes a million to 10 million ACV, (laughs) uh, you know, man, it's gonna be hard to switch. So, all the classic moats still apply. You know, this is still software, but, you know, you can unlock this capability and i- i- you know, this is a moment, you know?

Diana Hu 23:56 Another point to double down on the importance of evals is that that still applies in the world of O1 as founders are wondering, "How am I gonna still build the best product on top of O1? Does it change?" And everything we discussed in the episode with Jake Heller applies because Gigaml is this company that, uh, Harshi worked with.

Harj Taggar 24:17 Yeah, and Gary too.

Jared Friedman 24:19 (laughs)

Diana Hu 24:19 And Gary, right?

Harj Taggar 24:19 Yeah.

Garry Tan 24:20 I adopted them.

Diana Hu 24:21 Can you tell us-

Harj Taggar 24:22 (laughing) They have. Yeah, they, that's the history.

Diana Hu 24:25 Maybe tell us a bit about what they, what they do.

Harj Taggar 24:27 The full, full backstory is we funded them for a completely different, um, idea, something like, they're an Indian founding team and the original idea was something he- helping Indian co- Indian high school students apply to US colleges, a very like niche idea.

Garry Tan 24:39 (laughs) Yeah.

Harj Taggar 24:40 Um-

Garry Tan 24:40 But they're super cracked IIT, AI engineers-

Harj Taggar 24:43 Yeah.

Garry Tan 24:43 ... researchers.

Harj Taggar 24:44 Yeah. And it just ba- it just happened, like we were like, "This is not a great idea. AI is, you know, changing the world," and you lit- like you'll-... research (laughs) like, that you've been doing at university or college is all aligned with, like, um, i- in particular, like, fine-tuning LLM models or something?

Jared Friedman 25:00 O- o- originally, it wasn't even the AI version of helping Indian high school students apply to-

Harj Taggar 25:05 No. It was just-

Jared Friedman 25:05 ... to, to college. (laughs)

Harj Taggar 25:06 It was, yeah. (laughs) No, it had nothing to do with AI, actually.

Jared Friedman 25:07 (laughs)

Harj Taggar 25:07 So it's like a classic YC story, where it was like-

Jared Friedman 25:09 (laughs)

Harj Taggar 25:09 ... "These are two clearly brilliant engineers."

Jared Friedman 25:12 (laughs)

Harj Taggar 25:12 "Um, we don't like the idea at all, but we should just fund them anyway and hope something works out." And the idea they actually pivoted to initially, which they raised a seed round for, um, was helping companies fine-tune, um, open source models so that they could get to, like, equivalent performance. Um, uh, as, you know, at the time, it was really only OpenAI. But, um, what I think, in general, what we found is that those have not proven to be great businesses, because just, like, the cost of the models has gone down, and also, like, you know, the performance of the open source models has gone up.

Garry Tan 25:45 You just haven't had to fine-tune as much as people thought you would need to.

Harj Taggar 25:48 Yeah.

Diana Hu 25:49 'Cause the models are just keep getting better. It's kind of betting on the different, on the opposite direction-

Harj Taggar 25:54 Yeah.

Diana Hu 25:54 ... on AGI that, "Let's just trust that these models are going to keep getting better and better, which, ergo, doesn't require as much fine-tuning."

Harj Taggar 26:01 Yep. And so they pivoted again into, "Well, like, let's just find a vers- like, let's just find, like, a... We're really good at AI now. We're, like, world experts in, like, fine-tuning and squeezing performance out of these models, so let's just, like, find a vertical application for that." And they went into, um, AI customer support, which is, like, competitive, but again, I just think e- if you're an in- intensely technical team, you can still find ways to squeeze out, like, a comparative edge against other teams in the space. And I think that's what they've done.

Diana Hu 26:35 The problem with customer support is you're dealing with very kind of squishy problems. There's just so many edge cases. It's just the space of things that could go wrong, as a customer rep, is enormous.

Jared Friedman 26:47 Well, it, it seems competitive, but the thing is, hardly any adoption has actually happened.

Harj Taggar 26:54 Yeah.

Jared Friedman 26:55 Like, it's not like the world has replaced all the customer support agents with AI yet. We can all see that it's going to happen, but it hasn't happened yet. And so from that standpoint, like, it's wide open.

Harj Taggar 27:05 What I, what I found, at least when I spoke to the Geeka ML team last, is that, um, part of the reason for the lack of adoption is that rules-based systems work fairly well for most, like, of the simple cases, and there's just not trust or belief that you can, like, build AI that's good enough to solve, like, the real messy stuff. And so most companies that were pitched on, like, an AI customer support agent were like, "Well, you can't actually go all the way and solve, like, the, like, h- hardest problems that take up most of the time. And the rules-based system works, like, totally fine for everything else." And so it, I remember when they were first pitching this idea, people would just be like, "This is just overkill. We don't need it. A rules-based system works totally fine." But, like, seems like that is no longer the case.

Jared Friedman 27:47 Yeah, 'cause they now have some really legit customers. Whose...

Harj Taggar 27:50 Uh, Zepto just signed up.

Jared Friedman 27:52 Okay.

Garry Tan 27:52 So last time I did office hours with them, they said that they automated 30,000 tickets per day.

Harj Taggar 27:58 Wow.

Garry Tan 27:58 And so they, you know, I think Zepto had more than 1,000 people working, uh, on those 30,000 tickets per day, which is, you know, 30 tickets a day. And then, uh, the interesting thing was, you know, on the one hand, you know, this is probably one of the things that frankly everyone, when they think about AI, they're a little bit worried. Like, you know, "Are these jobs going to go away?" And, uh, the interesting thing about the Zepto customer support, uh, job is that it's so not a fun job that I think the turnover rate was something like a few months.

Harj Taggar 28:32 Interesting.

Garry Tan 28:32 Like, you know, most customer support agents only wanted to work there for, you know, six months or less. So, you know, this actually is an interesting case of when, you know, when something is incredibly rote, it's literally replacing butter passing. Like, these are sometimes not really actually good jobs.

Harj Taggar 28:51 (laughs)

Garry Tan 28:51 And, um, you know, hopefully those people can go and do something way more awesome with their time and, you know, beautiful brains than-

Harj Taggar 28:59 (laughs)

Garry Tan 28:59 ... you know, these rote jobs that-

Jared Friedman 29:00 Than, than apologizing for Zepto orders that got misplaced. (laughs)

Harj Taggar 29:03 (laughs)

Garry Tan 29:03 Exactly. Right.

Diana Hu 29:05 But the crazy thing they figured out with o1 is that, to your point, Harsh, their previous implementation before o1 was GPT plus rules and all that, and it would not be able to handle m- most of the cases.

Harj Taggar 29:18 Yeah.

Diana Hu 29:18 It would have about a 70% error rate. Now what they did is, doing the technique like Jake Heller described with really going hardcore on the evals plus o1, during the hackathon, they got to only 5% error, which is, that's an order of magnitude improvement.

Jared Friedman 29:37 Yeah.

Harj Taggar 29:37 The other role in this is incredible, too, right? This is what I'm saying, like, the complex, like, the things that are, like, very complicated that take up lots of time and are expensive to solve, would, like, essentially, like, they could not do them. They were-

Diana Hu 29:47 They were shrunk. (laughs)

Harj Taggar 29:47 (laughs) Yeah, yeah, so basically just-

Jared Friedman 29:48 Yeah.

Harj Taggar 29:48 ... like 0%. And, and that's what I'm... That- that's what they were encountering when they were selling this, is a lot of people are like, "Well, actually, all of the... The stuff that we want to automate are these, like, complicated edge cases that waste lots of time." Um, and, like, they just, they- they couldn't actually do any of that. But, like, now they're at, like, 15%, and that's with, like, o1-preview alone. Oh, so that, that's-

Jared Friedman 30:08 No.

Harj Taggar 30:08 Is that 15% of just error?

Jared Friedman 30:09 No, that's error.

Harj Taggar 30:09 So now they're at, like, um-

Jared Friedman 30:10 They're at 85% accuracy.

Harj Taggar 30:11 85%. Yeah, so they went from-

Diana Hu 30:12 Yeah, so they went from shrugging. (laughs)

Harj Taggar 30:12 ... zero to 85%. (laughs)

Jared Friedman 30:14 They went from 0% accuracy to 85% accuracy, yeah.

Garry Tan 30:17 So the interesting thing here is that o1, it's not even o1 yet. It's o1-preview. And then, uh, it's such a new technique that I think they're trying to protect their advantage right now.

Harj Taggar 30:31 Yeah.

Garry Tan 30:31 So, uh, you know, if you use o1 in ChatGPT, it looks like it will tell you what's really going on. But apparently they have a fake model (laughs) that just spits out things to give you the impression that it's breaking it up into steps. And, uh, they've actually, you know, hidden it because they don't want other people to have access to that data yet. Um, but the next step seems like it needs to be some interpretability, some directability. And then for that to happen, you know, I'd be curious if o2 ends up having that. Like, you want to be able to see, "Okay, well, show me the work, show me the steps, and oh, like, that step, the third step, you know, can we rerun this? But I want this to branch in this way."

Diana Hu 31:11 Or edit it. I think this is one of the things that would be the next unlock is, um, right now it has the plan that it comes out, the chain of thought, but you cannot edit it. So imagine now, right now today, o1 just outputs, whatever, 15 steps to the problem that you need to solve.And imagine now being able to edit each of the steps. Then you get into the super, super fine-tuned next level of, uh, Jay Keller.

Garry Tan 31:32 So this is the... it's the worst that these models are ever going to be right now, right this moment. And, you know, literally week to week, w- you know, there are things that you couldn't do maybe a month ago that you could do really, really well right now. So that sounds like a pretty crazy moment in history.

Diana Hu 31:50 So we've been talking a lot about the kinds of companies and ideas that get this wave of uplift from this model improvement for o1. What are the kinds of ideas that are the opposite, that are not getting

Diana Hu 32:06 benefited as much from o1 and perhaps maybe people even should pivot? Because they're getting... they might just get deprecated from the improvements of o1, o2, o3?

Harj Taggar 32:18 I wouldn't go all the way and suggest they should pivot, but I do think companies that are building AI coding agents or AI, um, program engineers are potentially, um... have stuff to think about here because it seems like o1 in particular is, like, outperforming on just, you know, solving programming problems essentially. Um, and I, I certainly know some of the teams I've worked with in the past, like, a lot of what they've invested in is, like, the chain of thought infrastructure behind this stuff, which is now just... o1 doesn't... is not actually, like, any leap forward for them. They've already, like, invested in that already. And so-

Garry Tan 32:56 I think that might be a function of, um, basically the opaque nature of what the chain of thought is. And once you get it to be directable, that's actually... I mean, frankly, that's what users and CodeGen are struggling with even right now. Like, once it starts going down a certain path, you can't really alter things. Like, you, you want it to ask you, "Hey, do you want me to do it like this or that?" And, you know, all of the systems are a little bit struggling with that right now.

Jared Friedman 33:23 I was gonna ask the, the inverse question, Diana, which is, like, each new model capability unlocks a new set of startup ideas. Like, um, a year ago, doing startup ideas where, like, the AI agent would talk on the phone just, like, didn't work. We had a bunch of companies that tried, and all the companies didn't, didn't work. And over the summer, it really started working.

Harj Taggar 33:45 Yeah.

Jared Friedman 33:45 And w- under the trends from the, the, the past two batches, like, anything around, like, phone calling is, like, blowing up right now because the models finally work. So, like, with this new o1 series of models, what are the startup ideas that, like, just became possible?

Diana Hu 34:00 To connect to Sam's essay, is a lot of things that are gonna g- make the atom world, physical world better because it's really good at math and physics. So any startup that's working around mechanical engineering, electrical engineering, chemical engineering, bioengineering, all of these things that really will make our lives better, I think really will... are getting an unlock as we've seen from the demos we highlighted.

Garry Tan 34:26 That's exciting. I mean, it can't just be helping people click a little bit faster. (laughs)

Harj Taggar 34:30 (laughs)

Garry Tan 34:31 It's gotta be things that actually create real world abundance for everyone. And, um, th- it might just be a little bit of a race. Like, I think there's sort of the fear of AI out there in society right now, and then,

Garry Tan 34:44 um, it's sort of up to the technologists to try to usher in this age of abundance sooner rath- rather than later. And if we can do that, then, um, abundance will win out over fear. So with that, I think we're out of time for this week of The Light Cone. We'll see you guys next time.