The Next Gen AI Models: Reliable, Consistent, Trustworthy — With Aidan Gomez

0:00An author of the original paper that launched the generative AI revolution joins us to sort out the technology's myths and facts. and when we'll see a return on all that investment. All that and more is coming up right after this.

0:15You're used to hearing my voice on the world bringing you interviews from around the globe. And you hear me reporting environment and climate news. I'm Carolyn Buehler. And I'm Marco Werman. We're now with you hosting The World Together. More global journalism with a fresh new sound. Listen to The World on your local public radio station and wherever you find your podcasts.

0:42Welcome to Big Technology Podcast, a show for cool-headed, nuanced conversation of the tech world and beyond. We have a great show for you today because we're joined by Aidan Gomez, the CEO of Cohere, an AI platform for enterprise. And he's also the co-author of the famous Attention is All You Need paper, which invented the transformer and started this whole AI thing. Aidan, great to see you. Welcome to the show. Thanks, Alex. Thanks for having me. Great to have you here. I want to begin with some myths and facts about AI. We have debates all the time on the show. Where's the technology going?

1:14Is it worth the investment? And no better person to ask than someone who was there at the beginning and now currently an entrepreneur in the space. So earlier this month, OpenAI raised$6.6 billion, but the reporting says they might be losing$5 billion per year. And in some ways, OK, you need the investment to build the models. But in another way, it's like, OK, where does this end? Because the compute, the data, the energy to train these models keeps getting bigger. The requirements for that keep getting bigger. And so does the money. And does this ever become sustainable? So what do you think?

1:51I certainly understand the urge for people to see the numbers being spent on training and be concerned that it's not going to recoup in value. But I think that those numbers are actually small relative to the long-term value that the technology will deliver. I think it is now time to prove that. So last year was very much the year of the proof of concept. People were getting familiar with the technology was their first time working with it. And so there were a lot of small tests and experiments. But this year is very much one of going to production and getting these models into the hands of people at scale.

2:32Of course, we're already seeing a high degree of ROI in the sense that there's now hundreds of millions of people who are using the technology. It's actually in their hands. It's part of their day-to-day. And so that certainly is ROI. And with what Cohere focuses on, the enterprise side, we're starting to see this technology get into the hands of employees and get into enterprises. It's a much slower process. It's a bigger lift. You have to integrate with existing systems within enterprises. You need to train employees on how to use this technology. But that's well underway now. And we're seeing quite dramatic growth in adoption.

3:12So I think we will find ROI and it's coming soon. And we'll talk more about the specifics of Cohere and ROI in the second half. But let's keep on this line because it goes to another one of these myths and facts, which is that the next set of models are going to be this godlike set of models. And, you know, you talked about how there's going to be like a lot of cost at the beginning. Right. And that's necessary cost to train these massive models. and the sense that I've gotten from my reporting, one of the things I've heard is companies have been willing to make those investments because they think that this next 12 or 18 months in model development is crucial and the capabilities will advance significantly as they put more compute data and energy into the process.

4:00So let's go to this myth effect, number two, which is, does the next set of models give us that godlike AI model? And so many people are expecting. I don't know about godlike. I don't think I'd ever use that term to describe what's coming. I think we're going to have some really powerful and useful tools emerge. I think that's what's coming. The idea that we're building AGI or something that's just going to solve all our problems for us, I think we need to set that aside. We'll get to that. But let me ask you more pointedly on this next generation of models. Okay, so you say we're going to see some new tools.

4:39What does the next generation? Because they are being trained on much more resources than they have been previously. So what tangible step forwards are you expecting to see from the next generation of models? I think the reliability and trust factor is huge. And also just the competency, right? The accuracy with which it gives you answers. And so all of those are going to increase. I don't see a step change coming, but I see a steady, continuous course towards very high accuracy, very high reliability AI. There's been so much hype in the industry. This is one of the things that sort of comes along in this discussion, which is like everybody I speak to who's on the ground says, yeah, we're not going to see a step change with like, let's say GPT-5, but exactly as you described, more reliability, more consistency.

5:33is there i mean is there a worry that some of the air is going to come out of the you know this ai moment if you know because again like i when i say godlike i'm not i don't believe that's going to happen but i'm reflecting what a lot of the hype is starting to expect and so if it's just steady you know steady improvements and reliability which is actually like you know we both agree pretty big but do you think that that sort of takes some of the steam out of this moment for ai because people will look at the step change as a failure given where the hype is. Well, listen, I'm not one of the people who's saying we're going to be building God-like AI.

6:14Yeah, I don't have much to say towards those claims. What I would say is that even if, just as a hypothetical, even if the technology froze and what we have today is all we get, there's so much good to be done. There is so much work to go to, to implement this technology across the economy, really boost productivity, drive better outcomes, build tools. So the technology does not have to move in order for incredible value to be realized. We just need to go do it. And it takes a lot of effort and time and work to go realize that value. okay and again more of that is coming in the second half where we go a little bit more tangible but let's stay with the theoretical or at least like the industry stuff what do you make of the fact that the gpus so we talk about like the ingredients again this is coming from your paper right the ingredients that are required for these models to get better they need data they need compute need energy and the compute right now is starting to go like through the roof in terms of the amount of compute that's being used to train models.

7:27So just for some context, so Meta's Lama 3 model, which was like state of the art, like 10 minutes ago, it used 16 ,000 GPUs to train that one. Now we're hearing that Elon Musk is building this super cluster, I think it's called Project Memphis, that has 100 ,000 GPUs. So multiples of what the cutting edge is being used to train on. So I'm curious if what you think that increase in GPUs are going to get us first and foremost, and then we'll talk about whether the right way to scale these models is with just throwing more compute, because I know you have a nuanced take on that. But like first and foremost, like if you go from 16 ,000 GPUs at the state of the art to 100 ,000, what do you think that delivers?

8:17It definitely delivers a bigger and better model. You have more compute. We know that scaling up improves things. There's questions around saturation and whether continuing to scale up is justified, whether there's going to be enough gains from that strategy to justify the increasing cost. My personal perspective is that building a massive model, it's not actually useful for the world if it's too big to be consumed, if it's too expensive to actually deploy. And so for Cohere, we've been very focused on building the right size of model. But if your question is, what is more compute unlock, it will be a better model.

8:57Objectively, we know that scaling leads to more capability, a smarter model that's more reliable. And so that's the output. And does that ever end? I mean, that's one of the big debates here is that, you know, basically, you could add compute and data basically to infinity and it will continue to improve? Or is there a tipping, you know, sort of a saturation point? I don't think within any achievable scaling up for humanity that we'll reach that tipping point. It just saturates. The gains become much, much smaller. And so you're much less willing to want to pay double the price for a minute difference.

9:38but it is pretty consistent that bigger is better and that just continues but it tapers off over time i mean open ai has talked about how like their goal is to build agi a lot of people in the industry talk about agi's north star i know cohere is more like let's make this practical for businesses but i want to get your sense because because that's not your north star i think you can speak a little bit more about like more honestly about what it means and whether it's achievable. So do you think that, let's just use this definition of AGI as intelligence that's as capable as humans in the tasks that humans do.

10:15Do you think that that is something that we should even be thinking about? Or is it a marketing tool? And is it achievable? I mean, with that definition of AGI, I think it's both achievable and a fairly reasonable target. So we can measure how good humans are in any particular task. And then, yeah, I mean, it's a reasonable goal to want to create technology that can perform that well in that task. So I think based on that definition, I think when we start to think about, you know, you described it as like godlike models, these are being described as well beyond human capabilities. So yes, I think my definition is probably artificial general intelligence.

10:59And I think that this godlike is the super intelligence thing that a lot of and I think a lot of people will use AGI as a synonym for a super intelligence, which seems wrong to me. But there is this belief that once we hit AGI, we've already reached super intelligence because if it can do everything that humans can do and doesn't get tired, doesn't need to sleep, doesn't need to get paid necessarily, you're already at super intelligence. But sorry, go ahead. Yeah. But no, I think that that definition of AGI is a reasonable one. I think it is exciting. I think that that's that's definitely the target What we want to do is we want to create machines that have this unique property that humans have of intelligence.

11:37And we want to be able to deploy them in the places to take work off of the shoulders of people and put it onto these machines to make work better and easier. And in order for you to do that, in order for you to trust the machine enough to shift that work over, it better be as good as the human. Otherwise, you're paying some price, you're reducing inaccuracy, things get worse, not better. And so that's a very reasonable objective. And when do you think we might reach it? In many respects, we're already there in many fields. The models are as good. We're still at work. Yeah, but I don't think those two things are in conflict.

12:24Really? Why not? Well, because I think that we will never see mass unemployment of humans. I think that this technology is going to unlock more opportunities. It will let us do more as opposed to scaling back what we do. humanity is very supply side constrained not demand side we always we want more we want better we want to be healthier we want to do more we want to have things be cheaper and so we have all this demand and we're trying to keep up with our own society's demand and this technology it's it's true promise is in bringing productivity and letting us do more now you can zoom in and you can like pick a specific field and you can say this field might be automated by ai and i think that's true and you know we should be thinking about retraining and shifting certain skill sets over to other uh new domains like retraining people but in general at the macro scale i think this technology will create much more opportunity uh than it will take away i mean if we have ai technology that can basically do work for us, whether it's knowledge work or whatever, right?

13:44I mean, we already have a lot of technology that can automate, you know, factory work. Why are we continuing to work? It brings purpose and meaning to a lot of lives and we enjoy it. I think that the right form of work is something that fulfills you and that is enjoyable, intellectually interesting, compelling. and that's really what I want to spend my time doing is that as opposed to number crunching or um and maybe someone else enjoys that crunching but uh for me I'd rather outsource that to my my excel spreadsheet right so yeah I I think that work in its best form is incredibly fulfilling and that's never that's something that humans will never give up we'll always want to do that But if we can hand off and if we can have an assistant that is, you know, on 24-7 and has access to all the information and tools that I have access to, and I can ask it to do things for me, that's a very compelling value proposition.

14:42It changes work in a way that is extremely positive, I think, for almost everyone. How far away do you think we are from having reliable assistants? Like a lot of people looked at OpenAI's 01 reasoning model and they're like, oh, this is just kind of like a step toward assistive AI. What do you think? I think the notion of using reasoning or letting the model have an inner monologue to work through problems, think through them, make mistakes, but then realize that, catch mistakes and correct them. I think that's a crucial piece in improving not only the accuracy or robustness or usefulness of the model, but also the trust in the model.

15:31Because you're able to inspect how it arrived at its conclusions, how it decided to do what it did, you actually trust it much more. It's explicitly written out. and so i think we've all known these sorts of tools would need to emerge and yeah i think it is a big step towards dramatically more reliable assistance ones that you can trust and work with and give feedback to i think it's really exciting okay and then where does that put you on the fear around ai i mean if ai can sort of go step by step figure out these processes realize where it went wrong go back take action right i think that's sort of where people get weirded out is when these things start to take action on their own, what can they do that we're not prepared for?

16:14So what do you think about that? Are you worried AI might cause harm to people? I think it's really important to remember that we get to choose where we deploy models. It's not like they get to choose where they work or what they have access to. We have to plug them in and we have the opportunity to implement safeguards. So to make sure that before these models are put in any very high stakes situations, that there there's oversight that a human has to approve high stake actions. It's not carte blanche and the model is now smart and we just plug it into everything and say, go at it. Uh, it's very much intentional and we're going to need to be thoughtful and careful about that.

16:58So I'm not scared of like a doomsday, like a Terminator scenario. I think that media has certainly instilled that with lots of sci-fi stories and it's a very compelling story, which is why well before AI was remotely competent, we were coming up with stories about how this might happen. Yeah, I think it's not just media, right? It's like also AI leaders are saying, you know, how many people signed that statement that said we should be treating AI risk the same way we treat climate change and nuclear? Why do you think there's so many people in the industry that are stirring up the fear around this stuff?

17:36I think that's a great that's a great question to ask them uh I did not sign that letter and so so it it puzzles you as well yeah yeah I mean I'm I'm empathetic to the fears because yeah like this is a very salient story that's why it's been so popular in sci-fi and and all this sort of stuff so I'm actually understanding of why people are so attached to those stories but as more and more evidence emerges that these models are much more controllable than we may have thought that they're a little bit less capable than we may have thought. It's harder and harder to make that narrative. And I think you see the discourse shifting now.

18:15I think the discourse has begun to shift away from doom and existential risk. And now it's much more about practical concerns, which I'm really happy to see stuff like, okay, this technology could be really useful for healthcare, but it could also cause harm if we don't do it the right way. And so specifically, how do we set up the safeguards to make sure that harm doesn't happen? Same thing with finance, right? And like distributing loans or something like that, or with people using them maliciously to pretend to be human and trick people. How do we prevent those things? That discourse is super productive, but that's very effective.

18:51And so things are shifting in that direction now. And I'm excited to see that change. Now, some of this fear comes from this line of people saying, oh, there are emergent behaviors in the models, right? That basically that they've found them able to sort of come up with things that are outside of their training set. And there's been some papers that say, okay, actually they don't really have any emergent properties or emergent behaviors. And as someone who wrote the paper that kicked this all off, what do you think about that can LLMs have any emergent behavior or discoveries that they weren't trained on?

19:27I think that they can, what's the right word? I think they can interpolate between skills. And so if they've seen how to do A and they've seen how to do B, they can get kind of the average of A and B, but they don't just go completely beyond anything that they've seen. I've never seen a model behave in a totally unexplainable way. They're really good interpolators. If you show them different domains. They can blend domains quite well. And, but yeah, I've heard the same thing about emergent behaviors. And I think the research is really inconclusive there. Uh, there's not a lot of compelling evidence that says we're going to have some total step change or capability take off even in the, the, like the latest state of the art research, a lot of it's about synthetic data and models teaching themselves.

20:16And so self-improvement is this notion of can a model actually teach itself without human intervention. This is now a huge part of model building. It's a big part of how we create data at Cohere. And before this started to become mainstream and actually part of the production process of creating these models, people were saying self-improvement. These things are just going to take off. They're going to become superhuman overnight and we won't be able to control it. Well, it turns out that doesn't actually happen. Right. So this intelligence explosion or intelligence takeoff is not something that happens.

20:49No, it's not happening. It's not happening. It improves for, it can self-improve for a while and it tapers off. And so, yeah, you get some good improvement out of it, which is why we use it. But then it plateaus. It doesn't just keep going forever. And so I think the evidence points firmly in the direction of a lot of those fears may have been misled. Now, talk a little bit about that. It's interesting you bring up the use of synthetic data and having the machine self-improve. Because one of the big questions about whether this plateaus is, you know, does the world run out of data to train the AI?

21:22And I was watching one of your recent interviews where you talked about how, you know, back in the day you could run up to anybody and they can add knowledge to a model. But as the models got smarter and smarter, it became less easy for people to add supplemental knowledge to them, which points to like sort of running out of available data to make these AI models smarter. So how does like AI generated data actually solve that problem? And where is synthetic data being used to make these models better? Yeah. So I think the example you gave is a good one, like where it's getting harder and harder to get the data that incrementally improves the model.

22:02And it's important to note that that's because the model is getting so much better. And so before we could just grab anyone off the street and they could teach them all something. And then that signal started to go away. And so we had to go to undergrad students in bio to teach the model about bio. And then we had to go to master's students and then PhDs. And we're kind of at that level where we're currently hiring PhDs to teach the model in their specific domain. But then after PhDs, where do you go, right? Like you, I guess, professors. What about after that? So I think the models are catching up with the state of knowledge across a bunch of different fields.

22:40I would say that synthetic data probably doesn't get us out of that that issue i actually i don't know if synthetic data outside of easily um verifiable domains like math it's hard to use synthetic data to drive outcomes so we'll be able to do it in so how is it being useful for you not not for making our models um fantastic philosophers or um making them fantastic social scientists or something like that for that we we rely on humans what we do use synthetic data for is for crafting how the model responds to stuff and in domains that are verifiable like math like coding in those places it's actually quite effective but that's still a huge domain of interest for people building and deploying these models we want them to be good at math and computer science and so more and more synthetic data is becoming a huge chunk of the data that we train on.

23:43Okay, fascinating. One last part of this discussion is sort of what methods help get this AI to improve. And there's been a question of whether LLMs can take it like all the way, or that you need to combine LLMs with different forms of training, whether that's reinforcement learning, I guess that's part of it already. But the other side of it is, do you have to like build world models with like robots going out in the real world and learning things like things like gravity and what happens when you bump into things which you just can't convey in text? So I'm curious if you think the current methods are able to get this field to the promised land or whether they need to be combined with others.

24:24There's definitely proof points out there which suggest large language models or like the transformer architecture is capable of handling a bunch of different modalities. And so you can merge not just text, but video and audio as well into the model. So you can give them a much more balanced experience of the world. You can show them the world. You can show them videos that demonstrate physics. You can let them see, hear, speak. And so as a platform, it does seem like this is a pretty good platform as far as they go. There's a more philosophical argument which is had among academics around is text enough or even is supervised learning enough is it enough for the model just to observe the world or does it need to take part in the world to really understand it for instance would you understand the world if you read all of the internet and you watched every video on youtube would you really understand it or do you need to actually be embodied be a little robot out there kicking a ball or, you know, running down the street.

25:32I actually take, I think the less popular view, which is the internet is enough. And by observation, you can actually learn enough to be extremely, extremely compelling. I think that's, if we're talking about AGI and doing things as well as humans do, I think that's enough. All right. I want to take a quick break, hear from our sponsor, come back, talk about ROI, and then just talk a little bit about your journey, Aiden, from being somebody who wrote that paper to where we are today. I think it'll be interesting for listeners. So we'll be back right after this. Did you know your credit card points and miles can lose value to inflation?

26:11Credit card companies often reduce the redemption value of your points and miles. Now, imagine a credit card with rewards that can grow in value. With the Gemini credit card, you can earn Bitcoin or one of over 50 other cryptos instantly with no annual fee. Every swipe at the store or gas pump earns you instant rewards deposited straight to your account. Plus, sign up now for a$200 Bitcoin bonus to kickstart your rewards. Visit Gemini.com slash card today. Check out the link in the description for more information on rates. Again, if you're looking to invest in Bitcoin but don't know where to start, the Gemini credit card makes it easy.

26:50The Gemini credit card is issued by WebBank in In order to qualify for the$200 crypto intro bonus, you must spend$3 ,000 in your first 90 days. Some exclusions apply to instant rewards in which rewards are deposited when the transaction posts. This content is not investment advice and trading crypto involves risk. The Gemini credit card cannot be used to make gambling related purchases.

27:14You're used to hearing my voice on the world bringing you interviews from around the globe. And you hear me reporting environment and climate news. I'm Carolyn Beeler. And I'm Marco Werman. We're now with you hosting The World Together. More global journalism with a fresh new sound. Listen to The World on your local public radio station and wherever you find your podcasts.

27:42And we're back here on Big Technology Podcast with Aiden Gomez. He's the CEO of Cohere, also the co-author of The Attention Is All You Need, paper that kicked this entire generative AI moment off, right? Invented the transformer. Before we get deeply into ROI, Aiden, just a personal question for you. I mean, are you, what does it feel like having seen, what does it feel like seeing your invention being taken in all these like wild directions and sort of being this key moment in a truly like step forward for the tech field? I mean, it's like beyond my wildest dreams. I think I don't take full credit for it at all.