고품질 AI 제품 구축: 무스타파의 인사이트

선도적인 AI 선구자인 무스타파는 가치 창출을 위해 고품질 데이터 수집과 AI 모델 미세 조정의 중요성을 강조합니다.

### 무스타파 술레이만: 미래를 위한 AI 우선 제품 구축
* 🚀 Microsoft의 AI CEO이자 AI 분야의 선도적인 개척자인 무스타파 술레이만은 중요한 문제를 해결하고 세상을 더 나은 곳으로 만들 수 있는 **인공 일반 지능(AGI)을 만들겠다는 사명**을 가지고 DeepMind를 공동 설립했습니다.
* 🧠 딥마인드의 공동 창립자 셰인 레그는 지능을 **"다양한 환경에서 잘 수행할 수 있는 능력"**으로 정의하며 일반성을 강조했고, 무스타파는 **AGI를 향한 탐구는 우리를 끊임없이 전진**하게 하는 움직이는 목표라고 믿습니다.
* 💡 고품질의 AI 모델을 구축하기 위해서는 **대규모 모델과 컴퓨팅**에서 고품질 데이터에 투자하는 것으로 초점이 이동하고 있으며, 스타트업의 경우 **미세 조정을 위해 가치 있는 데이터를 자연스럽게 수집하는 UI**를 만드는 것이 매우 중요합니다.
* 📚 무스타파는 모호함과 복수의 정답이 핵심인 부정확성을 미덕으로 삼는 문제 영역에 집중하고 **빠르게 수익화할 수 있는 영역**을 찾을 것을 제안합니다.
* 🤖 무스타파는 향후 몇 년 안에 **AI가 사람들의 삶에서 항상 존재하는 조력자가 되어 기술과 상호작용하는 방식을 바꾸고 기존 인터페이스를 뛰어넘을 것**으로 예상하고 있습니다.
* 📈 Microsoft에서 무스타파의 주요 목표는 **메모리와 개인화에 집중**하여 Microsoft Edge 및 Surface 디바이스에 배포된 AI 모델인 Co-Pilot의 품질을 향상하는 것입니다.
* 🛠 좁은 범위의 자율 액션 에이전트를 구축할 때는 컨텍스트에 따라 **쿼리를 적절한 모델로 리디렉션**할 수 있는 라우터 또는 분류기를 설계하는 것이 관건입니다.
* 🎙 무스타파는 소비자를 대상으로 하는 AI 제품의 경우 UI가 **"방해가 되지 않아야 한다"**고 생각하며, **음성 우선 인터페이스가 미래 UI의 핵심 요소**라고 보고 있습니다.

we're not going to have gbt 5 tomorrow and suddenly gbt 5 replaces all humans as the ultimate judge or teacher. You can get an 80% prototype going and it can look good, but a real consumer experience requires you to nail the 99 percentile experience. It has to be really high quality consistently, and as soon as the AI breaks out character, gets something wrong, that sort of destroys the illusion and it breaks trust, so then you lose your consumer. And so from a startup perspective, I think the real trick is finding existing data sources, or more importantly creating a UI that allows you to collect high quality data from an interaction with a product domain that you think is valuable, that is producing a mo of Highly valuable data that you can then use to Post train and fine-tune your model again and get into that feedback loop that is a path to creating an enormous amount of value, which I think is such a creative time in entrepreneur La MFA. Thanks so much for joining. Hey Seth, good to see you again. Uh, thanks for

having me so very excited to have Mustafa here. As many people in the audience will know, Mustafa is one of the leading pioneers in AI. He is currently the CEO of AI at Microsoft, venture partner at Greylock, former co-founder of Inflection AI, as well as co-founder of DeepMind, which was of course acquired by Google where he became VP of AI at Google. So Mustafa, I feel very privileged to be able to work with you. Um, you know, through Greylock, but also to have you here to kind of share your take on the current space, which is evolving so quickly. So maybe to kick it off, I'm curious to hear more about how you decided to focus your career on AI before it was obvious. It's kind of strange looking back on it. I sort of had to write my TED Talk recently, and I reflected on how crazy things are, you know, 15 years later when we first started talking about founding DeepMind in 2010. It's hard to overstate how

weird. We were like, it, you know, often people say being an entrepreneur is, you know, choosing to obsessively work on something that is really contrarian and tempting, something that everybody else thinks is impossible. I think in our case, people didn't just think it was impossible, they thought it was completely absurd. And I, I honestly am not quite sure how we came to be to have so much faith in our ability to try to do something so out of distribution and exceptional. What was kind of the initial insight around DeepMind? I mean, we didn't just start working on AI or machine learning, we were fully committed, uh, to working on artificial general intelligence, you know, producing a system that could exceed human capabilities and knowledge at all, uh, levels. And the reason we were motivated to do that is because we genuinely wanted to use AI to solve other problems and to make the world a better place. I think there was no academic lab or setting that could, you know, accommodate the scale of investment that we thought was necessary at that time. In academic labs, there

wasn't a focus on large-scale engineering. There certainly wasn't a focus on products. And um, even in the kind of sort of big National project Investments that you get from like working in, you know, a government, there was just nothing that was like a technical effort to try to understand what intelligence was at scale. And U, you know, deploy it at important problems. And so it was really only like startups that were as Brave and courageous, I think, as was necessary to be successful on this Mission. So it was just always obvious that I would do another company at that point. Yeah, um, like by that point, that was my third effort at a company, and it just felt like this is the only vehicle I kind of because you learned so much along the way by screwing it up and figuring out how to do it right. And I had come off the back of a bunch of years working in the nonprofit sector in government in conflict resolution and facilitation, had started two small companies, um, actually one that was focused on selling networking equipment and electronic Point of Sales.

Systems to restaurants, but this was like way before that was possible. That was a failure. Um, and the thing I sort of realized is that we need more knowledge and insight in our world to help us address the overwhelming complexity of our systems. Um, it's so difficult to make an intervention into a complex social system today, like an economy or, uh, you know, food production system or the financial system, and be confident that that intervention is going to make the impact that you think it is. And that's really, I think, one of the reasons why we need amazing AI. We need to be able to make good predictions about complexity in our world in order to create value and to change that world, to, you know, help people live healthier and better lives. Um, it sounds super cheesy, but like that is really what was motivating us back then and still don't. And I'm curious, so you had this kind of mission that unlocking abundant intelligence for the world would actually solve problems that matter. How do you actually define intelligence in this context? The thing that gave me

confidence that we might actually be able to make progress towards inventing intelligence was our third co-founder Shane leg had spent his entire PhD um researching various definitions of intelligence and trying to aggregate them into a single metric um that we could use to turn the science of intelligence and the Neuroscience of understanding biological intelligence into an engineering effort and really make that a measurable kind of quantifiable exercise. The definition that he came up with was the ability to perform well across a wide range of environment so again emphasizing generality and that was like a major now everyone takes the AGI part for granted as though G is the central part of intelligence but that's an assumption you know generality happens to be one of the characteristics of intelligence but it's not the only important characteristic and um it also turns out that it's also very hard to measure uh and to scale it down to something that you can really grasp whereas another sort of definition was the Turing test of course is that a system would be intelligent if it could deceive a human.

into thinking that it was itself a human during natural conversation, and in some ways, we've sort of crossed the threshold of that intelligence, right? Like we have systems now that are really very good at conversation, and at least for a few turns, could, you know, it's certainly better than humans in many respects, right? You can still tell that it's an AI or a chatbot and not a human, but in a few years' time, you really won't be able to tell. And yet, that doesn't really tell us anything about whether these systems are actually intelligent. Every time we cross a threshold in terms of a benchmark or a milestone in AI, you then sort of turn around and go, Okay, well, here are all the problems with that mechanism of measurement, and here's the next sort of thing that we need to measure. I think I've heard Reed say that AGI is the AI that we don't yet have. It's kind of always push, push forward in the future, exactly, exactly. And it's like the constant carrot that we dangle.

to kind of chase ahead, you know, sort of move ahead. So another measure that I have proposed is, you know, that we should focus more on the capabilities of the system, the actions, the things that it can do, the things that we can observe that it can have an impact in some environment, rather than this kind of abstract idea of, is it general or is it good at having a conversation, you know. So, you know, that would basically be like, can it produce, you know, human quality labor in a practical environment and actually earn money for that, or like, could it write software, for example. It's a very measurable thing. So that, you know, I sort of called that a modern Turing test and said that in the next five years, a system would be able to take a very abstract goal like, go create a new product, get that sort of designed, manufactured, drop-shipped, and then distributed and marketed, and try and earn a profit from it. And then you could measure that profit in terms of making a million dollars or something like that. I think it's going to be a system

that can actually do that certainly before 2030. Yeah, that's amazing and it's possible. You expect that that type of system would actually trade off on the generalizability, and it's actually, you know, built for a specific use case. Yeah, I mean, I think that it's more likely that we'll have really powerful systems that are specialized for specific use cases that have real deep domain expertise. Um, you know, then we're going to have this sort of very general-purpose system that can switch from being a marketer to a, you know, clinician to being a doctor, being a lawyer, whatever. I mean, you know, clearly, those, the general case is going to come afterwards. So I want to spend a moment just getting your take of just the state of large models today and maybe you wanted to start with just level setting for the audience of, you know, what was the inflection point that led us to the current state of, you know, the GPT3, GPT4-style models, the inflection 2.5-style models, this combination of the

Transformer architecture and scaled compute. I think that this revolution has been driven by Deep learning. We are still building deep learning models, albeit with a slightly different flavor. Now, the Transformer architecture from 2017, and you know, we're now turning those into composable units that are essentially going to act like parts of our software development ecosystem. You know, you're just, you're just going to turn to your AI, and it's going to actually generate code for you. I, we're already sort of seeing that, Gab co-pilot, and you know, we're seeing it like as a member of your team, able to take a natural language instruction and just of act in unison with you. And I think that people sort of don't quite realize is that these models are not going to be large forever. Um, in the history of all technologies that have been valuable, anything that is significant gets cheaper and easier to use over time, and that curve is sort of double exponential over the last like couple years. It's, it's incredible. I mean, we like Microsoft AI just released F 3.

fully open source. It is, um, close to but not quite at GPT-4 level. Um, it's fully open source. It's 3.8 billion parameters, right? So, that's like more than 100x smaller in terms of inference compute than basically the absolute frontier of models today. Like I said, it's not quite as good, but it's certainly, you know, as good as or better than GPT-3.5. Um, that's mind-blowing. I mean, that's something that can fit on your laptop or on a phone. Like, I mean, the future. And so, we should expect that trajectory to continue. Um, I think open-source models are going to be very close behind the closed-source proprietary API models. Um, months or maybe even just a year or a year and a half or something. And that's just going to change the whole creation landscape, basically. That's super interesting. And what enabled this model to be kind of almost equal in performance while also being much smaller? Well, so for the last couple of years, everybody has been

focused on reinforcement learning from human feedback. In the final stages of training, um at the stage of fine-tuning or post-training, as people say, you have a bunch of um trained rers or judges that compare two possible responses or completions from a model, and that pairwise comparison provides large-scale feedback for the kinds of behaviors that you want your model to exhibit. Everyone's familiar with that now, but what we've, you know, been focused on as soon as that was like showing promising signs. So for the last, like, 18-24 months, is reinforcement learning from AI feedback, where, you know, we really want very smart and capable models to do that pairwise comparison, um, because obviously we can automate that process, and we can produce an even larger number of supervised fine-tuned labels, um, to give more feedback to the pre-trained model across a wider range of experiences and moments that might be in tension with one another if you only have a small number of samples that come from expensive highly trained humans. So that was one method is reinforcement

learning from AI feedback, and then the second is generating training data from these models. So, you know, sometimes people refer to that as distillation, you know, where you're sort of trying to absorb as much of the best bits of a big powerful model as possible, um, and then you're using that to post-train, uh, or sort of align your smaller model. And parameter count is no longer the primary propy for capability. High-quality data is the real valuable asset here, in addition to the architectures. So, you know, for the last sort of eight, you know, six to 12 months, everyone's been focused on compute, compute, compute. Can I get compute? And obviously, that's kind of important, but um, or like, you know, large models, large models, but really it's investing in high-quality data. And so from a startup perspective, I think the real trick is finding existing data sources or create, or more importantly, creating a UI that allows you to collect high-quality data from an interaction with a product domain that you think is valuable that is producing a mo of high.

highly valuable data that you can then use to post-rain and fine-tune your model again, and get into that feedback loop that is a path to creating an enormous amount of value. And it does not require you to depend on the large-scale model providers, which I think is a great, you know, that's why it's such a creative time in entrepreneur. Yeah, that's super interesting. As a startup, you're competing with incumbents that have access to large sets of data, so I'm curious if you could share anything more about the nuances of opportunities that might exist for startups to get specific types of data that's maybe more valuable than others. Okay, so how do you go about collecting high-quality data because people have slightly, obviously, in pre-training, it's about volume of token and there the hyperscalers will have a longstanding advantage because they already own search engines or YouTube or whatever it is, right? Whereas in post-training, you need a small number of very high-quality tokens to, you know

align the model to the behavior that you want uh for your product, and you can collect that from scratch. Like when we built Pi, we created, and to this day, the most high-quality human conversational AI with the best EQ even today in the market, right? Yeah, and we didn't use any data from Big providers. We collected all of it ourselves from scratch by training um paid teachers. We call them AI teachers. Some people call them Raiders, but the crucial thing for a startup is you have to really, really, really pay attention to training those teachers. You have to pay them a lot of money. Um, like I'll just tell you from our perspective, we selected people who had an undergraduate education, nothing less, that largely spoke English as a first language, with some exceptions, that had a domain expertise that we thought was valuable, like maybe they said they were very passionate about history or they like, you know, had a good kind of cultural knowledge or they were movie Buffs or whatever it is. Um, they had to pass 20 hours of training.

testing by us so we would give them reading and comprehension exams. We would give them multiple choice questions. They would have to do sentence completions. They'd have to do spot the differences. They would have to do like really quite hard analytical tasks. And in order to just keep my team super humble about how valuable this task was, I would obviously also have all my team go through the same training, right, and go through the same tests. And I can tell you not even a majority of people par, yeah, I I was about to say I'm nervous. I said s over the test, yeah, you know, it's actually not easy. It's a pretty, it's a pretty tough thing to do because you're, if you think about the task, you're asking a human to read through two 10 turn conversations, look at the proposed answer by one model and by another model, and then absorb a huge behavior policy, like very detailed line by line. The AI should do X, shouldn't do Y in this situation, should do this, and then remember the training from the AI.

teacher that says you know all kinds of like subtle exceptions and stylistic tones, and brand things, and backstory things, and capability awareness, and then you have to sort of find the correct intersection of all of those to decide, is this paragraph more in line with the behavior policy, or is this one more in line, yeah, yeah, it's a painful talk that's super interesting. I'm curious how you view this evolving, right, because you mentioned we're kind of moving from reinforcement learning just from humans but also from AI. So how do you view kind of application layer startups between being vertically integrated versus which parts of the stack they really need to be experts in? Yeah, that's a good question. I mean, I think you have to be very principled in answering that question, and that's where the bet is of your startup. You have to decide which part of it am I going to bet on. Obviously, a bunch of people are building tooling and infrastructure, and that's fine. We all are familiar with that kind of, you know, strategy. I'm a great

believer in you know building and owning your own product, and as much as possible controlling the key bit of the value there, which in my opinion is the llm, and everything around that is secondary. The words that come out of the llm are what you have to focus on, and that means that you know I think it's reasonable to break off the pre-train model and get that from somebody else. Yeah, that's a good approach. Um, but I think you need to own your fine-tuning stack, and I would not give the fine-tuning stuff to somebody else. You have to train your teachers because that's not going to go away anytime soon. We're not going to have gbt 5 tomorrow, and suddenly gbt 5 replaces all humans as the ultimate judge or teacher. I think that's quite unlikely. It's going to be a lot better than gbt 4, but even the people that have tried to do RL aif with gbt 4, the quality is okay, and it is impressive.

very cool, but it isn't on the verge of replacing humans entirely. You can get an 80% prototype going, and it can look good, but a real consumer experience requires you to nail the 99th percentile experience. It has to be really high quality consistently, and as soon as the AI, you know, breaks, AI character gets something wrong, has a hallucination, whatever you want to call it, that sort of destroys the illusion, and it breaks trust. So then you lose your consumer, and so I think that the key thing for the startups, at least for the next year, is to get really good at data collection and data filtering and data quality that makes sense. I'm curious, what are the different UIs that you think kind of AI-first companies are going to be building, whether it's, you know, a chatbot, an agent, just regular SaaS that's kind of enabled by AI? Yeah, I mean, I think that, in my opinion, the UI needs to get out of the way, especially for the consumer. Obviously for SaaS, you can have all the bells and whistles

all the developer features, right? Blah, blah, blah. But for a consumer, the goal is to get the UI out of the way. So we create a very pared-back, quiet, soft, calming, you know, I think quite distinct looking AI with very, you know, UI with very few buttons. But we also have, uh, one of the best voices in the world. Um, you know, we, we got, you know, in the end, I think we had nine or 10 voices. Um, they were really high quality and very, very human-like. I mean, P's still alive and should try it out, but it's like that. I think voice first is a big part of the future UI. Yeah, I like the where you have to choose the voice that you like as part of the onboarding experience with P. Yeah, yeah, because it's sort of like connecting with your AI. It's a personalization moment and, um, 30% of all our conversations took place on voice. Um, and they were by far the longest, most engaged, most retained users. People should keep that in mind. It's

an very important Insight. Yeah, that's super interesting. I've heard you talk about like AI has IQ, they have EQ, and then you've talked about AQ, right? An action quotient. So I feel like this is a very interesting topical thing at the moment where people are talk about these kind of autonomous agents for certain use cases, um, which includes reasoning and planning. I'm curious, um, how far away are we from the kind of chat Bots we've experienced, that most people have experienced today, to a fully autonomous agent that you described at the beginning that's capable of, uh, executing an end-to-end task? What's missing between where we are today and, you know, that Vision in, you know, let's say the next six months to three years? So I think, um, first of all, I I don't think we're on a path towards fully autonomous, and I think that's actually quite undesirable. I think fully autonomous is quite dangerous, and you know, I got a lot stick after my TED talk because I said that that the autonomous capability was

dangerous and so on, and that's one that should be regulated, and you know, I don't really care. I mean, I still think that I, I, I think that if you have an agent that can formulate its own plans, come up with its own goals, acquire its own resources, just objectively speak, act completely independently of human. I'm objectively speaking, that is going to be more potentially risky than not. So, I think about it as these like narrow veins of autonomy where you give it a specific goal, and it has limited degrees of freedom to go off and act in some specific environment. So, like making automatically calling some API to check some registry, to collect some information, to observe some State, maybe writing something into a third party, you know, API that is not yours, but is again restricted with some specific degrees of freedom because I think the security risks here are significant. So, yeah, I just think we should tread carefully on the autonomous piece, but in terms of on the actions piece, um, it's still pretty hard to get these models to um.

follow instructions with subtlety and nuance over extended periods of time. I think that they can do it, you know, and there's a lot of cherry-picked examples that are impressive, you know, on Twitter and stuff like that, but to really get it to consistently do it in novel environments is pretty hard. And I think that it's going to be not one, but two orders of magnitude more computation of training the models, um, so not GPT-5, but more like GPT-6 scale models. So I think we're talking about two years before we have systems that can really take action that makes sense. And if you were to break down, I guess, some of the unsolved kind of research or technical problems to get there, well, an action is different really to predicting a sequence of words. So when you ask a model to produce, you know, to complete a sequence of action, let's say it's like three things in, you know, to basically let's say to book, you know, a restaurant that you and I can go to on a

certain day. Yep, the first action would be check the availability in both of our calendars, so that's a correct function call. Reconcile the correct moment, so that's the second action. Make sure that it's a restaurant that has availability, so that check is another one. And then, you know, go and sign in so that you can basically, you know, use the correct tool to book the right restaurant at the right time. Put your credit card details down, um, having obviously also checked that it's a restaurant that we both like, etc. So it's like four or five or six different steps of, you know, just to produce that one, quote unquote, action subcomponents right. In order to get that right, you're basically saying that the model has to produce perfect function calling for each element and do so in sequence. So it can't just be arbitrary, it has to be in sequence. And that's like saying it has to write a four-page document in response to one question that is exactly that document. And it can't be something that is approximate.

or similar to that document. So, we all think that obviously these models are magic at the moment. And they write beautiful poetry and creative copy and text, and give you good answers. And sometimes they're grounded and blah blah blah. But when, when for each one of those answers, there's a wide range of correct answers that it could have picked, right? Tens, hundreds, thousands, maybe. So, it isn't producing a specific perfect every single token that is outputed is correct answer for each one. It's not there yet, right? So, to get that level of precision, we have to scale up these two orders of magnitude. That's what's happened so far. The last, you know, five orders of magnitude of Transformers with every 10x of compute and data, we get more Precision. It's not just emerging capabilities, that's wrong. People say, Oh, it's surprising we had these emerging capabilities. That's an anthromorphic projection. They're not surprising emerging capabilities. It is just more precise attention to the correct mapping between a prompt and an output, right? So, you're

just honing in on something more specific now do you think we'll be able to get narrow forms of actions in a specific domain before we get to GPT-6? Yeah, definitely. I mean, there are some good actions today, right? You can definitely see these orchestrators making good API calls at the right time. The question is, can it do it with 99% accuracy? Because if it does it 80% then, you know, one in five times it's getting it wrong, it's not usable for a consumer, right? So, you either have to constrain the action space so that every time you're asking your model to take an action, it's only got five options to pick from and the consequences of it getting it wrong are low. Or, you know, you have to find a problem domain where four out of five accuracy is acceptable, right? And if you think about the architecture of building one of these agents, right, are there any differences to what we just talked about in terms of focusing on post-training getting the UI?

out of the way, in terms of the architecture of building one of these kind of narrow autonomous action agents. I mean, one thing for people to keep in mind, I think, is that um, there's a whole range of tools in the toolbox now, right? And so the art is in designing a a router or a classifier that can take some given input. It's either contextual information, metadata, or of course, the incoming query from the user, and redirect that to a model that is appropriate to that context, right? And that's important for inference management, uh, inference budget management, because obviously you can serve to, you can, you can redirect the query to um, smaller, cheaper models or higher quality models or models that are specialized in a particular domain that have been fine-tuned for a particular, you know, area of expertise, or indeed have certain capabilities like they might be good at retrieval, they can retrieve from some knowledge base or even from the open web, or they might trigger a model that has been fine-tuned for a voice response, because obviously the length and style of a voice response

would be quite different to one that was producing paragraphs and paragraphs in a traditional sense. So, the router is a critical part of the architecture as well. Yeah, so Mustafa, I'm curious if you were starting your first company today, given everything that exists and you know the crazy rate of change, where would you focus your attention? Um, I would look for problem domains that make a virtue out of imprecision. So, what is a problem where if you solve that problem, the valuable contribution that you've made is that ambiguity, imprecision, multiple possible answers are the key thing. If you pick a problem domain where you know the consequences are really high if you get it wrong and where there's really only like one or two correct answers, then you're mod always going to struggle. So, that's the first thing I would say, look for more of those things. And just on that point, there is a ton of activity in domains that do require more precision like law or accounting or tax. Do you think that is a futile effort at the moment? Actually, law doesn't require as much

precision as taking actions right. So even in law, most of the applications are retrieving similar cases or giving you summaries of cases, right? Or where there are five possible summaries, all of which could be correct, right? Right. Or where you're retrieving one case over another, you don't have to. So the law thing is actually a high-stake domain because the consequences if it gets it wrong are like really wrong, are bad, right? But like unlike, you know, generating marketing copy, but actually there's a lot of right answers in that domain. Medical is much harder, like clearly there are fewer right answers in medical and the consequences are really high. So that's a pretty difficult domain, you know, a bunch of my really good friends from Deep Mind hell um who had now been at Google just released a paper last week, unbelievable, unbelievable work that they have done showing that you know they can basically provide an amazing reasoning engine for clinicians. And I think in the future for patients too, like that that's coming soon. What, what other kind of factors

would you think about so I I would say where you can design an interface, which naturally collects valuable label data for fine-tuning. By virtue of the interface, that's really important because if you're successful, you want to be able to compound that success. Because the more users you get, the more data you get, the more quality, the higher quality model you can produce. And then you get that virtu of cycle. That's a really important, you know, part of it. And then, then it sounds obvious, but like I think a domain that where you can monetize much faster, you know, than you might think. Because you need to get people to pay for it pretty fast because GPUs are wildly expensive, as everyone knows. What's an example of that that comes to mind for you? I think the companies that are doing specialist services for not quite 10,000 fans, you know, but like, you know, 10,000 true fans, but like I mean, like people that really do need that kind of Niche, um, highly adapted, um, expert system in your pocket, you know.

don't know whether it's a mechanic or a dentist or you know someone who's passionate about a certain hobby or a chunk of IP. I think there's value in those kind of those sort of Specialists use case that people be prepared to pay for. Yeah, that makes sense. Curious to hear just a little bit about the products, the AI products you're working on at Microsoft. You know what the portfolio looks like. Yeah, so I'm responsible for Bing, um, uh, Edge, the browser, um, and all of co-pilot, which is deployed in basically every Microsoft Surface now. And it's been an impressive arrival, actually. The quality of products and their scale and reach is sort of much greater than you might think as a kind of default, you know, Silicon Valley person who had grown up in Google. Well, you don't become a three trillion dollar company for nothing. Yeah, but the rep that we give it in Zur and Valley relative what it has is like I think it needs a you know rethink and and also just like huge

scale and distribution. My main goal is to uplevel the quality of co-pilot. And so we're rapidly building, you know, some of the best models in the world, partnering very closely with OpenAI, building on top of all of OpenAI models and infrastructure, fine-tuning their models. And the next phase is that we're really going to start focusing on memory and personalization. I mean, your AI should remember everything about you, all your content, your personal data, or everything that you've said, and be there to support you and be your aid and your sidekick, you know, throughout your life. So that's what we're going to be focused on. That's fascinating. I'm curious, how do you think about the constraint of the existing applications, you know, in Microsoft Office versus the ideal version of co-pilot? Yeah, that's a good question. I mean, yeah, people have often said that like an AI subsumes all other interfaces and surfaces, and I think that probably overstates it, but it's the right direction, you know.

think there will come a time in a few years when the first thing you think is, You just say, 'Hey co-pilot, can you take care of this for me?' What's the answer to that? Where do I find it? Can you book this? Remember that. Buy this. Do this. You're just going to have this ever-present aid in your life. It's going to change what it means to use a keyboard. It's going to change what it feels like to have apps. It'll move us way beyond the search engine and, you know, the browser. And you're certainly not going to think, I need to go write a document or send a message message in a traditional way. You'll still have those things, but your AI will just manage a canvas of activity across your entire life, and largely be coordinating with other AIs and other services and collecting information for you. Yeah, that's fascinating. Okay, well, let's leave it at that. Mustafa, really appreciate you spending some time today and always. Enjoy catching up with you. Yeah, great to see you, Seth. Thanks a lot.

it was fun. Thanks. Ciao. Thanks for listening to product Le AI. You can find more information about today's interview and the entire series on the website productleipod.com. You can subscribe to the show on all major podcast platforms and watch the video version of this interview on YouTube. And if you want to get all the links and details delivered right to you, sign up for my LinkedIn newsletter. I'm Seth Rosenberg, and this is product Le AI. [Music] [Applause]

Mustafa Suleyman on Defining Intelligence

Seth Rosenberg

무스타파의 AI 기업가 정신 여정

무스타파는 기술을 활용해 현실 세계의 문제를 해결하고 긍정적인 영향을 미치고자 하는 열망으로 AI 분야에서 경력을 쌓았습니다.

그가 딥마인드와 함께 인공 일반 지능에 집중하게 된 동기는 학계와 정부 환경에서 부족한 AI 분야의 대규모 엔지니어링 및 제품 중심의 노력에 대한 필요성 때문이었습니다.

AI 기업가로서 무스타파의 사명은 풍부한 지능을 활용하여 복잡한 사회 시스템에서 가치 있는 예측과 개입을 하고 궁극적으로 사람들이 더 건강하고 더 나은 삶을 살 수 있도록 돕는 것이었습니다.

셰인 레그의 연구는 다양한 환경에서 잘 수행할 수 있는 능력을 기반으로 지능을 측정하는 메트릭을 만드는 데 중점을 두었습니다.

지능의 핵심 특성으로서 일반성이라는 개념이 도전받고 있으며, AI 시스템의 기능과 영향력을 측정하는 방향으로의 전환이 제안되고 있습니다.

딥러닝과 트랜스포머 아키텍처에 의해 주도되는 대규모 AI 모델의 개발은 코드를 생성하고 인간 운영자와 함께 작동할 수 있는 잠재력을 가진 전문화되고 구성 가능한 단위로 이어지고 있습니다.

오픈 소스 GPT-4가 AI의 미래에 미치는 영향

38억 개의 파라미터를 가진 오픈 소스 GPT-4는 GPT-4 수준에 근접하고 있으며, AI 환경에 큰 영향을 미칠 것으로 예상됩니다.

이 모델의 성능은 AI 피드백을 통한 강화 학습과 대규모 모델로부터의 학습 데이터 생성을 통해 달성됩니다.

스타트업이 대규모 모델 제공업체와 경쟁에서 성공하기 위한 열쇠는 사후 학습을 위한 고품질 데이터를 수집하고 모델을 미세 조정하는 데 있습니다.

고품질 대화형 AI를 위한 AI 교사 교육

Pi의 고품질 대화형 AI는 처음부터 특정 기준과 광범위한 교육을 받은 AI 교사를 사용하여 구축되었습니다.

이러한 AI 교사의 교육 과정에는 엄격한 테스트와 행동 정책 및 문체 톤에 대한 깊은 이해가 포함되었습니다.

스타트업은 AI 모델의 어느 부분에 베팅할지 결정할 때 원칙을 세워야 하며, 고품질 대화형 AI를 보장하기 위해 사전 학습된 모델을 위한 미세 조정 스택을 소유하는 데 집중해야 합니다.

AI에서 품질의 중요성과 UI의 미래

AI 스타트업은 소비자 신뢰를 구축하고 고객을 잃지 않으려면 고품질의 일관된 99번째 백분위수 경험을 제공하는 데 집중해야 합니다.

AI를 우선시하는 미래의 기업 UI는 개인화된 경험을 위해 고품질의 사람 같은 음성을 강조하면서 미니멀하고 차분한 디자인이 우선시될 것입니다.

완전 자율 AI 에이전트는 잠재적인 보안 위험 때문에 신중하게 접근해야 하며, 대신 특정 목표와 제한된 자유도를 가진 좁은 범위의 자율성에 초점을 맞출 필요가 있습니다.

특정 작업을 수행하도록 GPT-6 스케일 모델을 훈련하려면 GPT-5보다 두 배 이상의 계산이 필요합니다.

이러한 모델이 동작 시퀀스에 대해 정확하고 구체적인 출력을 일관되게 생성하는 데 필요한 정밀도는 여전히 중요한 과제입니다.

소비자 애플리케이션에서 이러한 모델을 사용하려면 좁은 형태의 동작에 대해 99%의 정확도를 달성해야 합니다.

좁은 범위의 자율 액션 에이전트 설계

- 좁은 범위의 자율 작업 에이전트를 설계하려면 다양한 도구를 사용하여 문맥 정보를 기반으로 쿼리를 가장 적합한 모델로 리디렉션하는 작업이 포함됩니다.

- 회사를 시작할 때는 부정확성으로부터 이점을 얻고 가치 있는 라벨링된 데이터를 수집할 수 있는 문제 영역에 집중하는 것이 중요합니다.

- 빠른 수익 창출과 모델 개선을 위한 고품질 데이터를 수집할 수 있는 사용자 축적이 가능한 도메인을 선택하는 것이 중요합니다.

Microsoft AI 제품과 코파일럿의 미래

Microsoft의 AI 포트폴리오는 빠르게 품질을 개선하고 메모리와 개인화에 초점을 맞추고 있는 Bing, Edge 브라우저, Co-Pilot을 포함합니다.

목표는 Co-Pilot이 사용자의 삶에서 항상 함께하는 도우미가 되어 기술과 상호 작용하는 방식을 바꾸는 것입니다.

Co-Pilot이 사용자의 생활 전반에 걸친 다양한 활동을 관리하고 다른 AI 및 서비스와 협력하는 것이 비전입니다.