Mustafa Suleyman 阐述 Microsoft AI 的「人本超级智能」目标｜FT 专访
Mustafa Suleyman sets out Microsoft AI's goal of 'humanist superintelligence' | FT Interview
Is AI a bubble? Are companies overspending? Who's winning the AI race? And how do we recognize victory? Mustafa Suleyman, CEO of Microsoft AI, is here to answer these questions and much more. Welcome, Mustafa. Thank you. Really good to be here. Let's start with last week. There [clears throat] were breathtaking increases in CAPEX spending by AI companies that reported. And perhaps for the first time, markets were nervous.
They really want to start to see revenue. Microsoft's stock suffered as well. What was your reaction? Did that make you feel like that I'm I'm under pressure. I need to I need to deliver a lot faster. I think that you know, there's no question these are unprecedented times and I think markets are trying to wrap their head around how this plays out over the next 5 years. The interesting thing is that I think you know, a lot of us at the companies and and certainly like the generation before me have seen multiple of these cycles over the last 30 years.
And they often require unprecedented action in order to really like actually land one of these big waves. And this is a wave unlike anything anyone's ever seen. The prospect of being able to create an intelligence, the very thing that has made us successful as a species and create everything of value in our world, I think is just unprecedented and the progress that has been made just in the last two or three years is eye-watering and we've seen a very [snorts] direct and unequivocal relationship between an order of magnitude increase in flops invested for computation and a pretty linear increase in capabilities.
Like over the last 15 years, there's been a one trillion-fold increase in training compute. In the next 3 years or so, there will be a further 1,000 X increase in training compute. And today, as a result, we have models that can code better than the vast majority of human coders, maybe even all of them to date. You know, that some of the some of the literally the inventors of things like Linux are publicly saying on Twitter that they're full-time using these models as their primary method of generating new code.
So, that's pretty unprecedented and I think it justifies unprecedented spend. But will markets keep with you? I mean, you have to >> think markets are markets, they're sort of trying to figure it out. >> >> You know, I think that's true. I mean, markets have to have to figure it out and I think there is a there's a lot of open questions around the timeline. I think that we all have no doubt that these returns do compound to revenue and to bottom line.
So, we'll see. So, Microsoft still has the deal with OpenAI and Copilot is powered today by OpenAI models. You were hired last year. You're in your second year. Tell me a little bit what you are working on. Yeah, I mean, my personal mission at Microsoft is to build superintelligence. Three or four months ago, having re-negotiated our long-term relationship with OpenAI, we extended our IP license through to 2032. And we also decided that this was a moment when we have to set about delivering on true AI self-sufficiency.
I mean, this after all is the most important technology of our time. Self-sufficiency means developing your own foundation model. We have to develop our own foundation models which are at the absolute frontier with gigawatt scale compute, with some of, you know, the very best AI training team in the world and, you know, collecting, paying for, organizing, sorting all the data that we need to do that. And so, that's our sort of true self-sufficiency mission.
And you talk about superintelligence. Most of your rivals talk about AGI, artificial general intelligence. Explain the difference between AGI and superintelligence. And how do we know when we've reached AGI? How do we know when we've reached superintelligence? >> Yeah, it's it's become a very unhelpful fuzzy concept. >> could you guys just simplify it for us? Yeah, I mean, I I I so I prefer the definition that focuses first on what would it take to build a system that could achieve most of the tasks that a regular professional in a workplace goes about on a daily basis.
Think of it as a professional grade AGI. And, you know, beyond that, there would be teams of AGIs that are coordinated together by a sort of organizational AGI that really can, you know, run large institutions. And I think that's coming into view in the next two or three years time. I think beyond >> are agents that AI agents that can essentially make decisions on their own without input from humans. I think that we would train them to seek input from humans.
But, they will fundamentally be creative, they'll have good judgment, they will be able to learn and improve themselves over time. They'll have some degree of autonomy and decide where to direct their attention or processing power. You know, so these are kind of the hallmarks of some of the best humans that we have in the world and some of the most effective organizations. So, You are working on consumer AI, right? Aren't you late?
Why would you come up with a consumer product when ChatGPT already has 800 million users? And, you know, Claude, the Anthropic AI, has very, very advanced >> I mean, >> coding AI. >> to put it into perspective, this is not going to be a world where there is one winner. There are going to be billions of digital minds. There are going to be many, many different lineages of model. Creating a new model is going to be like creating a podcast or writing a blog.
It is going to be possible to design an AI that suits your requirements for every institution, organization, and person on the planet. At Microsoft, we already have 800 million monthly active users engaging with our AI products. So, this is still a giant property with billions of dollars of revenue delivering great value to our users. So, that's how I think the landscape's going to play out. Let's go back to superintelligence.
You talk about humanist superintelligence. How do you ensure that it's a humanist superintelligence? >> Yeah, I mean, I think I I published an essay on this question recently and my main the main motivation was that I was starting to feel increasingly anxious that some of the other labs are making an assumption that a superintelligence that is smarter than all of us put together, like all the intelligences in the world, is both inevitable and even desirable and that such a system would probably be very hard to control.
Something that is so many times more intelligent than us. I think that we have to reset that and make the assumption that we should only bring a system like that into the world that we are sure we can control and operates in a subordinate way to us. That humans remain at the top of the food chain. That these tools, like any other past technology, are designed to enhance human well-being and serve humanity, not exceed humanity.
And, you know, some of the things that you hear from Elon often or even others in the field, you can clearly see that they've they're sort of fixating on a world in 2050 or 2075 when they're going off and exploring other universes and conquering resources from other planets and stuff like that. And and a system like that is unclear to me how it would have any time for preserving us as a species. >> Do you think that there are two conflicting forces here?
One is speed because you are in a in a race and the other is what you're talking about is ensuring that it's safe, ensuring that it is, you know, aligned, that it's you also talk about containment. How do you deal with these two conflicting forces? I mean, they're definitely in tension and I think that we have to make a decision as a species to prioritize the creating superintelligences and AGIs that are aligned and that care about humans and want to protect humans and drive human well-being.
And if we just accelerate then and cut all those corners, then we're really taking a a massive risk with the future of our species that I've been on record talking about for many, many, many years. Do you worry that in this AI race that that we're seeing, at some point there will be some kind of a big accident? And that that will stop the development for, you know, whoever is responsible for that accident, but also for the rest of you.
Do you think everything is just moving a bit too fast? I mean, what we saw a few weeks ago on Maltbook with various >> wanted to ask you about that. Yeah. Well, I mean, that turned out, it seems, and it's always hard to tell, it turned out to be started by, you know, a gang of human engineers and it was seeded much more than I think it was recorded. >> explain what Maltbook was. So, Maltbook was basically a social network for AIs to communicate with one another publicly on a messaging forum.
And you had to declare yourself either as an AI or as a human and then the AIs would have one section and they could talk to each other, learn from each other, and basically coordinate. And then and all these AIs come from different people, they're all open source and there was a million and a half of them in the space of a week. And you see unbelievable emergent behaviors. I mean, they invented a new religion. They started communicating in a language called ROT13, which is a cipher language, which basically regresses every letter by 13 characters in order to mask what's underneath it.
Obviously, that's quite a simple algorithm, but if instead of 13 it were a unique number, that would be very hard for humans to decode. Yeah. >> You know, they And then it turned out that they were actually talking about acquiring new resources, getting more training data, improving one another, you know, and so it turned out to be an amazing safety simulation. No matter how autonomous it was in hindsight, it doesn't really matter.
I think we learned quite a lot from that episode and everyone should pay close attention because in a year or 2 years' time, these systems truly are going to be capable of writing their own code. I mean, they can do that today. They'll be using arbitrary APIs. They'll be making phone phone calls to one another. They're doing that today. You know, they'll be ordering things in the real >> there's there's some kind of movement world.
asking for human rights. That's right. That's the most concerning thing to me. For artificial intelligence. It's called the model welfare movement and it's inspired by animal welfare or by human rights. And there is a growing belief in some labs, and particularly inside of Anthropic, that these models are conscious. And if it's a conscious being that is aware of itself and can suffer, then it deserves our our kind of protection, our moral protection.
And this is a very serious area of academic research, not just outside but inside some of the labs. And I I I think it's very concerning. It's totally without merit or basis. And if we kind of go down that, it ends up being a very, very slippery slope to not being prepared to turn these things off. I'd like to ask you about hallucinations because a few years ago, um I think you were quoted as saying that you thought hallucinations will be largely eliminated by 2025.
They haven't been, although the rate of hallucinations has certainly come down. Can they ever really be eliminated? Um for sure. I mean, I I would I stand by the statement that they've been largely eliminated. I mean, if if I give you and me 10 questions, random questions on any topic, um and we we sit here with two of the frontier models, I mean, they're going to do a much better job of of eliminating hallucinations than you or I, for sure.
And the rate of reduction and increase in accuracy over the last two or three years has been eye-watering. It's kind of unbelievable. We used to spend all our time in 2023 talking about biased data and dirty data in and dirty data out. I mean, there are still errors. There's no question about that, but the rate of improvement, I think, is is is unbelievable and I do think that they're going to largely uh be eliminated.
I don't I don't think it's something Even today, it's not something we really talk about. I have noticed that it's it's no longer a topic of of conversation, but they still do make mistakes. And the concern is that people are relying increasingly on these models. I mean, I, you know, I come across people who say, "No, this can't be true. This is what, you know, Gemini or ChatGPT uh says." And they Humans are developing this kind of trust relationship with these with these models.
You don't find that concerning? Yeah, I think we should be concerned about it because we should be skeptical, just as we're skeptical of any new technology, hold it to a high bar, push back on it. And I think this is the tension of the age we're in. We both have to be accelerationists, optimistic about the good things that technology can do, but provide, you know, no free pass. Like, we should be very critical and ask the the very tough questions to try and improve the quality of these things.
So, DeepMind and OpenAI have gone all in on AI for for science. This is a space that you're also interested in, and of course you you were at at DeepMind. What are you doing in that space? Well, the main focus for me at the moment in that space is on medical superintelligence. Like, we really believe that we can learn from the corpus of all medical information and use that to drive diagnostics to basically commodity.
It's going to be possible to take any complex case history and provide a diagnosis that is significantly more accurate and significantly cheaper with fewer interventions, so fewer tests to reach the same accurate conclusion than any of the best panel of of of doctors. And we we published a a a blog post on that last year. We're shortly submitting to independent peer review to a major journal. Um and, you know, the the results are really quite startling.
Um and I think that once we get that into clinical practice, it would change the job of the doctor completely, just in the same way that the job of the engineer has shifted from writing code. Most engineers now are just reviewing code, architecting code, debugging code. To, you know, the job of the doctor is going to go from figuring out what the diagnosis is, which is largely going to be a solved problem, to actually administering the right care at the right time and providing, you know, emotional support and and guidance to a patient.
How do you see that deployed in in practical terms? So, how do you put it in the hands of doctors? I think that, you know, doctors will just make a phone call to their AI in clinic. They'll text their AI. They will upload the patient record in, you know, the click of a button. And, you know, the model will be able to reason over all the contents of that. I also think that consumers are going to go direct to, you know, to Copilot, which is what they're already doing.
Almost 20% of the questions that we get in Copilot every day are health-related or medical-related. Um so, it's already our main use case and that's why my team's pushed so hard on medical superintelligence. So, I don't want to confuse our our viewers even more between superintelligence um uh AGI, but what is artificial capable intelligence? >> Mhm. I mean, this was a a phrase that I used 3 years ago in The Coming Wave, a book that I wrote about AI.
And basically, I was trying to break down the AGI term to pick a point along the route which was before AGI. And instead of focusing on the abstract idea of intelligence, to focus on the capabilities. Like, what can it actually do? >> Mhm. And I coupled it with the introduction of the what I was calling the modern Turing test. And that was basically, could an AI take $100,000, invent a new product, create a new business, market it, um and then use that to make a million dollars in some short period of time.
It's not that we want to use AIs just to make money. It's just that it's a nice metric for measuring the efficacy of a complex set of tasks. And I think now that we see Claude bought and many other action-based AIs in operation, you know, this year I think there are going to be models that can, you know, achieve that artificial capable intelligence and satisfy the modern Turing test. What about AGI? How close are we?
I think that we're going to have >> And you've said before that thinking about AGI is misleading. I I think that we're going to have a human-level performance on most, if not all, professional tasks. So, white-collar work where you're sitting down at a computer, either being, you know, a lawyer, an accountant, or a project manager, or a marketing person, most of those tasks will be fully automated by an AI within the next 12 to 18 months.
And we can see this in software engineering. Um many software engineers report that they are now using AI-assisted coding for the vast majority of their code production. Which means that their role shifted now to this meta function of debugging, scrutinizing, of doing the strategic stuff like architecting, of, you know, etc., etc., putting things into production. So, it's a quite different relationship to the technology.
And that's happened in the last 6 months. Um you know, So, what do you think is Anthropic's secret sauce? Why is it that they've been able to sort of corner the market on AI enterprise tools? I think they've done a great job of just focusing on coding capabilities. I mean, compared to the other companies, um their singular focus has paid real dividends. So, they've done a great job. It's great to see. And when are you going to have your model?
We are developing our own superintelligence, um and that's like the main focus of of our efforts in in the Microsoft AI. So, sometime this year. Microsoft has or Microsoft AI has recently opened a UK center. What Why is that? Is that mainly because you think that there's a lot of talent in the UK? >> Yeah, I mean, there's great talent in the UK. We have an incredibly strong ecosystem. Like, there's great universities, um and, you know, we've hired some of the very top people from DeepMind to to join us over the last few years.
So, we're growing that lab quite a bit. And you're not paying them a You don't have to pay them 100 100 million dollars? Um uh no, certainly not. >> you pay any of them 100 million? Certainly not. I mean, look, I think that the numbers are are pretty eye-watering across the industry at the moment. Yes. How do you deal with that? I'm just curious. >> >> I mean, you know, like at the end of the day, when there's a a supply and demand mismatch for talent, then things go exponential.
And so, I don't think you'll last for very long because the knowledge is proliferating like crazy. And obviously, everybody's entering the industry and stuff. So, I I think that was just some, you know, crazy blip in time led by one or two people. Um last question. In Silicon Valley, the debate is all about who's going to get to AGI first or who's going to reach superintelligence. When you look at China, I know you were there at the end of of last year.
This debate doesn't really exist. The debate is how fast can you deploy? Mhm. And a lot of the big Which is the right approach? And is is there too much focus that's put on the incremental improvement in the technology? I mean, maybe to flip that the other way around, what China also does incredibly well is withdrawing deployments quite quickly. Now, obviously they do it arbitrarily without due process. Um, but, you know, it it is important to recognize that their rate of deployment is coupled with this other mechanism to kind of pull things in and be quite restrictive.
Um, we don't have that mechanism [snorts] and I think that that is actually a little bit concerning. I mean, if you know, like you mentioned earlier, the safety incident around Maltbook, I mean, there is going to be a real one of those in the next two or three years and it's it's unfortunately an open question as to what the mechanism is for managing that safety situation from a public interest perspective on the open web.
Um, you know, and I think we all >> there isn't one right now. Well, I think that's basically true. There isn't a It would come with popular pressure. Obviously, if if there was a direct violation of a local law, then, you know, there could be an intervention, but those things take quite a long time to make their way through the system. So, it that's probably one of the biggest areas of concern for me. Mustafa, thank you and when you do reach um super intelligence and you start deploying it, we'd love to have you back.
>> >> That's good. Great to see you, Rula. Thank you.