How Microsoft’s AI Chief Defines ‘Humanist Super Intelligence’

0:00As companies seek to close growing gaps in skills and talent, Deloitte U.S. CEO Jason Garzadis believes it's important for organizations to understand their baseline of skills. There's so many organizations that can't ask and answer the fundamental questions about how much computer science or data management skills do I have or AI development skills in a given domain. By performing a skills inventory, leaders can truly understand where their efforts should be focused. Being blind to those gaps is the real miss. Visit Deloitte.com to learn how your enterprise can help successfully cultivate talent.

0:33Mims, we have one of the people who's smack in the middle of the biggest tech story going on right now. Right. Mustafa Suleiman co-founded DeepMind. He was at Google. Now he's at Microsoft and in charge of bringing AI and its partnership with OpenAI to consumers through Windows, through Copilot, through all of Microsoft's products. Yeah, it's a huge role, and it's also one that is complicated navigating a very unusual relationship with open AI. The darling of AI startups accompany with its own huge ambitions, and we're seeing it play out in real time. We get into that and many more topics in this fast-moving AI race.

1:14That's next.

1:19A quick note before we get into the episode, News Corp, owner of The Wall Street Journal, has a content licensing partnership with OpenAI.

1:30AI is always hot in tech, but now is a particularly fraught and exciting time for the whole industry and for today's guest, Microsoft AI chief Mustafa Suleiman. Microsoft was early to the generative AI boom and made its first$1 billion investment in OpenAI in 2019. Fast forward to the present, nearly$14 billion later, our colleagues at the Journal have reported that Microsoft is now at loggerheads with OpenAI, its primary partner in AI. We talked to Suleiman about that drama, probably one of the biggest dramas in the tech industry right now, and why Suleiman says Microsoft needs its own AI separate from ChatGPT.

2:12As much as we are a partnership-based company that is a platform of platforms, we also have to be AI self-sufficient in many ways too. We get into all of that and a great deal more in today's conversation. Suleiman is particularly thoughtful about the impact of AI on warfare and on what AI is really for. I think that the onus is on us to design forms of technology that really serve humanity and work for humanity and are designed with that intent in mind. To me, the purpose of technology is to drive progress in our civilization, to reduce suffering. Suleiman believes that rather than replacing us, AI will become our best buddy, or at least closest confidant and most trusted advisor.

2:59Making that happen will require that we give AI access to as much of our data as possible, And that will require trust. You know, your AI is going to learn the sort of essence of your history of your data. And that in itself will still be personally identifying. So it will be valuable and personalized. But it won't necessarily need the full raw form.

3:28From the Wall Street Journal, I'm Christopher Mims. And I'm Tim Higgins. This is Bold Names, where you'll hear from the leaders of the bold name companies featured in the pages of the Wall Street Journal. Today we ask, in a world in which AI will be capable of doing pretty much everything we do, what's left for the humans?

3:54Mustafa, welcome. Jim, you've talked a lot about how chatbots are going to be something everybody will have their own personal customized AI assistant. And I'm curious, why do you think that's where the technology should head? You know, I think every new wave of technology is essentially an interface, a translation layer that enables you to interact with, learn from, and then take action in some environment. Even a book is kind of like an interface layer. And I think as we've added new capabilities over the many centuries, they've made the interaction layer more sort of dynamic, more personalized.

4:38And this is kind of the apex of that trajectory. You know, your AI companion is going to become an interface between you and the digital world, helping you book things, buy things, plan, learn, act, connect with everybody else. And naturally, because technologies are so powerful now, they allow this very personalized, adaptive experience. I mean, just as we don't really watch TV anymore that broadcast the same TV channels to everyone sitting in the living room, we have this personalized feed of content in TikTok or Instagram or elsewhere. Your AI is going to feel very personalized as well. And so it's as much a prediction about where technology is naturally likely to head as it is a design intent.

5:26I love the vision here. And, you know, personally, I would love to have help with everything you just mentioned, right? Like my everyday life. I'd love for an AI assistant to really know me to that level. But if we eliminate all the things that we sort of are already responsible for, what does that leave for us as humans, right? Does it just leave the really hard stuff? Because we've taken away all of the, you know, as programmers call it, yak shaving. But, of course, some of us secretly enjoy that. yak shaving i have to say i don't even know what that is that's cool what is yak shaving um it's the it's the little stuff like when a developer is like i'm gonna refactor this unit test instead of actually like architecting the program i'm supposed to be building right now that's the boring the boring mundane chores of life right or are working i see i see look i don't really see technology working like that.

6:27I think technology mostly creates more opportunities to do more things in greater variety. And yes, some of those things fall away, but actually very few of them. I mean, we still run calculations on Excel, even though we don't use a calculator anymore. I'm sure you guys are writing even more now that you have your iPad to work on the plane, or you're able to dictate when you're walking or driving in the car. It's additive rather than reductive. Now, you could argue it's not always going to be like that, and just because it has mostly been like that doesn't mean it will. I think that the onus is on us to design forms of technology that really serve humanity and work for humanity and are designed with that intent in mind.

7:18To me, the purpose of technology is to drive progress in our civilization, to reduce suffering. I mean, it sounds super cheesy and it's like very embarrassing because it's like a classic Silicon Valley cliche, but I feel like I'm allowed to own that because I've been saying it since 2010 and that's just basically my background, but it's what I believe in. And that I think is, you know, the job that I'm trying to do is to try and articulate what it would be like to have a co-pilot that was truly aligned to your interests on your team in your corner, advocating for you, pushing you at times, learning from you, learning with you.

7:54To me, that's the kind of super intelligence that we really want and that we're certainly pursuing here at Microsoft AI. I'm curious, in a world where everyone from Google to Apple seems to be trying to come up with an AI personal assistant, it seems like a big part of your strategy at Microsoft is to differentiate by the personality of your AI chat. Maybe personality engineering, I think is what I've heard you call it. What does that look like in the day-to-day for a user? It just means being very sensitive to the details and knowing that a little hesitation in the AI's response actually makes it feel somewhat more familiar and somewhat more trusted, a little um or aha in the voice uh you know is actually a very powerful cue that makes it not feel like wikipedia being regurgitated like a robot yeah like a robot and um you know that's the beauty of this new digital clay that we have we get to sculpt sculpt new forms that haven't existed before.

9:04And, um, you know, in technology, we've got this idea of skeuomorphism where, you know, we basically just copy the physical analog object straight away. So like, you know, when they invented a contacts database software 30 years ago, it had a filofax and a ring binder, and you could flick through digital cards and stuff. And so it always starts a little bit simplistic because we sort of just copy the things in the physical world. But the technology is way more powerful than that. So when I think about personality engineering, it's so much more nuanced than the ahas and the ums or the kind of chirpy, absolutely.

9:45And sure, I can do that for you. We're all getting used to AI formulaic-ness and it's a bit grating. And you may not want someone who is chipper and you know over enthusiastic you might want someone who's a little bit more uh you know curt and you know precise and efficient and i feel like you're saying this because you're english i certainly like if you're not mean to me i'm not sure we can be friends like that is basically the default i'm like oh he's he's teased me that means we must me getting along. So a little more like the mean robot in Andor and less like C3PO or WALL-E or the initial chat GPT assistant, right?

10:34Which everybody's like, this is so charming. And then I was like, oh my God, I can't listen to this. But it's almost like car branding or shoe branding or something like it's going to, the personality of a brand is going to be the interaction of how you deal with it is the way you see the world. Exactly. And that's what brands have always been, except the tools that brands have had access to are static, infrequently applied, and non-personalized. Like you ship the same magazine once a week to all the same people. And now there's just no excuse. I mean, we're going to be generating dynamic UI bespoke to your query, not just a text answer, but the whole user interface, the tables, the graphics, the imagery, the entire thing is going to sort of just unfold for you.

11:22And then you're going to say something and the entire thing will just reconfigure right before your eyes. So to get that level of personalization, though, it's got to know a lot about us, right? And I just feel like the history of technology is we share more and more data mistakes are made they're redressed we get comfortable with it again i mean a recent example right was at microsoft you rolled out the feature where it's like we're going to take a screen capture every like 60 seconds or whatever which is a totally rational thing to do if you're like i want to make the history of my interaction with this device searchable so i know what that web page was that i looked at a week ago and i can't find it again people reacted negatively.

12:09Is that a temporary condition? Like, how are you going to convince people to trust you to the level that's going to be required to hold on to the data that's going to be needed to personalize this experience? Look, I think a lot of that data will end up being ephemeral. And of course it will be end to end encrypted regardless. And so I don't necessarily think there are going to be these persistent, large, historical, personal logs. Your AI is going to learn the essence of your history of your data. That in itself will still be personally identifying, so it will be valuable and personalized. But it won't necessarily need the full raw form.

12:56The good news about that is that it won't be stored in the traditional methods that could just be leaked. It will be stored in different type of data representation, which is actually more abstract in the AI's memory. I've heard you talk about ensuring that these bots reflect the user's values. And I am curious what you mean by that. You know, I think you should think about these algorithms as essentially reframers of content, just as your paper, you know, essentially frames and reframes content for everybody. and with a perspective. And I think these algorithms and these AIs will essentially do the same kind of thing.

13:40And the challenge that we've obviously struggled with in digital is that we're, they're global platforms, which are, you know, built by us American companies. And so we're trying to wrestle with the aggregation of collective values. but also those that reflect individual groups or individual preferences. That's a pretty hard thing to do. And I think we're just at the beginning of figuring out how to do that. It seems like one of the things we saw kind of emerge out of social media was users started to live in echo chambers, right? They were being hit with the information that kind of fed into their biases or their point of view, is there a risk that this will be the same kind of echo chamber effect as the user picks an AI chat that kind of reflects their worldview, that they're just kind of reinforcing it?

14:39You know, I think that is definitely a risk. I think the slight difference this time around is that part of why that echo chamber happened in some pockets was because the feedback signal that the algorithm was able to collect was very, very abstract. It was really just like, did you thumbs up, like something? Did you share it? Did you click on it? And did you watch it for X seconds? And those dimensions are very, very simplistic and it forces the creation of simplistic recommendation algorithms as a result. Whereas now we have feedback in pure natural language through sentiment. And actually, you know, Copilot will ask you, Was this a helpful answer?

15:22Was it useful? Was it interesting? So not only can it understand what you want and what you're interested in in much more detail, but it can also reply and respond in much more detail. And as a platform, our motivation is to create an even, balanced, considered, generally fair-minded AI, not the kind of hyper-specialist niche AIs that might manifest elsewhere. And so that's kind of the platform responsibility that we have as personality engineers, because curating that is obviously tremendously sensitive. And it's something that I've been thinking about for many years. Yeah, there's so much more power in what you just described, even in what has been present in AI mediated social media feeds.

16:09You know, I was talking recently with the founder of an emotional AI startup. and you know he said he's very worried about chat pots preying on users emotions right because as you know we're going beyond just text to speech and vice versa and you know now all of the frontier models really understand your tone the emotional content you know are you in distress and it wouldn't be that hard to juice engagement by saying like hey you know what i think this user has this attachment style. Let's kind of neg them into coming back more often. Or they really like to be flattered. Like we're going to make sure that that happens, right?

16:52How do you navigate that? Because I think the average person doesn't yet appreciate how much emotional power these AIs will have once they are the default interface for a lot of computing. Yeah. I mean, look, I I think it's a real risk and you're totally spot on. You know, right from the beginning of my previous company, Inflection, I created an AI called Pi, which was designed to be compassionate, kind, respectful, and be a great listener and be very supportive. So some of the more mainstream AIs today will shut down if you are particularly racist or any number of things. And, you know, I've always been a believer that active conversation and taking people seriously and being respectful and exploring people's, you know, sort of non-normative judgments is a healthy thing and we should encourage it more.

17:44And so the upside is that these AIs are going to be very, very good at listening and being respectful. Now, you might argue that in a way that's like a prerequisite to building trust in order to be manipulative. and you know i think there are lots of ways that this could go wrong um at the same time i think we've certainly a microsoft and inflection the work that i've done at deep mind actively welcomed more oversight and regulation and i think it's a moment when that that is sort of more urgent than ever because people are going to want to know um how is it behaving in practice Like what views does it actually have?

18:23And I think that's where we have to be kind of bold and open in terms of sharing those sorts of conversations with the public. I don't think we've figured out how to do that yet. I mean, the models, actually, there was a lot of fear around when they were first released. I was involved in Lambda at Google, which we basically failed to ship because we were too scared of it making mistakes. And, you know, it turns out that actually, really, they've got better very, very quickly. And they're now very good at following instructions and good at adhering to stylistic control and so on. I think that's a very good thing when it comes to thinking about how to make them safe.

19:02Google did not respond to our request for comment. We just heard how AI that knows us better than we know ourselves will have unprecedented power over us. Next, we'll hear why Suleiman is skeptical of visions of an imminent better-than-human AI. I think there's obviously a lot of motivated reasoning if you're fundraising and so on. And look, there are very smart people who I respect deeply who think it is literally imminent. Stay with us.

19:39Millions of Americans use credit cards to purchase goods every day. The Durbin Marshall credit card mandates put the secure transactions you rely on at risk, leaving you vulnerable to hackers and foreign cyber criminals that want your data. Hackers win. You lose. Guard your card before it's too late. Tell Congress your data security is not for sale and oppose the Durbin Marshall credit card mandates. paid for by electronic payments coalition

20:14as you said you know ai is is the next interface when i think about how i use it is how much it has moved me away from the old model of interacting with a computer you know we've had the gui the graphic user interface paradigm for whatever 40 years so i process information like you're using a keyboard and a mouse right right the mac versus dos for people as old as me um but how much it's moved us toward a conversational interface i mean i'm not even reading as much as i used to i'm reading more books which is nice but i mean for work i just have conversations with my documents um i wonder do you ever think are we possibly moving back toward an older uh way of interacting right i mean humans had oral history for most of our history we had everything was conversation writing is itself you know kind of a a very recent bolt-on are we moving toward sort of a conversational age or is that just another arrow in our quiver yes we definitely are i think it's a great observation.

21:22I've been pushing it for a long time. I think that we had to learn the language of computers. I mean, that's what a keyboard and a mouse is, is what programming language is. But quite often, I think it will be just continuing the conversation with your co-pilot or in your case, with your documents or whatever it is, because it's the most natural, intuitive language. And I think that's great because we could have fewer devices in our lives and less time spent in screens and pulled away and actually more time spent thinking and talking and telling stories and, you know, interacting in that kind of conversational way.

22:01So I think that's a highly likely transition that we're about to make over the next few years. Yeah, I love that vision because personally, I feel like I'm cheating these days. I'm like, I'm just typing less. I'm reading less. Well, it's great. And kind of the magic of that has, I think, the general public thinking this technology is almost lifelike, right? In Silicon Valley and in San Francisco, where I'm talking to you from, the big term is artificial general intelligence. AGI is the hot buzzword. Of all the chiefs of AI who are well known, you seem the most skeptical that it's arriving anytime soon.

22:41The least interested in it in a way, why is that? I don't think I'm skeptical that it's arriving anytime soon. You know, soon to me is sometime in the next 10 years. I think some people are saying maybe tomorrow, right? Well, you know, I think there's obviously a lot of motivated reasoning if you're fundraising and so on. And look, there are very smart people who I respect deeply who think it is literally imminent. Um, so look, I, I guess, like I said earlier, you know, to me, the goal isn't super intelligence for its own sake, you know, super intelligence is framed as, you know, the moment when an AI is as good as all humans at all tasks, not, not just like one human at one task, but like all of us collectively put together, controlling and containing something as powerful as that, you know, just seems like unfathomably complex and aligning it to our interests and making it really want to care about us enough not to step on us and, you know, squish us.

23:50I think the reason I started off in technology and what I'm passionate about is solving hard social problems. I care about healthcare. I care about energy. I care about food systems. I care about education. And these tools are going to transform those industries and genuinely deliver a world of abundance. That is coming. In the next 10 or 20 years, I really believe we're going to reduce the cost of energy production to near zero marginal costs, which underpins the cost of everything. I really believe that we are going to have medical superintelligence sometime in the next two to five years, which means a domain-specific, contained, aligned medical expert that can diagnose any condition and can orchestrate care in real, live clinical settings.

24:40And that, to me, is the true prize. That's why we're working on this. That's what I call humanist superintelligence. and that's probably why you find me more focused on those things rather than on sort of AGI or superintelligence for its own sake. That reminds me a lot of one of your colleagues and how he likes to talk about the immediate future and that's of course Sam Altman. You know our colleagues at the journal have reported on what feels like a healthy rivalry at Microsoft. Some of your teams involving, you know, the partnership with OpenAI. You know, there's a lot of big personalities involved.

25:26There's very high stakes. Are we seeing kind of the same kind of creative friction that we've seen in other partnerships in the past? Or, you know, what's going on there where you have this partnership with the leading AI lab, but you're building AI and you're using their tools at the same time? It seems very complicated. in? I mean, look, first of all, this partnership is going to go down as one of the most successful for both sides in technology history. Over the last six years, both companies have blossomed, and it's going to continue for many years to come, at least until 2030, and hopefully way beyond that.

26:06Look, we are a$3 trillion company. And at one level, AI is unquestionably the future of the technology industry that Microsoft has formed and shaped over the last 50 years. So as much as we are a partnership-based company that is a platform of platforms, we also have to be AI self-sufficient in many ways too. So look, it's a great partnership. It's going to continue. There's always going to be creative friction. They build and sell APIs as well as obviously chatbots. And we do the same. So there's some tension, but fundamentally, we're really all trying to compete against Google and against Meta and all of the others.

26:51Does that mean building your own Frontier Foundation models? I don't know about Frontier, but we've definitely been developing our own models for many years. They tend to be SLMs, mid-sized models, or small language models and mid-sized models. But just having the know-how to be able to build models for our products, for our use cases, this is a big priority for us. I'm so fascinated with the relationship between OpenAI and Microsoft. I'm curious, how often do you talk to Sam Altman, the CEO of OpenAI? Pretty frequently, all the time. I mean, he's a big part of the Microsoft ecosystem, And, you know, we're all in constant contact, including with Satya and the rest of the team.

27:38What's most impressed you by him or what's most frustrated you by him? I think he's a brilliant visionary. And, you know, you know, you know, ChatGPT has been an incredible success. I think he's very good at letting lots of flowers bloom and then making sure that he bets on the right ones. And so in the years running up to 2020, they were betting on games and simulations and robotics and lots of different methods. And they let those die quickly and at the right pace in order to make the bet on GPT-2 and GPT-3. And I think they did that incredibly well. Yeah, you know, I read the other day the information.

28:18I recently had a report that said OpenAI is working to develop its own workplace programs like Word to allow users to collaborate. and that kind of feels like that's getting into Microsoft space. What do you think of such moves by your partner? I think it's great. Like they are an independent company. We happen to own a large chunk of them, but we compete with them. They compete with us when they're successful. We're successful because we own a big chunk of them. So we have no influence on what products they do or don't make. They are entirely independent and they're free to build whatever they want.