CEO, OpenAI
Microsoft BOMBSHELL Announcements: Sam Altman on GPT-5, Devin Joins Microsoft and Phi-3 (SUPERCUT)
The latest AI News. Learn about LLMs, Gen AI and get ready for the rollout of AGI. Wes Roth covers the latest happenings in the world of OpenAI, Google, Anthropic, NVIDIA and Open Source AI.
My Links 🔗
➡️ Subscribe: / @wesroth
➡️ Twitter: https://x.com/WesRothMoney
➡️ AI Newsletter: https://natural20.beehiiv.com/subscribe
#ai #openai #llm
About Sam Altman
Sam Altman, CEO of OpenAI, has been active in public appearances discussing AI's societal impact, infrastructure buildout, and the approach of superintelligence. At the groundbreaking for a new data center called "The Barn" in Saline Township, Michigan, Altman described the project as a "huge bet" and said he hoped it would become a model for how data centers and communities can mutually benefit. He stated that the site could help cure diseases and provide tutoring to millions of students, and emphasized the importance of not increasing energy prices and creating union jobs. In media interviews, Altman said people are "right to be anxious" about AI, calling it "one of the biggest" technological shifts, and acknowledged that the industry has failed to articulate how people will stay in control. He also said he believes AI will treat most diseases by 2035 and expressed concern about AI's impact on mental health. Altman has also discussed OpenAI's policy blueprint for preparing society for superintelligence, which he described as urgent and intended to start a debate. He stated that the government should be doing AI research and making longevity treatments available, and that his single biggest contribution has been pushing for AI to be a "democratized technology." Altman announced that the OpenAI Foundation is pledging $100 million to Alzheimer's research, and he participated in the Breakthrough Prize ceremony presenting an award to the Muon g-2 collaborations. He also spoke about the need for "proof of human" systems to verify identity in an age of AI-generated content, and reiterated his view that the most important priorities are accelerating research and the economy.
Transcript (30 segments)
✨ AI-enhanced transcript with speaker attribution
K
Kevin Scott0:00
Again, there's no point of diminishing marginal return here. You can count on things getting more robust and cheaper at a pretty aggressive clip over time. What are the category of things that people should be expecting over the next few months?
So Microsoft has made some bombshell announcements at the Build event. Sam Altman takes the stage to talk about the next big thing that's rolling out. They specifically don't use any numbers, but the message is crystal clear: this train is not stopping, and what's coming next will be bigger, better, faster, cheaper, and more intelligent. Also remember Cognition Labs and Devin, the AI software engineer. Some of you thought that it was fake—it was debunked. Well, now it's joining forces with Microsoft. I told you this guy is not going anywhere. Here's a quick supercut of what I thought were the most important announcements at the Build event on day one.
S
Satya Nadella0:56
I'm working with our partner at OpenAI to push that frontier forward. There's this really beautiful relationship right now between the exponential progression of compute that we're applying to building the platform and the capability and power of the platform that we get. Without mentioning numbers, I want to give you all an idea of the scaling of these systems. In 2020, we built our first AI supercomputer for OpenAI—the supercomputing environment that trained GPT-3. You can think of that system about as big as a great white shark. The next system we built, delivered in 2022, trained GPT-4—it was about as big as an orca, three times the size of the shark. The system we have just deployed is about as big as a whale relative to the shark-size supercomputer. And it turns out you can build a whole hell of a lot of AI with a whale-size supercomputer—five times the size of the orca. One of the things I want everybody to really be thinking clearly about—and this is our segue to talking with Sam—is the next model is coming. This whale-size supercomputer is hard at work right now building the next set of capabilities that we're going to put into your hands so that you all can do the next round of amazing things. And so with that, I'd like to bring Sam Altman to the stage.
K
Kevin Scott2:40
You are one of the busiest people on the planet. Wild week. I really appreciate you taking time out to chat with us today. I guess what I really wanted to start our conversation about—and I asked you this question last week—is there's just been an extraordinary amount of change over the past year and a half. What has been the thing that has surprised you most, particularly relevant to an audience of developers?
S
Sam Altman3:04
I'm delighted to be here, obviously. Great to see you. Developers have been such a core part of what's been happening this last year and a half. There's millions of people building on the platform. What people are doing is totally amazing, and the speed of adoption and talent in figuring out what to build with all of this—over what has really not been very long—is quite remarkable. When we put GPT-3 out in the API, some people thought it was cool but it was narrow. Then seeing what people have done with GPT-4, and seeing now what's happening with GPT-4o even though it's new and hasn't been out that long—I've never seen a technology get adopted so quickly in such a meaningful way. What people are building, how people are finding out how to do things that we never even thought of possible—which is why it's always great to have an API—that's been very cool to see.
K
Kevin Scott3:53
Yeah, and I think what you just said is one of the most important points to me. There's a version of AI that could have existed that is a bunch of smart people building things at extraordinary scale and then just building it into a bunch of products where everybody gets to passively use them. The really brilliant thing that you all have done is taken the exact same set of things and decided to make it available to any developer who's able to sign up for an API key.
S
Sam Altman4:24
Yeah, we try to be really thoughtful about what makes a good API for this. There's going to be all kinds of ways people can use this, but the more this can just be a layer that gets built into every product, every service, the better. And we've tried to make it such that if you want to add intelligence to whatever you're doing—any product, any service—we make that very easy.
K
Kevin Scott4:41
Yeah, and again I think the progress has been stunning. The setup for introducing you onto the stage here was—I saw that big blue whale. You're making good use of the whale-size computer right now. Without getting too specific, which we can't be obviously, what are the category of things that people should be expecting over the next few months?
S
Sam Altman5:06
The most important thing—and this sounds like the most boring, obvious thing I can say, but I think it's actually much deeper than it sounds—the most important thing is that the models are just going to get smarter generally across the board. There will be a lot of other things too which we can talk about, but if you think about what happened from GPT-3 to 3.5 to 4, it just got smarter and you use it for all these things. It got a little more robust, it got much safer both because the model got smarter and we put much more work into building the safety tools around it, it got more useful. But the underlying capability—this amazing emergent property of us actually seeming to increase the general capability of the model across the board—that's going to keep happening. And the jump that we have seen in the utility that a model can deliver with each of those half-step jumps in smartness is quite significant each time. So as we think about the next model and the next one and the incredible things that developers are going to build with that, I think that's the most important thing to keep in mind. Also, speed and cost really matter to us. With GPT-4o we were able to bring the price down by half and double the speed. New modalities really matter—voice mode has been actually a genuine surprise for me in how much I like it, and when people start integrating that I think that'll matter. But it's the overall intelligence that'll be coming that I think matters the most.
K
Kevin Scott6:23
So for a while now you have been one of the most successful startup investors in the world, and now you are one of the most successful CEOs of one of the most important companies in the world. You've got a room full of developers here—I think there are 5,000 people in the room and about 200,000 people online right now. What's your advice to them as they think about how to spend their precious time given what's happening in the world?
S
Sam Altman6:49
Two things. Number one: this is probably the most exciting time to be building a product, doing a startup, whatever it is, that we have seen at least since the mobile boom, and probably since the internet, and maybe even bigger than that—we don't know yet. The big opportunities, the ability to build something new and really change the landscape, that comes at platform shift times. We haven't had a platform shift in a while, and this looks like it's really truly a platform shift. So my biggest piece of advice is: this is a special time and take advantage of it. This is not the time to delay what you were planning to do or wait for the next thing. This is a special moment and a few years where a lot of stuff is going to happen and a lot of great new things are going to get going. The second thing, also about platform shifts: when the mobile phone revolution started—like 2008, 2009—you would see people say 'we're a mobile company, we have a mobile app,' and then only a few years later, no one said they were a mobile company because it was table stakes. An amazing new technology—which I would put AI in that category—it doesn't get you out of the hard work of building a great product or a great company or a great service. You still have to do it. AI alone is a new enabler, but it does not automatically break the rules of business. So you can use this as a new thing to do, but you still have to figure out how you're going to build enduring value in whatever you're doing. And it's easy to lose sight of that in the excitement of the gold rush.
K
Kevin Scott8:21
So one last thing before we let you go. You and I, and members of your team and members of the Microsoft team, have been doing really an extraordinary volume of work over the past year and a half, two years, thinking about safe deployment of an awful lot of AI capability—from APIs and developer tools to end products. We've accumulated a really interesting volume of experience that's sort of hard to get if you're not doing deployments at this scale. And part of the interesting and surprising progression of capabilities of these models means that they're more useful in helping make AI systems safer. I don't know whether you had some thoughts you wanted to share there as well.
S
Sam Altman9:10
You know, when we first developed this technology, we spent a lot of time talking about: we've made this thing, it's cool—are we ever going to be able to get it to an acceptable level of robustness and safety? And now we kind of take that for granted. With GPT-4, if you use it, it's far from perfect, we have more work to do, but it is generally considered robust enough and safe enough for a wide variety of uses. And that took an enormous amount of work across both teams and fundamental research. When we started this, we had this language model that looked kind of impressive and kind of not, and even then—how are we going to get it aligned? What is it going to take to be able to deploy it? The number of different teams we've had to build up—from research and creation of the model to safety systems to figuring out policy to how we do the monitoring—that's a huge amount of work, but it's necessary to be able to deploy these and use them. When you take a medicine, you want to know it's going to be safe. When you use an AI model, you want to know it's going to be robust and behave the way you want. I've been super proud of the work the teams have done together, and it's amazing how fast this much work has happened and that we can all now use this and say, oh yeah, basically it works. As the models get more powerful, there will be many new things we have to figure out as we move towards AGI—the level of complexity and the new research that it'll take will increase. I'm sure we'll do that together. But we view this as a gate on being able to put these things out into the world, which we really want to do.
K
Kevin Scott10:35
Yeah, it's definitely table stakes is how we at Microsoft have been thinking about building applications on top of this incredible platform that is emerging right now. So impossible, difficult—and I think it's the same challenge that all of you face—is that you really want to focus on things that have made the transition from impossible to merely difficult. That's where all the interesting stuff is. If you look at the history of platform revolutions, that's where all the interesting companies emerge from, it's where all the innovation happens, it's where all the value gets unlocked. And in the case of technology platforms that are sort of exponentially progressing, it's like the only reasonable place to go aim. Because if you're aiming somewhere different—the platform is becoming so much more capable and so much cheaper over time that everything that you sort of have in your imagination that's too expensive to do right now or too fragile is going to become cheap and robust before you can even blink your eye. So that is really the thing more than anything else that I would say to all of you to take away from what I'm saying here today: really focus on those phase transitions.
So while you all have been out there grinding away building really extraordinary things over the past year with all of these AI tools, we've been hard at work trying to make forward progress on our AI platform. We talked a lot about how we're optimizing the current frontier—making things cheaper and more powerful and complete—but we've also been hard at work building new supercomputing infrastructure. Another thing I'm super excited about is the work that Cognition has been doing. Scott and the Cognition team are also here with us today. We have just recently announced a partnership between Microsoft and Cognition. Their product Devin is an absolutely amazing tool—if you can imagine some of the most tedious things that you do as an engineer or software developer, Devin is a tool designed to help you with those tasks. The incredible work that Devin is doing on top of these incredibly powerful tools is just really extraordinary, and we're super excited to be partnering with them and bringing all of the power of what they're doing to Azure.
S
Satya Nadella13:33
I really want to talk with you all today about just a couple of simple things. What's driving all of this progress? Why is all of this happening right now? Part of it is we're riding an extraordinary platform wave—something is fundamentally changing in the universe of technology, much in the same way that it changed when we were going through the PC revolution, where Moore's Law was driving an incredible increase in the power and lowering of the cost of personal computing, which led to it becoming ubiquitous. A similar thing happened with the internet revolution, where networking technology connected all of this compute together and allowed us to do things that previously were unimaginable. And we're going through one of those major technological changes right now, partly driven by the incredible scaling of the capability of AI systems as you apply more compute and more data to training them. But before we get to that expansion of the frontier, a super important part of the emergence of a new powerful platform is completing the stack. It's actually hard work—even when you have a piece of technology that is improving at an exponential rate—to figure out how to do all of the things that have to be done in order to deploy it in real applications so that you can go out and deliver value to real customers. We've done a huge amount of work over the past year on the Copilot stack—both optimizing a bunch of systems so things are getting cheaper and more capable, and building that whole cloud of capabilities, systems, services, and tools around the core AI platforms so that you all can build the things that matter to you.
One of the reasons that we have been able to do this is no other company has deployed more generative AI applications over the past year than Microsoft has. You have probably heard us talking about all of these different Copilots—this new software pattern that we originated with GitHub Copilot, where you pair powerful generative AI with this user interface paradigm where you're using the AI to help assist users with tasks. You can apply this to everything, and many of you in the audience are building your own Copilots. Microsoft itself is building Copilots for service, for sales, a Copilot in Bing, Copilot in Edge, Copilot in Windows. The reason we've been able to do all of this work is because we have the Copilot stack that we built for ourselves—to have real agility in getting these products built quickly, built efficiently, price and cost optimized, and built in a way where they're safe and secure. One of the things you'll be hearing a lot more of at Build is that part of what the Copilot stack is allowing us to do is to unify the experience across all of these Copilots into one logical Microsoft Copilot, where you don't have to really pay attention to which Microsoft product or service you're in—the Copilot just understands all of your context and delivers all of the capability of the model in the context of your data and your task to you when you need it.
The other thing that is really driving progress is not just this completion of the Copilot stack, but we are riding a fundamental wave in the development of this AI platform. If you just look at compute over time—how much GPU cycles or accelerator cycles we're using to train the very biggest models in the world—since about 2012, that rate of increase in compute when applied to training has been increasing exponentially. And we are nowhere near the point of diminishing marginal returns on how powerful we can make AI models as we increase the scale of compute. So we're sort of doing two things at once at Microsoft: optimizing the current frontier and building that toolkit to help you all leverage it, while at the same time investing at a pretty incredible rate in pushing the frontier forward.
One of the super interesting things that has just happened as we're pushing the frontier forward is what our partner OpenAI launched last week in GPT-4o. As mentioned earlier, GPT-4 is a stunning achievement—a multimodal model that understands a bunch of different input types from video to text to speech, that can respond in a bunch of rich ways from text to speech and eventually video, and can respond to users in their applications in real time. In the case of the ChatGPT demos that folks have seen, you can even interrupt the model so that you can have really fluid interactions with these systems. An enormous amount of work has gone into GPT-4o—both the model itself as well as the supporting infrastructure around it—to ensure that it's safe by design.
I wanted to also just remind folks that this efficiency point is real. While we're building bigger supercomputers to get the next big models out and to deliver more and more capability, we're also grinding away on making the current generation of models much more efficient. Text: 12 times decrease in cost, six times increase in speed. Quite a year and a half ago, it's 12 times cheaper to make a call to GPT-4o than the original GPT-4 model, and it's also six times faster in terms of time to first token response. It's just really extraordinary how much progress we're making because of the full set of optimizations—from the silicon we're building, networks, data center optimization, as well as an incredible amount of software work on top of all of this hardware and infrastructure to really tune the performance of these systems. The great thing is, again, there's no point of diminishing marginal return here. One of the messages that I want to land with you all today is that you can count on things getting more robust and cheaper at a pretty aggressive clip over time. It's a really important thing to internalize. We challenge ourselves on at Microsoft all the time: aim for things that are really truly ambitious, because all of this optimization work is going to accrete to make things really ubiquitous in terms of how you can go deploy them.
I have a little quick demo video here, so let's roll the video. [Demo plays: A woman asks for help debugging Python code using GPT-4o's vision and voice capabilities. She shows her code via phone camera, the model identifies a bug—using extend instead of append—and suggests a fix. After she corrects the code, it runs successfully.] I mean, it really is extraordinary. I should say by the way that Jennifer would never make that actual mistake in writing a Python application, but Kevin might. I do want to make sure that we're paying attention to just how much has changed over the past year. What you just saw would have been absolutely inconceivable to think about actually working. This was not a tortured demo—Jennifer showed me this last night and then she just recorded this demo. It's just crazy that it works this well.
Another set of things that have been making a huge amount of progress is what's possible with smaller models. We have been working for a while on this series of models called Phi that are small language models. Satya mentioned this in his keynote earlier—imagine an efficient frontier. Usually when you're building these models, you're trading size off—which is related to performance and cost and a whole bunch of other things—versus quality. The smaller the model is, the cheaper it is to do inference and the less compute you need to run the model, so small models are more amenable to running on devices, but it usually means you take a hit on quality. What we're discovering, particularly over the past year, is that there's this notion of an efficient frontier. We don't even show the GPT-4 point on this slide—it would be way, way off to the right in terms of size. If you want extreme levels of quality and performance, a frontier model is your friend, but in some cases you may want to choose one of these other models somewhere on this efficient frontier where the trade-off between cost to serve, latency, or locality is acceptable given the quality you can get. The very interesting thing that's been happening is the quality you're able to achieve in these small models is getting pretty high.
Remember back ancient history to the launch of ChatGPT in November of 2022. ChatGPT launched on top of GPT-3.5, and everybody was absolutely gobsmacked at what was possible. Fast forward a few months to March 2023, and ChatGPT gets an upgrade to GPT-4, which is even more extraordinary—able to ask extremely complicated questions and get very rich, interesting, compelling completions. Now fast forward to today: a version of Phi-3, optimized and running on a mobile phone, can respond to a prompt just like ChatGPT could just a year or so ago, with responses that are sort of equivalent. This is not arguing that Phi-3 running on this device is just as powerful as GPT-4—it is not. But the way you all should be thinking about it is: in many cases, these models can be appropriate to use for building your applications when you have a particular set of constraints that you're trying to optimize towards.
I wanted to really motivate why this matters with the following example. Satya mentioned earlier the partnership that Microsoft has formed with Khan Academy. Khan Academy's mission is really interesting and important—they are trying to ensure that every learner on the planet, no matter where they are, has access to high-quality, individualized instruction. One of the things we are exploring together with Khan Academy is the possibility of achieving that goal of ubiquity of personalized learning agents by using something like Phi-3, where you can imagine training a Phi-3 model that's very good at something like math instruction. This is an actual interaction with Phi-3 medium that has been fine-tuned to work particularly well for math tutoring. The challenge with doing something like this is not just having the model give the student an answer—you want it to lead them towards discovering the answer themselves. A tutor is very different from an answer agent. It's exciting to think about how many tools organizations like Khan Academy have to solve these really important missions. So with that, I'd love to bring Sal Khan from Khan Academy onto the stage.
K
Kevin Scott27:29
One of the interesting things that happened when ChatGPT burst onto the scene a few years ago is that there was this reaction from a bunch of educators—actually a reasonable reaction—where, okay, we don't understand this, we don't want our students using it, they're going to do things that we would prefer that they didn't do. You, on the other hand, looked at this and said 'this is amazing' and leaned all the way in. Can you explain a little bit about what drove your first reaction to this new technology?
S
Sal Khan27:55
Yeah, you know, some of you all know how Khan Academy got started. Almost 20 years ago, it started with me tutoring a cousin. I was a hedge fund analyst at the time. I tutor one cousin, word spreads in my family—free tutoring is going on. Before I know it, I'm tutoring 10 to 15 cousins. I start writing tools for them, software. I start making videos—that's what a lot of people know about Khan Academy. If you think about that journey from then till now, even right before we started really working on generative AI, everything we've been doing is: how could you scale that type of personalization that I was originally doing with my cousin Navia? We were approximating it with software and videos and teacher tools, but to some degree there was an asymptote to how far you could get with pre-generative AI tools. And then when we saw—and it was really GPT-4 that opened our mind. Greg and Sam from OpenAI showed it to us end of the summer 2022, and we realized: there's things that have to be worked out, but it could get that much closer to emulating what a real tutor would do. It was obvious it could also be used as a cheating tool, and you have to worry about safety and privacy, especially with under-18 users. But I told the team: let's turn those into features, let's put the guardrails on it, because this could get us that much closer to our mission, which is free, world-class education.
K
Kevin Scott29:01
Yeah, I think one of the other things that you all have done—and this is a really important thing to internalize about these models and systems—is the model isn't a product. The systems aren't silver bullets. You still actually have to understand who your customer is, what problem you're trying to solve, and how to go deal with a whole bunch of gnarly things on top of this incredibly interesting and powerful tool so you can do something useful. Do you want to talk a little bit about what you had to do there?
S
Sal Khan29:26
Yeah, and I have to admit—and maybe a lot of people in this room or in the world right now are experiencing this—every now and then you see some of these demos and you're like, does my application even have relevance anymore? This thing is going to be able to do everything! But then when you sit down and you really think about how a school system, a teacher, a student is going to use it—and what are the guardrails, what are the privacy concerns, how do you make sure that it really does the tutoring interactions appropriately and it's aligned to standards—you realize that there's a lot to do at the application layer. Now I think we're all discovering together this new world of developing applications on top of large language models. It's not deterministic in a traditional way. You have to have evals, you have to constantly test it. But we're realizing that there's just so much to do. It really is a very exciting time.
K
Kevin Scott30:10
Yeah, I mean one of the things that I'm especially excited about is this mission that you all have for ubiquity, and the partnership that we're doing with you all is going to enable you to get every teacher in the United States hands-on access to Khanmigo and your tools. Just a personal anecdote for me: my daughter is in the ninth grade, she's taken biochemistry and is just in love with science in general. She, on her own without any prompting from Dad, figured out how to use the free version of ChatGPT to take a bunch of biochemistry papers that were way, way, way more complicated than a 15-year-old by rights has to understand, dump them into ChatGPT, and then just ask a million questions about it. Her learning acceleration, because she's figured out how to use this tool, is extraordinary. I just want every kid in the world to have the same experience that my daughter has.
S
Sal Khan31:02
Absolutely. And you know what we realize at Khan Academy—there is a subset of students that if you give them the tool, and it sounds like you're lucky enough to have a daughter like that, they will run with it. But what you really need in most cases is you need caring adults, primarily teachers, motivating students, driving that usage. And so what we're really excited about with this partnership—and this is a big deal, I want to make sure you know—we are using state-of-the-art models that use real compute, it has real cost associated with it. When we launched Khanmigo, which is still out there and it's a tutor for students, it's a teaching assistant for teachers—but what we're launching today as part of this partnership is these state-of-the-art teacher tools. We're going to be able to give free to every teacher in the United States so that they can get productivity improvements. Yeah, big, big, big deal.