We started OpenAI years ago because we felt like something really interesting happening in AI and we wanted to help steer it in a direction. It’s honestly just really amazing to see how far this whole field has come then. And it’s really gratifying to hear from people like Raymond are using the technology we are building, and others, for so many things. We hear from people who are excited, we from people who are concerned, we hear from people who both those emotions at once. And honestly, that’s how we feel. Above all, it like we’re entering an historic period right now where we as a world are going to define technology that will be so important for our society going forward. I believe that we can manage this for good.
So today, I want to show you current state of that technology and some of the underlying design that we hold dear.
So the first thing I’m going show you is what it’s like to build a tool for AI rather than building it for a human. So we have a DALL-E model, which generates images, and we are exposing as an app for ChatGPT to use on your behalf. And can do things like ask, you know, suggest a post-TED meal and draw a picture of it.
(Laughter)
Now you all of the, sort of, ideation and creative back-and-forth and taking of the details for you that you get out of ChatGPT. here we go, it’s not just the idea for the meal, but very, very detailed spread. So let’s see what we’re going to get. But doesn’t just generate images in this case — sorry, it doesn’t text, it also generates an image. And that is something that really expands the power what it can do on your behalf in terms of carrying out your intent. And I’ll out, this is all a live demo. This is all generated by the AI as speak. So I actually don’t even know what we’re going to see. This wonderful.
(Applause)
I’m getting hungry just looking at it.
Now we’ve extended ChatGPT with tools too, for example, memory. You can say “save this later.” And the interesting thing about these tools is they’re very inspectable. you get this little pop up here that says “use DALL-E app.” And by the way, this is coming to you, all ChatGPT users, over months. And you can look under the hood and see what it actually did was write a prompt just like a human could. And so you sort have this ability to inspect how the machine is using these tools, which allows us to provide to them.
Now it’s saved for later, and let me you what it’s like to use that information and to integrate with other applications too. can say, “Now make a shopping list for the tasty thing I was suggesting earlier.” And make a little tricky for the AI. “And tweet it out for all the TED out there.”
(Laughter)
So if you do make this wonderful, wonderful meal, I definitely to know how it tastes.
But you can see ChatGPT is selecting all these different tools without me having to it explicitly which ones to use in any situation. And this, I think, shows a new way thinking about the user interface. Like, we are so used to thinking of, well, we have apps, we click between them, we copy/paste between them, and usually it’s a great experience within app as long as you kind of know the menus and know the options. Yes, I would like you to. Yes, please. Always good to be polite.
(Laughter)
And by this unified language interface on top of tools, the is able to sort of take away all those details you. So you don’t have to be the one who spells out single sort of little piece of what’s supposed to happen.
And I said, this is a live demo, so sometimes the unexpected happen to us. But let’s take a look at Instacart shopping list while we’re at it. And you can we sent a list of ingredients to Instacart. Here’s everything need. And the thing that’s really interesting is that the UI is still very valuable, right? If you look this, you still can click through it and sort modify the actual quantities. And that’s something that I think shows that they’re going away, traditional UIs. It’s just we have a new, way to build them. And now we have a that’s been drafted for our review, which is also a very important thing. We click “run,” and there we are, we’re the manager, we’re to inspect, we’re able to change the work of AI if we want to. And so after this talk, you will be able to this yourself. And there we go. Cool. Thank you, everyone.
(Applause)
So we’ll back to the slides. Now, the important thing about how we build this, it’s not just building these tools. It’s about teaching the AI how to use them. Like, what do we want it to do when we ask these very high-level questions? And to do this, we use old idea. If you go back to Alan Turing’s 1950 paper the Turing test, he says, you’ll never program an answer to this. Instead, you can learn it. You build a machine, like a human child, and then it through feedback. Have a human teacher who provides rewards and as it tries things out and does things that are either or bad.
And this is exactly how we train ChatGPT. It’s two-step process. First, we produce what Turing would have called a child machine through an unsupervised learning process. just show it the whole world, the whole internet and say, “Predict what comes next in text you’ve seen before.” And this process imbues it with all sorts wonderful skills. For example, if you’re shown a math problem, the only way to complete that math problem, to say what comes next, that green nine up there, is to solve the math problem.
But we actually have to do a step, too, which is to teach the AI what to with those skills. And for this, we provide feedback. We the AI try out multiple things, give us multiple suggestions, then a human rates them, says “This one’s better that one.” And this reinforces not just the specific that the AI said, but very importantly, the whole process the AI used to produce that answer. And this allows it to generalize. allows it to teach, to sort of infer your intent and apply it in scenarios that hasn’t seen before, that it hasn’t received feedback.
Now, sometimes the we have to teach the AI are not what you’d expect. For example, we first showed GPT-4 to Khan Academy, they said, “Wow, is so great, We’re going to be able to teach wonderful things. Only one problem, it doesn’t double-check students’ math. If there’s some math in there, it will happily pretend that one plus equals three and run with it.” So we had to collect some data. Sal Khan himself was very kind and offered 20 hours of his time to provide feedback to the machine alongside our team. And over the course of a couple months we were able to teach the AI that, “Hey, you really should push on humans in this specific kind of scenario.” And we’ve actually made lots and lots of improvements to the this way. And when you push that thumbs down in ChatGPT, that actually is kind like sending up a bat signal to our team say, “Here’s an area of weakness where you should gather feedback.” And when you do that, that’s one way that we really to our users and make sure we’re building something that’s more for everyone.
Now, providing high-quality feedback is a hard thing. If you think about a kid to clean their room, if all you’re is inspecting the floor, you don’t know if you’re teaching them to stuff all the toys in the closet. This is nice DALL-E-generated image, by the way. And the same sort of reasoning to AI. As we move to harder tasks, we will have to scale ability to provide high-quality feedback. But for this, the AI itself is happy to help. It’s to help us provide even better feedback and to scale ability to supervise the machine as time goes on. let me show you what I mean.
For example, you can ask GPT-4 a like this, of how much time passed between these two foundational blogs on unsupervised and learning from human feedback. And the model says two passed. But is it true? Like, these models are not 100-percent reliable, they’re getting better every time we provide some feedback. But can actually use the AI to fact-check. And it can actually its own work. You can say, fact-check this for me.
Now, in this case, I’ve actually the AI a new tool. This one is a tool where the model can issue search queries and click into web pages. And it actually out its whole chain of thought as it does it. It says, I’m just to search for this and it actually does the search. It then it finds publication date and the search results. It then is issuing search query. It’s going to click into the blog post. And all this you could do, but it’s a very tedious task. It’s not a that humans really want to do. It’s much more fun be in the driver’s seat, to be in this manager’s position you can, if you want, triple-check the work. And out come so you can actually go and very easily verify any piece of this chain of reasoning. And it actually turns out two months was wrong. Two and one week, that was correct.
(Applause)
And we’ll cut back the side. And so thing that’s so interesting to about this whole process is that it’s this many-step collaboration between a human and AI. Because a human, using this fact-checking tool is doing it in order to produce data another AI to become more useful to a human. And I think this really the shape of something that we should expect to be much more common in future, where we have humans and machines kind of very carefully and designed in how they fit into a problem and how we want to that problem. We make sure that the humans are providing the management, oversight, the feedback, and the machines are operating in a way that’s inspectable and trustworthy. together we’re able to actually create even more trustworthy machines. I think that over time, if we get this process right, we will be able to solve problems.
And to give you a sense of just impossible I’m talking, I think we’re going to be able to rethink almost aspect of how we interact with computers. For example, think about spreadsheets. They’ve around in some form since, we’ll say, 40 years ago with VisiCalc. I don’t think they’ve changed that much in that time. And here is a specific spreadsheet of all the AI on the arXiv for the past 30 years. There’s about 167,000 of them. And you see there the data right here. But let me you the ChatGPT take on how to analyze a data set like this.
So we can give ChatGPT to yet another tool, this one a Python interpreter, so it’s to run code, just like a data scientist would. so you can just literally upload a file and ask questions about it. very helpfully, you know, it knows the name of the file and it’s like, “Oh, this CSV,” comma-separated value file, “I’ll parse it for you.” The only here is the name of the file, the column names you saw and then the actual data. And from that it’s able to infer these columns actually mean. Like, that semantic information wasn’t in there. It has to sort of, together its world knowledge of knowing that, “Oh yeah, arXiv is a site that submit papers and therefore that’s what these things are and that these are integer and so therefore it’s a number of authors in the paper,” like all of that, that’s work a human to do, and the AI is happy to help with it.
Now I don’t even know I want to ask. So fortunately, you can ask the machine, “Can you make exploratory graphs?” And once again, this is a super high-level with lots of intent behind it. But I don’t know what I want. And the AI kind of has to infer I might be interested in. And so it comes with some good ideas, I think. So a histogram of the of authors per paper, time series of papers per year, cloud of the paper titles. All of that, I think, will be pretty interesting to see. the great thing is, it can actually do it. Here go, a nice bell curve. You see that three is kind the most common. It’s going to then make this nice of the papers per year. Something crazy is happening in 2023, though. like we were on an exponential and it dropped off cliff. What could be going on there? By the way, this is Python code, you can inspect. And then we’ll see word cloud. So you can see these wonderful things that appear in these titles.
But I’m unhappy about this 2023 thing. It makes this year look bad. Of course, the problem is that the year is not over. So I’m going to push on the machine. [Waitttt that’s not fair!!! 2023 isn’t over. percentage of papers in 2022 were even posted by 13?] So April 13 was the cut-off date I believe. Can you use to make a fair projection? So we’ll see, this is the of ambitious one.
(Laughter)
So you know, again, I like there was more I wanted out of the here. I really wanted it to notice this thing, maybe it’s a little of an overreach for it to have sort of, inferred that this is what I wanted. But I inject my intent, I this additional piece of, you know, guidance. And under the hood, AI is just writing code again, so if you want to what it’s doing, it’s very possible. And now, it the correct projection.
(Applause)
If you noticed, it even updates title. I didn’t ask for that, but it know what I want.
Now we’ll cut back to the again. This slide shows a parable of how I think we … A vision of how may end up using this technology in the future. person brought his very sick dog to the vet, and veterinarian made a bad call to say, “Let’s just wait and see.” the dog would not be here today had he listened. In the meanwhile, provided the blood test, like, the full medical records, to GPT-4, said, “I am not a vet, you need to talk a professional, here are some hypotheses.” He brought that information to a second vet used it to save the dog’s life. Now, these systems, they’re not perfect. You overly rely on them. But this story, I think, shows that a human with a medical and with ChatGPT as a brainstorming partner was able achieve an outcome that would not have happened otherwise. think this is something we should all reflect on, about as we consider how to integrate these systems our world.
And one thing I believe really deeply, is getting AI right is going to require participation from everyone. And that’s deciding how we want it to slot in, that’s for setting the of the road, for what an AI will and won’t do. And if there’s one thing take away from this talk, it’s that this technology just looks different. Just different from people had anticipated. And so we all have to literate. And that’s, honestly, one of the reasons we released ChatGPT.
Together, believe that we can achieve the OpenAI mission of ensuring that general intelligence benefits all of humanity.
Thank you.
(Applause)
(Applause ends)
Chris Anderson: Greg. Wow. I mean … I suspect that within every mind here there’s a feeling of reeling. Like, I suspect that a very number of people viewing this, you look at that and you think, “Oh my goodness, much every single thing about the way I work, need to rethink.” Like, there’s just new possibilities there. Am I right? thinks that they’re having to rethink the way that do things? Yeah, I mean, it’s amazing, but it’s really scary. So let’s talk, Greg, let’s talk.
I mean, I guess my question actually is just how the hell have you done this?
(Laughter)
OpenAI has a hundred employees. Google has thousands of employees working on artificial intelligence. is it you who’s come up with this technology shocked the world?
Greg Brockman: I mean, the truth is, we’re all building shoulders of giants, right, there’s no question. If you at the compute progress, the algorithmic progress, the data progress, all those are really industry-wide. But I think within OpenAI, we made a lot of very deliberate choices from early days. And the first one was just to confront reality as it lays. And that we thought really hard about like: What is it going to to make progress here? We tried a lot of that didn’t work, so you only see the things that did. And I think the most important thing has been to get teams of people who are very different from each other work together harmoniously.
CA: Can we have the water, by the way, just brought here? I think we’re to need it, it’s a dry-mouth topic. But isn’t there something just about the fact that you saw something in language models that meant that if you continue to in them and grow them, that something at some point emerge?
GB: Yes. And I think that, I mean, honestly, I the story there is pretty illustrative, right? I think high level, deep learning, like we always knew that was we wanted to be, was a deep learning lab, and exactly how to do it? I think in the early days, we didn’t know. We tried a lot things, and one person was working on training a model to the next character in Amazon reviews, and he got result where — this is a syntactic process, you expect, know, the model will predict where the commas go, the nouns and verbs are. But he actually got state-of-the-art sentiment analysis classifier out of it. This model could tell you a review was positive or negative. I mean, today we are like, come on, anyone can do that. But this the first time that you saw this emergence, this sort of semantics that emerged from underlying syntactic process. And there we knew, you’ve got to scale thing, you’ve got to see where it goes.
CA: So think this helps explain the riddle that baffles everyone looking at this, these things are described as prediction machines. And yet, we’re seeing out of them feels … it just feels impossible that could come from a prediction machine. Just the you showed us just now. And the key idea of emergence is that you get more of a thing, suddenly different things emerge. It all the time, ant colonies, single ants run around, when you enough of them together, you get these ant colonies that show completely emergent, different behavior. Or city where a few houses together, it’s just houses together. But as grow the number of houses, things emerge, like suburbs and cultural centers and traffic jams. Give one moment for you when you saw just something that just blew your mind that you just did not see coming.
GB: Yeah, well, so can try this in ChatGPT, if you add 40-digit numbers —
CA: 40-digit?
GB: 40-digit numbers, the will do it, which means it’s really learned an circuit for how to do it. And the really interesting thing is actually, if you have add like a 40-digit number plus a 35-digit number, it’ll often it wrong. And so you can see that it’s really learning the process, it hasn’t fully generalized, right? It’s like you can’t the 40-digit addition table, that’s more atoms than there in the universe. So it had to have learned general, but that it hasn’t really fully yet learned that, Oh, can sort of generalize this to adding arbitrary numbers of arbitrary lengths.
CA: So what’s here is that you’ve allowed it to scale up and at an incredible number of pieces of text. And it learning things that you didn’t know that it was to be capable of learning.
GB Well, yeah, and it’s nuanced, too. So one science that we’re starting to really good at is predicting some of these emergent capabilities. And to do actually, one of the things I think is very undersung this field is sort of engineering quality. Like, we had rebuild our entire stack. When you think about building a rocket, tolerance has to be incredibly tiny. Same is true machine learning. You have to get every single piece of the stack engineered properly, and then you start doing these predictions. There are all these incredibly scaling curves. They tell you something deeply fundamental about intelligence. If you look at GPT-4 blog post, you can see all of these in there. And now we’re starting to be able to predict. So we were to predict, for example, the performance on coding problems. basically look at some models that are 10,000 times 1,000 times smaller. And so there’s something about this is actually smooth scaling, even though it’s still early days.
CA: So is, one of the big fears then, that arises from this. If it’s fundamental to what’s here, that as you scale up, things emerge that you maybe predict in some level of confidence, but it’s capable of surprising you. Why isn’t just a huge risk of something truly terrible emerging?
GB: Well, I think all of these questions of degree and scale and timing. And I one thing people miss, too, is sort of the integration the world is also this incredibly emergent, sort of, very thing too. And so that’s one of the reasons that we it’s so important to deploy incrementally. And so I think that what kind of see right now, if you look at this talk, a lot of I focus on is providing really high-quality feedback. Today, the tasks that do, you can inspect them, right? It’s very easy to look that math problem and be like, no, no, no, machine, seven the correct answer. But even summarizing a book, like, that’s hard thing to supervise. Like, how do you know if this book summary any good? You have to read the whole book. No one wants to that.
(Laughter) And so I think that the important will be that we take this step by step. And that say, OK, as we move on to book summaries, we have supervise this task properly. We have to build up a track record with these machines that they’re able actually carry out our intent. And I think we’re going to have to produce even better, more efficient, reliable ways of scaling this, sort of like making machine be aligned with you.
CA: So we’re going to hear later this session, there are critics who say that, you know, there’s real understanding inside, the system is going to always — we’re never going to know that it’s not errors, that it doesn’t have common sense and so forth. Is it your belief, Greg, that it is true any one moment, but that the expansion of the scale and the human feedback that you about is basically going to take it on that journey of actually getting to things like and wisdom and so forth, with a high degree of confidence. you be sure of that?
GB: Yeah, well, I that the OpenAI, I mean, the short answer is yes, I believe is where we’re headed. And I think that the OpenAI approach here has always just like, let reality hit you in the face, right? It’s like this field is the field of promises, of all these experts saying X is going to happen, Y is it works. People have been saying neural nets aren’t going to work for 70 years. haven’t been right yet. They might be right maybe 70 years plus one something like that is what you need. But I think our approach has always been, you’ve got to push to the limits of technology to really see it in action, because that you then, oh, here’s how we can move on to a new paradigm. And just haven’t exhausted the fruit here.
CA: I mean, it’s quite a controversial stance you’ve taken, the right way to do this is to put it out in public and then harness all this, you know, instead of your team giving feedback, the world is now giving feedback. But … If, know, bad things are going to emerge, it is out there. So, you know, the original story I heard on OpenAI when you were founded as a nonprofit, well you were there as great sort of check on the big companies doing their unknown, possibly evil thing with AI. And were going to build models that sort of, you know, somehow them accountable and was capable of slowing the field down, if need be. Or at least that’s of what I heard. And yet, what’s happened, arguably, is the opposite. That your release of GPT, ChatGPT, sent such shockwaves through the tech world that Google and Meta and so forth are all scrambling catch up. And some of their criticisms have been, are forcing us to put this out here without proper or we die. You know, how do you, like, make the that what you have done is responsible here and not reckless.
GB: Yeah, think about these questions all the time. Like, seriously the time. And I don’t think we’re always going to get it right. But one I think has been incredibly important, from the very beginning, when we were about how to build artificial general intelligence, actually have it benefit all humanity, like, how are you supposed to do that, right? And that plan of being, well, you build in secret, you get this powerful thing, and then you figure out the safety of it and you push “go,” and you hope you got it right. I don’t how to execute that plan. Maybe someone else does. But for me, was always terrifying, it didn’t feel right. And so I think this alternative approach is the only other path that I see, which that you do let reality hit you in the face. And think you do give people time to give input. You have, before these machines are perfect, before they are super powerful, you actually have the ability to see them in action. And we’ve seen from GPT-3, right? GPT-3, we really were afraid that the one thing people were going to do with it was misinformation, try to tip elections. Instead, the number one was generating Viagra spam.
(Laughter)
CA: So Viagra spam is bad, but there are things are much worse. Here’s a thought experiment for you. you’re sitting in a room, there’s a box on the table. believe that in that box is something that, there’s a strong chance it’s something absolutely glorious that’s going to give beautiful to your family and to everyone. But there’s actually also a percent thing in the small print there that says: “Pandora.” there’s a chance that this actually could unleash unimaginable evils on the world. you open that box?
GB: Well, so, absolutely not. I think you don’t do it that way. honestly, like, I’ll tell you a story that I haven’t actually before, which is that shortly after we started OpenAI, I I was in Puerto Rico for an AI conference. I’m sitting in hotel room just looking out over this wonderful water, all these people having a time. And you think about it for a moment, if you could choose for basically that Pandora’s box be five years away or 500 years away, which would you pick, right? On the hand you’re like, well, maybe for you personally, it’s better to have it be five years away. if it gets to be 500 years away and get more time to get it right, which do you pick? And know, I just really felt it in the moment. was like, of course you do the 500 years. brother was in the military at the time and like, puts his life on the line in a much more way than any of us typing things in computers and this technology at the time. And so, yeah, I’m really sold on the you’ve got approach this right. But I don’t think that’s quite playing the field it truly lies. Like, if you look at the history of computing, I really mean it when I say that this is industry-wide or even just almost like a human-development- of-technology-wide shift. And the more you sort of, don’t put together the pieces that are there, right, we’re still faster computers, we’re still improving the algorithms, all of these things, are happening. And if you don’t put them together, you get an overhang, means that if someone does, or the moment that someone manage to connect to the circuit, then you suddenly have this powerful thing, no one’s had any time to adjust, who knows kind of safety precautions you get. And so I think that thing I take away is like, even you think about development of other sort technologies, think about nuclear weapons, people talk about being like zero to one, sort of, change in what humans could do. But I actually think if you look at capability, it’s been quite smooth over time. And so the history, I think, of technology we’ve developed has been, you’ve got to do it and you’ve got to figure out how to manage it for moment that you’re increasing it.
CA: So what I’m hearing is that you … the model you want to have is that we have birthed this extraordinary child that may superpowers that take humanity to a whole new place. It is our responsibility to provide the guardrails for this child to collectively teach it to be wise and to tear us all down. Is that basically the model?
GB: think it’s true. And I think it’s also important to say may shift, right? We’ve got to take each step as we it. And I think it’s incredibly important today that we all do get literate in this technology, out how to provide the feedback, decide what we from it. And my hope is that that will to be the best path, but it’s so good we’re having this debate because we wouldn’t otherwise if it weren’t there.
CA: Greg Brockman, thank you so much for coming to and blowing our minds.
(Applause)