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