In a talk from the cutting edge of technology, OpenAI cofounder Greg Brockman explores the underlying design principles of ChatGPT and demos some mind-blowing, unreleased plug-ins for the chatbot that sent shockwaves across the world. After the talk, head of TED Chris Anderson joins Brockman to dig into the timeline of ChatGPT's development and get Brockman's take on the risks, raised by many in the tech industry and beyond, of releasing such a powerful tool into the world.

We started OpenAI 7 years ago because we felt like something really interesting was happening in AI.

We wanted to help steer it in a positive direction.

It's honestly just really amazing to see how far this whole field has come since then.

And it's really gratifying to hear from people like Raymond, who are using the technology we are building and others, for so many wonderful things.

We hear from people who are excited, we hear from people who are concerned, we hear from people who feel both those emotions at once.

Above all, it feels like we're entering an historic period right now, where we as a world are going to define a technology that will be so important for our society going forward.

And I believe that we can manage this for good.

So today I want to show you the current state of that technology and some of the underlying design principles that we hold dear.

So the first thing I'm going to show you is what it's like to build a tool for an AI rather than building it for a human.

So we have a new DALI model, which generates images, and we are exposing it as an app for chat GPT to use on your behalf.

And you can do things like ask, you know, suggest a nice post-TED meal and draw a picture of it.

Now you get all of the sort of ideation and creative back and forth and taking care of the details for you that you get out of ChatGPT.

It's not just the idea for the meal, but a very, very detailed spread.

But chat GPT doesn't just generate images in this case.

It doesn't just generate text, it also generates an image.

And that is something that really expands the power of what it can do on your behalf in terms of carrying out your intent.

And I'll point out, this is all live demo.

This is all generated by the AI as we speak.

So I actually don't even know what we're going to see.

Now I'm getting hungry just looking at it.

Now we've extended chatGPT with other tools too.

And the interesting thing about these tools is they're very inspectable.

So you get this little pop-up here that says, use the DALI app.

And by the way, this is coming to all ChatGPT users over upcoming months.

And you can look under the hood and see that what it actually did was write a prompt, just like a human could.

And so you sort of have this ability to inspect how the machine is using these tools, which allows us to provide feedback to them.

And let me show you what it's like to use that information and to integrate with other applications too.

You can say, now make a shopping list for the tasty thing I was suggesting earlier.

I'm going to make it a little tricky for the AI.

And tweet it out for all the TED viewers out there.

So if you do make this wonderful, wonderful meal, I definitely want to know how it tastes.

But you can see that ChatGBT is selecting all these different tools without me having to tell it explicitly which ones to use in any situation.

And this, I think, shows a new way of thinking about the user interface.

Like, we are so used to thinking of, well, we have these apps, We click between them, we copy-paste between them.

And usually it's a great experience within an

of know the menus and know all the options.

And by having this unified language interface on top of tools, the AI is able to sort of take away all those details from you.

So you don't have to be the 1 who spells out every single sort of little piece of what's supposed to happen.

And as I said, this is a live demo, so sometimes the unexpected will happen to us.

But let's take a look at the Instacart shopping list while we're at it.

And you can see we sent a list of ingredients to Instacart.

And the thing that's really interesting is that the traditional UI is still very valuable.

If you look at this, you still can click through it and modify the actual quantities.

And that's something that I think shows that they're not going away, traditional UIs.

It's just we have a new augmented way to build them.

And now we have a tweet that's been drafted for our review, which is also a very important thing.

We're able to change the work of the AI if we want to.

And so after this talk, you will be able to access this yourself.

Now the important thing about how we build this, it's not just about building these tools.

It's about teaching the AI how to use them.

Like, what do we even want it to do when we ask these very high-level questions?

If you go back to Alan Turing's 1950 paper on the Turing test, he says, look, you'll never program an answer to this.

You could build a machine like a human child and then teach it through feedback, have a human teacher who provides rewards and punishments as it tries things out and does things that are either good or bad.

And this is exactly how we train ChachiPT.

First, we produce what Turing would have called a child machine through an unsupervised learning process.

We just show it the whole world, the whole internet, and say, predict what comes next in text you've never seen before.

And this process imbues it with all sorts of wonderful skills.

For example, if you're shown a math problem, the only way to actually complete that math problem, say, what comes next, that green 9 up there, is to actually solve the math problem.

But we actually have to do a second step, too, which is to teach the AI what to do with those skills.

We have the AI try out multiple things, give us multiple suggestions, and then the human rates them, says, this one's better than that 1.

And this reinforces not just the specific thing that the AI said, but very importantly, the whole process that the AI used to produce that answer.

And this allows it to generalize, it allows it to teach to sort of infer your intent and apply it in scenarios that it hasn't seen before, that it hasn't received feedback.

Now, sometimes the things we have to teach the AI are not what you'd expect.

For example, when we first showed GPT-4 to Khan Academy, they said, wow, this is so great, we're going to be able to teach students wonderful things.

Only 1 problem, it doesn't double-check students' math.

If there's some bad math in there, it will happily pretend that 1 plus 1 equals 3 and run with it.

Sal Khan himself was very kind and offered 20 hours of his own time to provide feedback to the machine alongside our team.

And over the course of a couple of months, we were able to teach the AI that, hey, you really should push back on humans in this specific kind of scenario.

And we've actually made lots and lots of improvements to the models this way.

And when you push that thumbs down in chat GPT, that actually is kind of like sending up a bat signal to our team to say, here's an area of weakness where you should gather feedback.

And so when you do that, that's 1 way that we really listen to our users and make sure we're building something that's more useful for everyone.

Now, providing high-quality feedback is a hard thing.

If you think about asking a kid to clean their room, if all you're doing is inspecting the floor, you don't know if you're just teaching them to stuff all the toys in the closet.

This is a nice dolly-generated image, by the way.

And the same sort of reasoning applies to AI.

As we move to harder tasks, we will have to scale our ability to provide high-quality feedback.

But for this, the AI itself is happy to help.

It's happy to help us provide even better feedback and to scale our ability to supervise the machine as time goes on.

For example, you can ask for, you know, a GPD 4 question like this, of how much time passed between these 2 foundational logs on unsupervised learning and learning from human feedback, and the model says 2 months passed.

These models are not 100 percent reliable, although they're getting better every time we provide some feedback.

But we can actually use the AI to fact-check, and it can actually check its own work.

Now in this case, I've actually given the AI a new tool.

This 1 is a browsing tool where the model can issue search queries and click into web pages.

And it actually writes out its whole chain of thought as it does it.

It says, I'm going to search for this, and it actually does the search.

It then finds the publication date in the search results.

It then is issuing another search query, it's going to click into the blog post.

And all of this you could do, but it's a very tedious task.

It's not a thing that humans really want to do.

It's much more fun to be in the driver's seat, to be in this manager's position, where you can, if you want, triple-check the work and outcome citations, so you can actually go and very easily verify any piece of this whole chain of reasoning, and it actually turns out 2 months was wrong, 2 months in 1 week.

so the thing that's so interesting to me about this whole process is that it's this many-step collaboration between a human and an AI, because a human using this fact-checking tool is doing it in order to produce data for another AI to become more useful to a human.

And I think this really shows the shape of something that we should expect to be much more common in the future, where we have humans and machines kind of very carefully and delicately designed in how they fit into a problem and how we want to solve that problem.

We make sure that the humans are providing the management, the oversight, the feedback, and that the machines are operating in a way that's inspectable and trustworthy.

And together, we're able to actually even create even more trustworthy machines.

And I think that over time, if we get this process right, we will be able to solve impossible problems.

And to give you a sense of just how impossible I'm talking, I think we're going to be able to rethink almost every aspect of how we interact with computers.

They've been around in some form since, let's say, 40 years ago with Visicalc.

I don't think they've really changed that much in that time.

And here is a specific spreadsheet of all the AI papers on the archive.

For the past 30 years, there's about 167, 000 of them.

But let me show you the chat GPT take on how to analyze a data set like this.

So we can give chat GPT access to yet another tool, this 1, a Python interpreter, so it's able to run code, just like a data scientist would.

And so you can just literally upload a file and ask questions about it.

And very helpfully, it knows the name of the file.

And it's like, oh, this is CSV, a comma-separated value file.

The only information here is the name of the file, the column names, like you saw, and then the actual data.

And from that, it's able to infer what these columns actually mean.

Like that semantic information wasn't in there.

It has to sort of put together its world knowledge of knowing that, oh yeah, archive is the site that people submit papers, and therefore that's what these things are, and that these are integer values, and so therefore it's the number of authors in the paper.

All of that, that's work for a human to do, and the AI is happy to help with it.

Now, I don't even know what I want to ask, so fortunately, you can ask the machine, can you make some exploratory graphs?

And once again, this is a super high-level instruction with lots of intent behind it, but I don't even know what I want, and the AI kind of has to infer what I might be interested in.

And so it comes up with some good ideas, I think.

So a histogram of the number of authors per paper, time series of papers per year, word cloud of the paper titles.

All of that I think will be pretty interesting to see.

And the great thing is, it can actually do it.

You see that 3 is kind of the most common.

It's going to then write a, it's going to make this nice plot of the papers per year.

Something crazy is happening in 2023, though.

Looks like we were on an exponential and it dropped off a cliff.

And by the way, all this is Python code you can inspect.

And then we'll see the word cloud, so you can see all these wonderful things that appear in these titles.

But I'm pretty unhappy about this 2023 thing.

Of course, the problem is that the year's not over.

So I'm going to push back on the machine.

So April 13th is the cutoff date, I believe.

So we'll see, this is a kind of ambitious 1.

So, you know, again, I feel like there was more I wanted out of the machine here.

Maybe it's a little bit of an overreach for it to have inferred magically that this is what I wanted.

I provide this additional piece of guidance.

And under the hood, the AI is just writing code again, so if you want to inspect what it's doing, it's very possible.

And if you notice, it even updates the title.

I didn't ask for that, but it knows what I want.

This slide shows a parable of how I think we, you know, a vision of how we may end up using this technology in the future.

A person brought his very sick dog to the vet, and the veterinarian made a bad call to say, let's just wait and see, and the dog would not be here today had he listened.

In the meanwhile, he provided the blood tests, like the full medical records, to GPT-4, which said, I am not a vet, you need to talk to a professional, here are some hypotheses.

He brought that information to a second vet, who used it to save the dog's life.

But this story, I think, shows that the human with a medical professional and with Chats GPT as a brainstorming partner was able to achieve an outcome that would not have happened otherwise.

I think this is something we should all reflect on and think about as we consider how to integrate these systems into our world.

And 1 thing I believe really deeply is that getting AI right is going to require participation from everyone.

And that's for deciding how we want it to slot in, that's for setting the rules of the road for what an AI will and won't do.

And if there's 1 thing to take away from this talk, It's that this technology just looks different, just different from anything people had anticipated.

And so we all have to become literate, and that's honestly 1 of the reasons we released ChatGBT.

Together, I believe that we can achieve the OpenAI mission of ensuring that artificial general intelligence benefits all of humanity.

I mean, I suspect that within every mind out here, apart from there's a feeling of reeling, like I suspect that a very large number of people viewing this, you look at that and you think, oh my goodness, pretty much every single thing about the way I work, I need to rethink.

Who thinks that they're having to rethink the way that we do things?

Yeah, I mean, it's amazing, but it's also really scary.

We have, I mean, I guess my first question actually is just how the hell have you done this?

Google has thousands of employees working on artificial intelligence.

Why is it you who's come up with this technology that's shocked the world?

Yeah, well, I mean, the truth is we're all building on shoulders of giants, right?

There's no question if you look at the compute progress, the algorithmic progress, the data progress, all of those are really industry-wide.

But I think within OpenAI, we made a lot of very deliberate choices from the early days.

And the first 1 was just to confront reality as it lays.

And that we just sort of thought really hard about what is it going to take to make progress here.

We tried a lot of things that didn't work, so you only see the things that did.

And I think that the most important thing has been to get teams of people who are very different from each other to work together harmoniously.

Can we have the water, by the way, just brought here?

I think we're going to need it, it's a dry mouth topic.

But isn't there something also, just about the fact that you saw something in these language models that meant that if you continue to invest in them and grow them, that something at some point might emerge.

And I think that I mean, honestly, I think the story there is pretty illustrative.

I think that at a high level, deep learning, We always knew that was what we wanted to be, it 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 of things, and 1 person was working on training a model to predict the next character in Amazon reviews, And he got a result where this is a syntactic process.

You expect the model will predict where the commas go, where the nouns and verbs are, but he actually got a state-of-the-art sentiment analysis classifier out of it that this model could tell you if a review was positive or negative.

Today, we are just like, come on, anyone could do that.

But this was the first time that you saw this emergence, this sort of semantics that emerged from this underlying syntactic process.

And there we knew, you've got to scale this thing, you've got to see where it goes.

CAWTHON So I think this helps explain the riddle that baffles everyone looking at this, because these things are described as prediction machines, and yet what we're seeing out of them feels impossible, that that could come from a prediction machine, just the stuff you showed us just now.

And the key idea of emergence is that when you get more of a thing, suddenly different things emerge.

When you bring enough of them together, you get these ant colonies that show completely emergent and different behavior.

Or a city where a few houses together is just houses together, but as you grow the number of houses, things emerge, like suburbs and cultural centers and traffic jams.

Give me 1 moment for you, when you saw just something pop that just blew your mind, that you just did not see coming?

Yeah, well, so you can try this in chat, JPT.

If you add 40-digit numbers, 40-digit numbers, the model will do it, which means it's really learned an internal circuit for how to do it.

And the really interesting thing is, actually, if you have it add, like, a 40-digit number plus a 35-digit number, it'll often get it wrong.

And so you can see that it's really learning the process, but it hasn't fully generalized, right?

It's like, you can't memorize the 40-digit addition table.

That's more atoms than there are in the universe.

So it had to have learned something general, but that it hasn't really fully yet learned that, oh, I can, like, sort of generalize this to adding arbitrary numbers of arbitrary lengths.

CAWTHON, So what's happened here is that you've allowed it to scale up and look at an incredible number of pieces of text, and it is learning things that you didn't know that it was going to be capable of learning.

of the things I think is very undersung in this field is sort of engineering quality.

Like we had to rebuild our entire stack and get, you know, like when you think about building a rocket, like you know every tolerance has to be like incredibly tiny.

You have to get every single piece of the stack engineered properly.

And then you can start doing these predictions.

There are all these incredibly smooth scaling curves that tell you something deeply fundamental about intelligence.

If you look at our GPT-4 blog post, you can see all of these curves in there.

And now we're starting to be able to predict.

So we're able to predict, for example, the performance on coding problems.

We basically look at some models that are 10, 000 times or 1, 000 times smaller.

And so there's something about this that is actually smooth scaling, even though it's still early days.

If it's fundamental to what's happening here, that as you scale up, things emerge, that you can maybe predict in some level of confidence, but it's capable of surprising you, why isn't there just a huge risk of something truly terrible emerging?

PM- Well, I think all these are questions of degree and scale and timing.

And I think 1 thing people miss, too, is sort of the integration with the world is also this, like, incredibly emergent, like, sort of very powerful thing, too.

And so that's 1 of the reasons that we think it's so important to deploy incrementally.

And so I think that what we kind of see right now, if you look at this talk, a lot of what I focus on is providing really high-quality feedback.

Today the tasks that we do, you can inspect them, right?

It's very easy to look at that math problem and be like, no, no, no, machine, like, 7 was the correct answer.

But even summarizing a book, that's a hard thing to supervise.

How do you know if this book summary is any good?

See all Greg Brockman transcripts on Tedtalk

The inside story of ChatGPT's astonishing potential