Sam Altman on OpenAI's Core AI Subscription Vision and Road to AGI

Source material

Transcript

Hi, and welcome to Training Data.

We are mixing it up for this week's episode and dropping a conversation that was filmed live at Sequoia's annual AI conference in San Francisco with OpenAI founder and CEO Sam Altman.

Sam is interviewed by our partner Alfred Lin.

We hope you enjoyed this special conversation with Sam about the Genesis story of ChatGPT, his predictions for agents in scientific discovery and robotics, and more.

And stay tuned for a few more special AI Ascent drops on our podcast feed later this week.

Our next guest needs no introduction, so I'm not going to bother introducing him—Sam Altman.

I will just say Sam is now 3 for 3 and joining us to share his thoughts at the 3 AI Ascent that we've had, which we really appreciate.

So I just want to say thank you for being here.

This was our first office.

That's right.

Oh, that's right.

Say that again.

Yeah, this was our first office, so it's nice to be back.

Let's go back to the first office here.

It started in 2016.

2016.

We just had Jensen here who said that he delivered the first GGX1 system over here.

He did, yeah.

It's amazing how small that thing looks now.

Oh, versus what?

Well, the current boxes are so huge, but yeah, it was a fun throwback.

How heavy was it?

That was still when you could kind of like lift one yourself.

He said it was about 70 pounds.

Yeah, I mean, it was heavy, but you could carry it.

So did you imagine that you'd be here today in 2016?

No, it was like we were sitting over there and there were 14 of us or something.

And you were hacking on this new system?

I mean, even now it was like we were sitting around like looking at whiteboards trying to talk about what we should do.

Like this was a—it's almost impossible to sort of overstate how much we were like a research lab with no—with a very strong belief in direction and conviction, but no real kind of like action plan.

I mean, not only was like the idea of a company or a product sort of unimaginable, the specific like LLMs is an idea.

We're still very far off.

And so it took— We're like trying to play video games.

Trying to play video games.

Are you still trying to play video games?

Now we're pretty good at that.

All right.

So it took you another six years for the first consumer product to come out, which is ChatGPT.

Along the way, how did you sort of think about milestones to get something to that level?

As like an accident of history, the first consumer product was not ChatGPT.

That's right.

It was Dolly.

The first product was the API.

So we had built—you know, we kind of went through a few different things.

We had a few directions that we really wanted to bet on.

Eventually, as I mentioned, we said, well, we got to build a system to see if it's working.

We're not just writing research papers.

We're going to see if we can play a video game well.

We're going to see if we can do a robot hand.

We're going to see if we can do a few other things.

And at some point in there, one person and then initially—and then eventually a team got excited about trying to do unsupervised learning to build language models.

And that led to GPT-1 and then GPT-2.

And by the time of GPT-3, we both thought we had something that was kind of cool, but we couldn't figure out what to do with it.

And also, we realized we needed a lot more money to keep scaling.

You know, we had done GPT-3.

We wanted to go to GPT-4.

We were heading into the world of billion-dollar models.

It's like hard to do those as a pure science experiment unless you're like a particle accelerator or something, even though it's hard.

So, we started thinking, okay, we both need to figure out how this can become a business that can sustain the investment that it requires.

And also, like, we have a sense that this is heading towards something actually useful.

And we had put GPT-2 out as model weights, and not that much had happened.

One of the things that I had just observed about companies, products in general is if you do an API, it usually works somehow on the upside.

This was, like, true across many, many YC companies.

And also that if you make something much easier to use, there's usually a huge benefit to that.

So, we're like, well, it's kind of hard to run these models that are getting big.

We'll go write some software, do a really good job of running them.

And also, we'll then, rather than build a product because we couldn't figure out what to build, we will hope that somebody else finds something to build.

And so, I forget exactly when, but maybe it was, like, June of 2020, we put out GPT-3 in the API.

And the world didn't care, but Silicon Valley did.

They're like, oh, this is kind of cool.

This is poignant.

It's something.

And there was this weird thing where we got almost no attention from most of the world.

And some startup founders were like, oh, this is really cool.

Or, like, I mean, some of them are like, this is AGI.

The only people that built real businesses with the GPT-3 API that I can remember were these company, a few companies that did, like, copywriting as a service.

That was kind of the only thing GPT-3 was over the economic threshold on.

But one thing we did notice, which eventually led to chat GPT, is even though people couldn't build a lot of great businesses with the GPT-3 API, people loved to talk to it in the playground.

And it was terrible at chat.

We had not at that point figured out how to do RLHF to make it easy to chat with.

But people loved to do it anyway.

And in some sense, that was the kind of only killer use other than copywriting of the API product that led us to eventually build chat GPT.

By the time GPT-3.5 came out, there were maybe, like, eight categories instead of one category where you could build a business with the API.

But our conviction that, like, people just want to talk to the model had gotten really strong.

So we had done Dolly, and Dolly was doing okay.

But we knew we kind of wanted to build, especially along with the fine tuning we were able to do, we knew we wanted to build this model, this product to let you talk to the model.

And it launched in 2022?

I think, yes.

Yeah, about six years from when the first...

November 30th, 2022.

So there's a lot of work leading up to that.

And 2022 launched today.

It has over 500 million people who talked to it on a weekly basis.

Yeah.

All right.

All right.

So, by the way, get ready for some audience questions, because that's what that was sans request.

You've been here for three every single one of the senses Pat mentioned, and there's been some lots of ups and downs.

But it seems like the last six months, it's just been shipping, shipping, shipping.

You shipped a lot of stuff, and it's amazing to see the product velocity, the shipping velocity continue to increase.

So this is like multi sort of part question.

How have you gotten a large company to like increase product velocity over time?

I think a mistake that a lot of companies make is they get big and they don't do any...

They don't do more things.

So they just like get bigger because you're supposed to get bigger and they still ship the same amount of product.

And that's when the molasses really takes hold.

I am a big believer that you want everyone to be busy.

You want teams to be small.

You want to do a lot of things relative to the number of people you have.

Otherwise, you just have like 40 people in every meeting and huge fights over who gets what tiny part of the product.

There was this old observation of business that a good executive is a busy executive, because you want people like muddling around.

But I think it's like a good...

At our company and many other companies, like researchers, engineers, product people, they drive almost all the value.

And you want those people to be busy and high impact.

So if you're going to grow, you better do a lot more things.

Otherwise, you kind of just have a lot of people sitting in a room fighting or meeting or talking about whatever.

So we try to have relatively small numbers of people with huge amounts of responsibility.

And the way to make that work is to do a lot of things.

And also, we have to do a lot of things.

I think we really do now have an opportunity to go build one of these important internet platforms.

But to do that, if we really are going to be people's personalized AI that they use across many different services and over their life and across all of these different main categories and all the smaller ones that we need to figure out how to enable, then that's just a lot of stuff to go build.

Anything you're particularly proud of that you've launched in the last six months?

The models are so good now.

They still have areas to get better, of course, and we're working on that fast.

But I think at this point, chat GBT is a very good product because the model is very good.

I mean, there's other stuff that matters too.

But I am amazed that one model can do so many things so well.

You're building small models and large models.

You're doing a lot of things, as you said.

So how do this audience stay out of your way and not be roadkill?

I mean, I think the way to model us is we want to build -- we want to be people's core AI subscription and way to use that thing.

Some of that will be like what you do inside of chat GBT.

We'll have a couple of other kind of really key parts of that subscription.

But mostly we will hopefully build this smarter and smarter model.

We'll have these surfaces like future devices, future things that are sort of similar to operating systems, whatever.

And then we want -- we have not yet figured out exactly, I think, what the sort of API or SDK or whatever you want to call it is to really be our platform.

But we will.

It may take us a few tries, but we will.

And I hope that that enables just an unbelievable amount of wealth creation in the world and other people to build onto that.

But yeah, we're going to go for the core AI subscription and the model.

And then the core surfaces and there will be a ton of other stuff to build.

So don't be the core AI subscription, but you can do everything else.

We're going to try.

If you can make a better core AI subscription offering than us, go ahead.

That would be great.

It's rumored that you're raising $40 billion or something like that at $340 billion valuation.

It's rumored.

I don't know if -- I think we announced it.

I just want to make sure that you announced it.

What's your scale of ambition from here?

We're going to try to make great models and ship good products.

There's no master plan beyond that.

We're going to -- Sure.

No, I see plenty of open AI people in the audience.

They can vouch for it.

We don't sit there and have -- I am a big believer that you can kind of do the things in front of you.

But if you try to work backwards from we have this crazy complex thing, that doesn't usually work as well.

We know that we need tons of AI infrastructure.

We know we need to go build out massive amounts of AI factory volume.

We know that we need to keep making models better.

We know that we need to build a great top of the stack kind of consumer product and all the pieces that go into that.

But we pride ourselves on being nimble and adjusting tactics as the world adjusts.

The products that we're going to build next year we're probably not even thinking about right now.

And we believe we can build a set of products that people really, really love.

And we have unwavering confidence in that.

And we believe we can build great models.

I've actually never felt more optimistic about our research roadmap than I do right now.

What's on the research roadmap?

Very smart models.

But in terms of the steps in front of us, we kind of take those one or two at a time.

So you believe in working forwards, not necessarily working backwards?

I have heard some people talk about these brilliant strategies of how this is where they're going to go and they're going to work backwards.

And this is take over the world.

And this is the thing before that.

And this is that.

And this is that.

And this is that.

And this is that.

And here's where we are today.

I have never seen those people really massively succeed.

Got it.

Who has a question?

There's a mic coming your way.

Being thrown.

What do you think the larger companies are getting wrong about transforming their organizations to be more AI native in terms of both using the tooling as well as producing products?

It's been, you know, it's smaller companies are clearly just beating the crap out of out of larger ones when it comes to innovation here.

I think this basically happens every major tech revolution.

There's nothing to me surprising about it.

The thing that they're getting wrong is the same thing they always get wrong, which is like people get incredibly stuck in their ways.

Organizations get incredibly stuck in their ways.

If things are changing a lot every quarter or two and you have like an information security council that meets once a year to decide what applications are going to allow and what it means to like put data into a system like it's just it's so painful to watch what happens here.

But like, you know, this is this is creative destruction.

This is why startups win.

This is like how the industry moves forward.

I am I'd say I feel like disappointed but not surprised at the rate that big companies are willing to do this.

They will.

My kind of prediction would be that there's another like couple of years of fighting pretending like this isn't going to reshape everything.

And then there's like a capitulation and a last minute scramble and it's sort of too late.

And in general startups just sort of like blow past people doing it the old way.

I mean, this happens to people too, like watching watching like a you know, someone who started maybe you like talk to an average 20 year old and watch how they use chat GPT and then you go talk to like an average 35 year old and how they use it or some other service.

And like the difference is unbelievable.

It reminds me of like, you know, when the smartphone came out and like every kid was able to use it super well and older people would just like took like three years to figure out how to do basic stuff.

And then of course people integrate.

But but the the sort of like generational divide on AI tools right now is crazy.

And I think companies are just another symptom of that.

Anybody else have a question?

Just to follow up on that.

What are the cool use cases that you're seeing young people using with chat GPT that might surprise us?

They really do use it like an operating system.

They have like complex ways to set it up to connect it to like a bunch of files and they have like fairly complex prompts memorized in their head or like, you know, in something where they paste in and out.

And the I mean, that stuff I think is all cool and impressive.

And there's this other thing where like they don't really make life decisions without asking like chat GPT what they should do.

And it has like the full context on every person in their life and what they've talked about.

And you know, that like the memory thing has been a real change there.

But but yeah, I think it grows oversimplification.

But like older people use chat GPT as a Google replacement.

Maybe people in their 20s and 30s use it as like a life advisor something and then like people in college use it as an operating system.

How do you use it inside of open AI?

I mean, it writes a lot of our code.

How much?

I don't know the number.

And also when people say the number, I think is always this very dumb thing because like Microsoft code is 3020 30% measuring by lines of code is just such an insane way to like I don't I maybe the meaningful thing I could maybe the thing I could say is it's writing meaningful code.

Like it's right.

I don't know how much but it's like writing the parts that actually matter.

That's interesting.

Next question.

Hey, Sam.

Oh, Mike, going is this Okay, hey, Sam.

I thought it was interesting that the answer to Alfred's question about where you guys want to go is focused mostly around consumer and being the core subscription.

And also most of your revenue comes from consumer subscriptions.

Why keep the API in 10 years?

I really hope that all of this merges into one thing.

Like you should be able to sign in with open AI to other services.

Other services should have an incredible SDK to like take over the chat GPT UI at some point.

But like to the degree that you are going to have a personalized AI that knows you that has your information that knows what you want to share later and you know has all this context on you.

You'll want to be able to use that in a lot of places.

Now I agree that the current version of the API is very far off that vision.

But I think we can get there.

Yeah, maybe I have a follow up question to that when you kind of took mine.

But like a lot of us who are building application layer companies, we want to like use those building blocks, those different API components, maybe the deep research API, which is not a released thing, but could be and build stuff with them like, is that going to be a priority like enabling that platform for us?

How should we think about that?

Yeah, I I think I hope something in between those that there is sort of like a new protocol on the level of HTTP for the future of the internet where things get federated and broken down into like much smaller components and agents are like constantly exposing and using different tools and authentication, payment, data transfer.

It's all like built in at this level that everybody trusts everything, talk to everything.

And I don't quite think we know what that looks like.

But it's like coming out of the fog.

And as we get a better sense for that, again, it'll probably take us like a few iterations toward that to get there.

But that's kind of where I would like to see things go.

Hey, Sam, back here.

My name is Roy, I'm curious.

The AI would obviously do better with more input data.

Is there any thought to feeding sensor data?

And what type of sensor data, whether it's temperature, you know, things in the physical world that you could feed in that it could better understand reality?

People do that a lot.

People like put that into, you know, people have whatever they build things where they just put sensor data into like an API and like an O3 API call or whatever.

And for some use cases, it does work super well.

I'd say that the latest models seem to do a good job with this and they used to not.

So we'll probably bake it in more explicitly at some point.

But there's already like a lot happening there.

Hi, Sam.

I was really excited to play with the voice model in the playground.

And so I have two questions.

The first is how important is voice to open AI in terms of like stack ranking for infrastructure?

And can you share a little bit about how you think it'll show up in the product and chat GPT the core thing?

I think voice is extremely important.

Honestly, we just we have not made a good enough voice product yet.

That's fine.

Like it took us a while to make a good enough text model to we will crack that code eventually.

And when we do, I think a lot of people are going to want to use voice interaction a lot more.

I am super when we first launched our current voice mode, the thing that was most interesting to me was it was a new stream on top of like the touch interface and I you could talk and be like clicking around on your phone at the same time.

And I continue to think there's something amazing to do about like voice plus gooey interaction that we have not cracked.

But before that, we'll just make voice really great.

And when we do, I think there's not only is it cool with existing devices, but I sort of think voice will enable a totally new class of devices if you can make it feel like truly human level voice.

A question about coding.

I'm curious, is coding just another vertical application or is it more central to the future of open AI?

That one's more central to the future of open AI.

Coding, I think, will be how these models kind of right now, if you ask chat GPT a response, you get text back, maybe you get an image.

You would like to get a whole program back.

You would like, you know, custom rendered code for every response, or at least I would.

You would like the ability for these models to go make things happen in the world.

And writing code, I think will be very central to how you like actuate the world and call a bunch of APIs or whatever.

So I would say coding will be more in a central category will obviously expose it through API on our platform as well.

But you know, chat GPT should be excellent at writing code.

So we're gonna move from the world of assistance to agents to basically applications all the way through.

I think it'll feel, yeah, like very continuous.

But yes.

So you have conviction in the roadmap about smarter models.

Awesome.

I have this mental model there's some ingredients like more data, bigger data centers, a transformer architecture test time compute, what's like an underrated ingredient or something that's going to be part of that mix that like maybe isn't in the mental model of most of us.

I mean, that's kind of the each of those things are really hard.

And you know, obviously, like the highest leverage thing is still big algorithmic breakthroughs.

And I think there still probably are some 10 x's or 100 x's left, not very many, but even one or two is a big deal.

But, you know, yeah, it's kind of like algorithms, data, compute, those are sort of the big ingredients.

Hi.

So my question is, you run one of the best ML teams in the world.

How do you balance between letting smart people like Isa chase deep research or something else that seems exciting versus going top down and being like, we're going to build this, we're going to make it happen.

We don't know if it'll work.

There are some projects that require so much coordination that there has to be a little bit of like top down quarterbacking.

But I think most people try to do way too much of that.

I mean, this is like, there's probably other ways to run good AI research or good research labs in general.

But when we started open AI, we spent a lot of time trying to understand what a well run research lab looks like.

And you had to go really far back in the past.

In fact, almost everyone that could like help advise us on this was dead.

It had been like a long time since there had been good research labs.

And people ask us a lot like why does open AI repeatedly innovate and why do the other AI labs copy or why do bio lab X not do good work and bio lab Y does do good work or whatever.

And we sort of keep saying like, here's the principles we've observed.

Here's how we learned them.

Here's what we looked at in the past.

And then everybody says great, but I'm going to go do the other thing.

That's fine.

Like you came to us for advice.

Like you do what you want.

But I find it remarkable how much these few principles that we've tried to run our research lab on, which we did not invent, we shamelessly copied from other good research labs in history, have worked for us.

And then people who have had some smart reason about why they were going to do something else that didn't work.

So it seems to me that these large models one of the really fascinating things is like a lover of knowledge about them is that they potentially embody and allow us to answer these like amazing, long standing questions in the humanities about cyclical changes in artistic, interesting things, or even like, you know, to what extent systematic prejudice and other sorts of things are really happening in society.

And can we sort of detect these and very subtle things which we could never really do more than hypothesize before.

And I'm wondering whether open AI has a thought about or even a roadmap for working with academic researchers, say, to help unlock some of these new things we could learn for the first time in the humanities and in the social sciences.

We do.

Yeah, I mean, it's amazing to see what people are doing there.

We do have academic research programs where we partner and, you know, do some custom work.

But mostly people just say like, I want access to the model or maybe I want access to the base model.

And I think we're really good at that.

One of the kind of cool things about what we do is so much of our incentive structure is pushed towards making the models as smart and cheap and widely accessible as possible that that serves academics and really the whole world very well.

So, you know, we have we do some custom partnerships, but we often find that what researchers or users really want is just for us to make the general model better across the board.

And so we try to focus, you know, kind of 90 percent of our thrust vector on that.

I'm curious how you're thinking about customization.

So you mentioned the federated like signing with open AI, bringing your memories, your context.

I'm just curious if you think customization and like these different post training on like application specific things is a band-aid for or trying to make the core models better and how you're thinking about that.

I mean, in some sense, I think that like platonic ideal state is a very tiny reasoning model with a trillion tokens of context that you put your whole life into.

The model never retrains, the weights never customized, but that thing can like reason across your whole context and do it efficiently.

And every conversation you've ever had in your life, every book you've ever read, every email you've ever read, every everything you've ever looked at is in there, plus connected to all your data from other sources.

And, you know, your life just keeps a pending to the context and your company just does the same thing for all your company's data.

We can't get there today.

But I think of kind of like anything else is a compromise off that platonic ideal.

And that is how I would eventually I hope we do customization.

One last question in the back.

Hi, Sam.

Thanks for your time.

Where do you think most of the value creation would come from in the next 12 months?

Would it be maybe advanced memory capabilities or maybe security or protocols that allow agents to do more stuff and interact with the real world?

I mean, in some sense, the value will continue to come from really three things like building on more infrastructure, smarter models, and building the kind of scaffolding to integrate this stuff into society.

And if you push on those, I think the rest will sort itself out.

At a higher level of detail, I kind of think 2025 will be a year of sort of agents doing work, coding in particular, I would expect to be a dominant category.

I think there'll be a few others too.

Next year is a year where I would expect more like sort of AI's discovering new stuff.

And maybe we have AI's make some very large scientific discoveries or assist humans in doing that.

And I am kind of a believer that most of the sort of real sustainable economic growth in human history comes from once you've like kind of spread out and colonized the earth, most of it comes from just better scientific knowledge and then implementing that for the world.

And then 27, I would guess is the year where like that all moves from the sort of intellectual realm to the physical world and robots go from a curiosity to like a serious economic creator of value.

But that was like an off the top of my head kind of guess right now.

Can I close with a few quick questions?

One of which is GPT-5.

Is that going to be just all smart in all of us here?

I mean, if you think you're like way smarter than 03, then maybe you have a little bit of a ways to go.

But 03 is already pretty smart.

Two personal questions.

Last time you were here, you just come off a blip with open AI.

Given some perspective now in distance, you got any advice for founders here about resilience, endurance, strength?

It gets easier over time.

I think like you will face a lot of adversity in your journey as a founder and the kind of challenges get harder and higher stakes, but the emotional toll gets easier as you kind of go through more bad things.

So it's, you know, in some sense like it does.

Yeah, even though like abstractly the challenges get bigger and harder, your ability to deal with them, the sort of resilience you build up gets easier like with each one you kind of go through.

And then I think the hardest thing about the big challenges that come as a founder is not the moment when they happen.

Like a lot of things go wrong in the history of a company.

In the acute thing, you can kind of like, you know, you get a lot of support, you can function a lot of adrenaline, like that's, you know, you're kind of like, even the really big stuff like your company runs out of money and fails, like a lot of people will come and support you.

And you kind of get through it and go into the new thing.

The thing that I think is harder to sort of manage your own psychology through is the sort of like fallout after.

And I think if there's, you know, people focus a lot about how to work in that one moment during the crisis and the really valuable thing to learn is how you like pick up the pieces.

There's much less talk about that.

I think there's, I've never actually found something good to point founders to to go read about, you know, not how you deal with the real crisis on day zero or day one or day two, but on day 60, as you're just trying to like rebuild after it.

And that's, that's the area that I think you can like practice and get better at.

Thank you, Sam.

Yeah, you're officially still on paternity leave.

I know.

So thank you for coming in and speaking with us.

Appreciate it.

Thank you.