Sam, Jakub, and Wojciech on the future of OpenAI with audience Q&A

Hello, I'm Sam. This is our chief

scientist, Yakob. And we have a bunch of

updates to share today about OpenAI. Um,

obviously the news of today is our new

structure. We're going to get to that

near the end, but there's a lot of other

important context we would like to share

first. Given the importance of a lot of

this, we're going to go into uh an

unusual level of transparency about some

of our specific research goals and

infrastructure plans and product, but we

think it's uh you know sort of very much

in the public interest at this point to

cover all of this. So our mission at

OpenAI in both the nonprofit and our new

PBC is to ensure that artificial general

intelligence benefits all of humanity.

As we get closer to building this, we

have new insights into what that is

going to mean.

There was a time earlier on in OpenAI

where we thought that AI would or AGI

would be sort of this oracular thing in

the sky and it would make all these

wonderful things for us and we now have

a sharper view of that which is we want

to create tools and then we want people

to use them to create the future. We

want to empower people with AI as much

as possible and then trust that the

process that has been working for human

history of people building um better and

better things with newer and better

tools will continue to go on. We we can

now see a vision where we help build a

personal AGI that people can use

anywhere with all of these different

tools, access to all these different

services and systems to help with work

and personal life and their personal

life. And as AI gets better and better,

as AI can even do things like discover

or help discover new science, what

people will be able to create with that

um to make all of society better and

their own lives more fulfilled, we think

should be quite incredible.

There are three core pillars we think

about for OpenAI. Research, product, and

infrastructure. We have to succeed at

the research required to build AGI. We

have to build a platform that makes it

easy and powerful to use. And we have to

build enough infrastructure such that

people can use at a low cost all of this

amazing AI that they'd like. Here's a

little like cartoon of how we think

about our world. So, at the bottom layer

here, we have chips, racks, and the

systems around them, the data centers

that these go into, and the energy. Uh

we'll talk more about the first three

today in energy another time. Then we

train models on top of these. Then we

have an open AI account on top of that.

We have a browser now uh called Atlas

and we have devices coming um in the

next few years that you'll be able to

take AI with you everywhere. And we have

a few first party apps like Chachi Pine

and Sor and we'll have more over time.

But mostly what we're excited about is

this big puzzle piece in the upper

right. We're finally getting to a world

where we can see that people are going

to be able to build incredible services

uh with AI starting with our API with

apps and chatbt new enterprise platform

that we'll have over time uh an open

account and way way more and people will

be able to fit all of the kind of like

current things in the world and many

more into this new AI world and we want

to enable that and we we believe the

world will build just a huge amount of

value for um for all of us. So that's

kind of what we see the economic picture

looking like. But one of the things that

we've thought about for a long time and

we really see happening now or starting

to happen now, glimmers of it, green

shoots, whatever you want to call it, is

the uh impact that AI will have on

science. And um although the economic

impact from that previous slide will be

huge for the long-term quality of life

uh and improvement that and change in

society, AI that can autonomously

discover new science or help people

discover new science uh faster will be I

think one of the most important things

and something that we're really trying

to wrap our heads around. So I'm going

to hand this over to Yakob to talk about

research and as I mentioned uh we're

going to share a lot about our internal

goals and our picture of where thing

where things are.

Thanks Sam. Um at the core we are a

research laboratory focused on

understanding a technology called deep

learning. And so a particular focus of

ours is um understanding uh what happens

as you scale up training deep learning

systems.

Um and one consequence we discuss a lot

there uh is AGI artificial general

intelligence. But we find that um in

some way even this maybe understates a

bit um the magnitude of um the possible

progress and change here. Um and so in

particular we believe uh that it is

possible that deep learning systems are

less than a decade away from super

intelligence. So systems that are

smarter uh than all of us on

uh a large number of of of of critical

axis

and

um this is of course a a serious thing

right there's uh a lot of implications

of this to grapple with and one uh

particular focusing impact of this

technology um and the technologies

leading up to it and something that we

organize our entire research program

around is the potential uh to accelerate

scientific discovery um to accelerate

the development of new technology.

Um

we believe that this will be uh perhaps

the most significant uh long-term impact

of AI development. um and it will

fundamentally change the pace of

progress on on on developing new

technologies.

Um and so thinking about how far along

we are uh uh towards these goals, uh one

good way to think about progress is to

look at the time horizon that it would

take people to accomplish the task that

the models can perform. And so this is

something that has been extending

rapidly over the past few years. Um so

where the current generation of models

is at right now is about five hours. So

um if you you can see this by looking at

the models matching the best people in

competition such as the international

Olympics or informatics

and we believe that this horizon will

continue to extend rapidly and this is

in part as a result of algorithmic

innovation and a part uh just scaling

deep learning further and in particular

scaling um along this um new axis um in

context compute also called test time

compute uh where we really see orders

and orders of magnitude to go. Um so

this is roughly how much time the model

spends thinking right and if you look at

how much time the model's currently

spent thinking about problems and if you

think about how much compute how much

time you would like to spend on problems

that really matter such as scientific

breakthroughs you should be okay using

entire data centers. Uh and so there is

there is really quite a way to go there.

Um and anticipating this progress uh we

of course make plans around it

internally and we want to provide some

transparency around our thinking there

and so we want to take this maybe

somewhat unusual step of sharing our

internal goals and goal timelines uh

towards these very powerful systems and

you know these particular dates we

absolutely may be quite wrong about them

uh but this is how we currently think

this is currently how how how we plan

and organize and So as a research

organization that is working on

automating research, naturally we are

thinking about how does this impact our

own work and uh how will AI systems that

accelerate development of future AI

systems look like? How can they empower

um research like alignment? And so uh we

are making plans around getting to quite

capable AI research interns that can

meaningfully accelerate our researchers

by expanding

a a significant amount of compute um by

September of next year. So we believe

that is actually quite close. Um and

then we uh look towards getting a system

capable of autonomously delivering on

larger research projects and a

meaningful uh fully automated AI

researcher uh by March of 2028.

Um and so of course as we look towards

these very capable systems uh we think a

lot about safety and alignment right and

in fact a lot of our work uh both on

deployments and safety uh but also just

on on on understanding deep learning and

development capability side we can think

of preparation for these very capable

models. Um safety is a multiaceted

problem and so the way we generally

structure our thinking uh are these five

layers ranging from uh factors that are

most internal to the model to ones that

are most external. And so at the core

right and what we believe is the most

important uh long-term safety question

for super intelligence is value

alignment.

So to put this um

this

value alignment uh you can think of as

what is what is really the thing that

the AI fundamentally cares about uh can

it adhere to some high level principles

u what will it do if it's given unclear

and conflicting objectives u does it

laugh humanity

um and the reason we believe that kind

of this this high level uh um

um

objectives or or or or principles

driving the AI are so important is that

as we get to the systems that are

thinking for very long, uh as they

become very smart, as they tackle

problems that are uh um at the edge or

perhaps beyond human ability, uh really

getting to like complete specifications

becomes quite difficult. Uh and so we

have to rely on this on this deeper

alignment.

Um then there's goal alignment. Um, does

the agent uh interact with people? How

does it interact with people? How does

it um do it following instructions? Um,

then reliability. Um, can the AI

correctly calibrate its predictions? Uh,

can it be reliable on easy tasks,

express uncertainty on hard ones, can it

deal with environments that are a little

bit unfamili unfamiliar?

Um then we have adversarial robustness

which is very related to reliability but

it's about adversarial settings. So can

can the AI withstand target attacks from

human or AI adversaries.

And then the outer layer is systemic

safety which are guarantees about the

behavior of the overall system that do

not rely on the AI's intelligence or

alignment. Um so for example this can be

security or what data does the AI have

access to? um or um um uh um what

devices it can use. And so we invest in

multiple research directions across

these domains. And we have seen uh quite

a lot of progress also come from just

the general uh development and improving

understanding of deep learning uh as a

whole. And uh I want to I I want to take

a slightly deeper technical dive here.

Um and talk about a particular

direction. Uh value alignment is a hard

problem, right? It's definitely not

solved yet. Um however, there is a new

promising tool that aids our study of

it. Uh and that is chain of fun

faithfulness. Uh it's something we

invest in very heavily. Um, starting

from our first reasoning models, we've

been pursuing this new direction in

interpretability. And the idea is to

keep parts of the model's internal

reasoning free from supervision. So

don't look at it during training and

thus let it remain representative of the

model's internal process. Um,

so we refrain from from from from kind

of guiding the model to think good

thoughts and and and and and so let it

let let it remain a bit more faithful to

to to what it actually thinks, right?

And this is not guaranteed to work, of

course, right? We cannot make uh

mathematical proofs about deep learning.

And so this is something we study. Uh

but there are two reasons to be

optimistic. One reason is that we have

seen very promising empirical results.

Uh this is a technology we employed a

lot internally. Uh we use this to

understand um how our models u um train

h how their propensities evolve over

training. Uh also we have had successful

external collaborations on investigating

the models propensity scheme for

example.

Um and secondly uh it is scalable and in

the sense that explicitly we make the

scalable objective not adversarial to

our ability to monitor the model. Um

and of course an objective not being

adversarial to the ability to monitor

the model is only half the battle. Um

and you know ideally you want it to to

to get it to help with monitoring the

model. And so this is something we're

we're researching quite heavily.

Um but one important thing to underscore

about train of thought faithfulness is

it's somewhat fragile. Um it really

requires drawing this clean boundary uh

and having this clear abstraction uh and

having restraint in in what ways you can

access the chain of thought and this is

something that is present uh at OpenAI

from algorithm design to the way we

design our products right so so if you

look at the chain of thought summaries

in chat GPT uh if we didn't have the

chain of summarizer if we just make the

chain of f fully visible at all times

right that would make it kind part of

the overall experience over time it will

be very difficult to not subjected to

any supervision.

Um and so longterm we believe that by

preserving some amount of this

controlled privacy for the models uh we

can retain the ability to understand

their inner process and we believe this

can be a very impactful technique uh as

we move towards these very capable

longunning systems. Um and I'll hand

back to Tom.

>> Okay, that's very hard to follow uh with

the rest of this and obviously that's

the most important part of what we have

to say. But um you know just to to

reiterate uh we may be totally wrong. We

have set goals and missed them miserably

before. But with the picture we see we

think it is plausible that by September

of next year we have sort of a intern

level AI research assistant and that by

March of 2028 which I believe is almost

5 years to the month after the launch of

GPT4. um we have like a a legitimate AI

researcher and this is the core thrust

of our research program. There are two

other areas we want to talk about uh

product and then infrastructure. On the

product side,

as we make this incredible progress with

deep learning, we want to make it useful

to people to sort of invent the future

as we mentioned. And um what that's

looked like traditionally for us as an

AI super assistant inside of Chacht, but

we're now really going to evolve to a

platform uh that other people will build

on top of and all of the pieces that

need to fit in uh of the world um you

know will be built by others. Before we

go talk about that, wanted to just show

a quick video of how people are using

GPT5. Some of the ways people are using

GPT5 in Chacht today.

>> I'm a quantum physicist.

>> I'm a nail technician, human

iminologist.

>> I'm a steel worker,

>> a designer and developer, a professor of

economics.

>> I basically go fishing for a living.

>> GPT5 is able to predict the outcomes of

experiments that we haven't even done.

Can you create a part numbering system

that's easy for my guys in the shop?

>> I want to catch dungeonous crab in the

Bay Area. Here I would ask a question

about the application of a certain

quantum operator. This one gives me a

very detailed mathematics.

>> I ask for basically a camera app where I

can draw real time in the air.

>> We have a theme. We have a direction to

go. Let's go from a million or infinite

amount of ideas to like give me 10 or

20.

>> You just start testing it. You know, you

tell it jokes. You ask it questions like

what should economists do?

>> Different baits I can use, different

water depths, all this information that

it would take years to figure out on

your own. A lot of trial and error. It's

tremendous for brainstorming. It's back

and forth.

>> I can kind of easily follow the

reasoning. I don't need to trust the

result. I can just look what did you do?

>> GBD5 just like did this um in one shot.

>> A 101 part numbers. It would have taken

weeks for me to number this and it would

have made me go crosseyed.

>> Uh Yakob will join me back for Q&A in a

little bit, but we're going to have one

special guest on before before the end.

Um we love that. We want much more of

that. We want that everywhere. So we

want open AI to be a platform that

people and companies can build on top

of. We can sort of see our way now to an

AI cloud where this is not just in chat

GPD. This is not just services that we

create, but we want to expose our

technology for as many people to build

the things that people will depend on

and use and create with as possible. I

think this this quote or at least this

like idea is originally from Bill Gates.

um at least that's where I first heard

it that you know you've built a platform

when there's more value created by

people building on the platform than by

the platform builder builder and that's

our goal um and that's our goal like

next year we we really think we can now

take this technology and this user base

and this sort of framework we've built

and get the whole world to build amazing

new companies and services and

applications on top of it

to do that uh there will be many things

that we have to evolve towards but

there's two foundational principles as

we as we move towards being in this

platform that I wanted to touch on. Um,

one is about user freedom. If this is

going to be a platform that all sorts of

people are building, uh, on using, um,

creating with, people around the world

have very different needs and desires

and there will of course be some very

broad bounds, but we want users to have

a lot of u control and customization of

how they of how they use it. Now, I

made, you know, one of my many stupid

mistakes when I tried to talk about this

recently. Uh, I wish I had used an

example other than erotica. thought

there was an understandable difference

between erotica and pornbots. But in any

case, we were trying to show the point

we're trying to get across is that

people need a lot of flexibility and

people want to use these things in

different ways and we want to treat our

adult users like adults. In our own

first party services, we may have, you

know, tighter guidelines, but AI is

going to become such an important part

of people's lives. The freedom of human

expression is going to need to be there.

Along with that, we think that world

will need to think about privacy in a

different way than they have for

previous kinds of technology. Privacy is

important for all sorts of technology,

of course, but privacy for AI will be

especially important. People are using

this technology in a different way than

they've used the technologies of the

past. They're talking to it like they

would to their doctor, their lawyer,

their spouse. Um, they're sharing the

most intimate details of their lives.

And of course, we need strong technical

protections on that privacy of that

privacy. But we also think we need

strong um policy protections of that

privacy. We've talked about concepts

like AI privilege. Um but really strong

protections if AI is going to be this

fundamental platform in people's lives

seem super important to us.

Okay. And then I want to go on to

infrastructure. So I know there's been a

lot of confusion about sort of where we

are in our infrastructure buildout and

we figured we would just be super

transparent about that. So where we are

today, um all of our commitments total a

little bit over 30 gawatts of uh of

infrastructure buildout. Um and that's

about a $1.4 trillion total uh financial

obligation for us over the next many

years. Um this is what we've committed

to so far. We of course hope to do much

more, but given the picture we see

today, given what we think we can see

for revenue growth, our ability to raise

capital, this is what we're currently

comfortable with. This requires a ton of

partnerships. We've talked about uh many

of our great chip partners. Uh there are

people building the data centers for us,

land, energy. Uh there will be chip fab

facilities. This is already getting to

require quite a lot of supply chain

innovation. And we're thrilled to get to

work with uh AMD, Broadcom, Google,

Microsoft Nvidia Oracle SoftBank

many others to really make this happen.

But this is still early. if the if the

work that Yakob talks about comes to

fruition, which we think it will, and if

the economic value of these things um

happen and people want to use all these

services, we're going to need much more

than this. So, I want to be clear, we're

not committing to this yet, but we are

having conversations about it. Our

aspiration is that we can build an

infrastructure factory where we can

create 1 gawatt a week of uh compute and

we aspirationally would like to get that

cost down significantly um to like $20

billion a gigawatt over the 5year life

cycle of uh you know that equipment. To

do this will require a ton of

innovation, a ton of partnerships,

obviously a lot of revenue growth. Um

we'll have to repurpose our thoughts

about robotics to help us build data

centers instead of doing all the other

things. Um, but this is where we'd like

to go and over the coming months we are

going to do a lot of work to see if we

can get here. Um, it will be some time

before we're in a financial position

where we could actually pull the trigger

and get going on this. 1 gawatt is like

a big number, but I figured we would

show a little video to put this into

perspective. Um, this is a data center

that we're building in Abene, Texas.

This is the first Stargate site. We're

doing several of these now around the

country, but this one is the furthest

along. There's like many thousands of

people that work here every day just

doing the construction at the site.

There's probably hundreds of thousands

or millions of people that work in the

supply chain to make all this happen to

design these chips, to fab these chips,

to put them together. Um there, you

know, there's all of the work that goes

into this for energy. There's an

enormous amount of stuff that has to

happen um for each one gigawatt. And we

want to figure out how we can make this

way more efficient and way cheaper and

way more scalable so that we can deliver

on the infrastructure that the research

roadmap requires and that all of the

ways that people will want to use this

need um to enable that uh we have a new

structure. So um maybe you saw before

this like crazy convoluted uh diagram of

all of the open entities. Now it's much

simpler. We have a nonprofit called the

open air foundation that is in control

of uh where the board sits uh or where

uh let's come back to the board uh where

the board also sits and uh owns a slice

of our PBC public benefit corporation

called OpenAI group. So uh nonprofit and

control public benefit corporation sits

under it. Um we hope for the OpenAI

Foundation to be the biggest nonprofit

ever. As I mentioned now uh a few times,

science is one of the ways that we think

the world most improves along with the

institutions that broadly distribute the

benefits of that. So the science will

not be the only thing that the nonprofit

funds, but it will be an important first

major area of the things that we do. The

nonprofit will govern the PBC. It will

initially own about 26% of the PBC

equity, but that can increase over time

with warrants if if we perform really

well. and it will use these resources to

uh pursue what we think are the best

benefits of AI uh given where the

technology is and what society needs.

The PBC uh will operate more like a

normal company. It will have the same

mission. Um it will be bound to that

mission. Um and you know in matters of

safety will only be bound to that

mission but it will be able to attract

the resources that we need for that

gigantic infrastructure buildout to

serve the research and product goals

that we have. So, the initial focus of

the the foundation, we'll do more things

over time, but we want to knock

something out of the out of the park,

uh, hopefully first, is a $25 billion

commitment to use AI to help cure

disease. There are a lot of ways this

can happen. Um, generating data, um, you

know, using a lot of compute grants to

grants to scientists and also for AI

resilience. Um AI resilience is a new

and I think very important area and I'd

like to invite uh our co-founder of

Voyche up to talk about what this will

what this will look like.

>> Hello

glad to be here.

>> Thanks for being here.

>> So um term AI resilience is um little

bit broader than what we historically

thought about AI safety. So in case of

the uh resilience we think that advanced

AI comes with risks and disruptions and

we would like to have an ecosystem of

organizations that uh can help to um

solve a number of these problems. Um so

let me give you an example to better

illustrate it. Um we all believe that AI

uh will advance in biology and as it

advances in biology there is a there is

a risk that some bad actor could use AI

to create manmade pandemics. So the on

the safety level the mitigation would be

to make sure that the models uh block

the queries that have to do with a

viology. Um however um if you consider

the entire entire um AI uh industry it's

very likely that even if open AI blocks

it someone could use uh different models

out there and still uh produce pathogens

and um the in case of resilience we

don't want just to block it but also

have a uh rapid response if the problem

would occur. So when I think about the

uh risks uh and disruptions there are

just many the the mental health is one

of them bio is another one another one

uh job displacement might be another one

and we think that we need the ecosystem

and maybe a good analogy that I like is

cyber security so at the beginning of

internet

uh it was actually it was a place that

people didn't feel comfortable putting

their credit card numbers because it was

so easy to get hacked and uh when there

was a virus people were giving each

other a call to disconnect the computer

from internet and we got a long way at

the moment there's entire infrastructure

of cyber security companies they are

protecting uh um the critical

infrastructure governments corporations

and individual users to such extent that

people are willing to put the most

personal data uh online to have life

savings um be online. Um yeah, so the

the cyber security got really far and we

think that something analogous will be

present for AI that there will be AI

resilience layer and uh I'm really

excited that the uh nonprofit uh will

help out uh to stimulate it to create

such an ecosystem.

>> So am I. I think this is an important

time to be doing this and I'm very

excited to that you're going to like

figure out how how we how we go off and

make it happen. So again, these are not

the only things that the nonprofit will

fund, but we're excited about these as

the first two using AI to develop cures

and uh treatments for diseases and this

new AI resilience effort as we figure

out what the deployment of AGI into

society is going to look like.

So we mentioned that those are our three

pillars, but you know what what if this

all works? We we think it is plausible

that in 2026 we start to see the models

of that year begin to make small

discoveries. By 2028 medium or maybe

even larger discoveries and you know who

knows what 2030 and 2032 are going to

look like. If AI can keep advancing

science as has happened in the past um

we think the future can be very bright.

Of course, we think it's very important

that humans can self-determine our way

through this future. But the open space

that new scientific advances give us is

quite impressive. So, we asked Sora to

help us imagine a radically better

future by looking at the past. And we

are particularly interested in how the

history of science builds on itself

discovery after discovery. This is what

we hope will happen with AI. You know,

this is going to be 200 years of

science. But if you can do these 200

years of compounding discoveries of the

scaffolding building up on each other

not in 200 years but in 20 years or in

two years and if you look at how much

this this is accelerated think about

what could be possible before you can

imagine a world uh where a radically

better future becomes quite possible.

You have a data center here that is

discovering a cure for cancer. A data

center there that's making the best

entertainment ever. A data center here

that's helping you find your future

husband or wife. This one is building

rockets and helping you colonize space.

This one is helping to solve the climate

crisis. So, we did all this stuff the

oldfashioned way. Um, and now with the

help of AI, we'll be able to shape what

comes next, uh, with maybe much more

power. So, we talked a little about AI

medicine. We're very excited about

robots. Um, we really think energy is

very, very important to the world. We

want to figure out what personalized

education can mean, design novel

materials, and probably a ton of other

things that we can't even think of yet.

So as we head into this next phase of

open AI and more importantly than that

this continual progress in deep learning

um we thank you for joining us today and

we're going to try something new now

which is we're going to just answer

questions. If this works it's something

we are we'll try more in the future.

Yakob is going to rejoin for this Q&A.

Thank you very much.

>> Um but uh this is a new format for us.

So bear with us as we try it this first

time again. If if um if this is useful

it's something we'll do uh again a lot

more. and we're going to try to just

answer questions in the order they are

most upvoted. Um,

are we good to go? All right, let's see

how this works. So, uh, you can put

questions in the Vimeo link and we will

just start answering them. So, from

Caleb, we've warned that the tech is

becoming addictive and eroding the tech

is becoming addictive and eroding trust

yet Sora mimics Tik Tok and Chach may

add ads. Why repeat the same patterns

you criticized and how will you build

build rebuild trust through actions and

not just words? We're definitely worried

about this. Uh I worry about it not just

for things like Sora and Tik Tok and ads

and chatbt which are maybe known

problems that we can design carefully

but you know we have certainly seen

people develop relationships with chat

bots that we didn't expect and there can

clearly be addictive behavior there

given the dynamics and competition in

the in the world. I suspect some

companies will offer very addictive new

kinds of products. Um and I think you'll

have to just judge us on our actions.

We'll have to you know we'll make some

mistakes. We'll try to roll back models

that are problematic. If we ship Sora

and it becomes super addictive and not

about creation, we'll, you know, cancel

the product and you'll you'll have to

just judge us on that. My hope and

belief is that we will not make the same

mistakes that companies before us have

made. Uh I don't think they meant to

make them either. It's uh you know,

we're all kind of discovering this

together. We probably will make new ones

though and we'll just have to evolve

quickly and have a tight feedback loop.

We we can imagine all sorts of ways this

technology does incredible good in the

world. also obvious bad ones and um you

know we're guided by a mission where

we'll just continuously evolve evolve

the product. Um one thing that we are

quite hopeful about for in terms of um

what we optim optimize for in in uh

products like chat GPT or or Sora is

thinking about um optimizing for the

very long term which is naturally very

aligned with uh how we think in general

about extending uh the horizon on which

the models can work productively. Um

and so we believe that quite a lot of

development is possible there and we can

eventually get the models that really

optimize for um long-term uh

satisfaction and and and well-being

instead of just short-term signals.

Um okay, next question. Will we have an

option to keep the for model uh

permanently after adult mode is

installed? We don't need a safer models

responsible adults. Um,

we have no plans to sunset 40. Uh, we

are not going to promise to keep it

around till the heat death of the

universe either, but we we understand

that it's a product that some of our

users really love. We also hope other

people understand why um it was not a

model that we thought was healthy for

miners to be using. Um, we hope that we

build better models over time that

people like more. You know, the people

you have a relationship with in your

life, they evolve and get smarter and

change a little bit over time. And we

think that we hope that the same thing

will happen. But yeah, no, no plans to

uh no plans to sunset 40 currently.

Uh wow, we have a lot of for questions.

All right, we're not going to in the

interest of time, we will not go through

uh all of these, but but yeah, we don't,

you know, we want people to have models

that they want to use. We don't want

people to feel like we're routing them

around models. Um, and you know, we we

want adults to make choices as adults

as long as we think we're not, you know,

selling heroin or whatever, which also

you you shouldn't you shouldn't do. Um,

so people that want to have emotional

speech, uh, as we've said, we want to

allow more of that and we plan to. Okay,

here's a good anonymous question for

Yakob. When will AGI happen?

Um

so

I think I think

in in some number of years we'll look

back at these years and we'll say you

know this was kind of the transition

period when AGI happened. Um I think you

know what one way we thought about

um I think as as some said like early on

adopting we thought about AGI kind of

emotionally as this like thing that is

like the kind of ultimate solution of

all the problems and and and um it's

it's this like single point um for which

there is before and after and um I think

um we found that it's a bit more

continuous than that. Um

and and so in particular for like

various kind of benchmarks that you know

seemed at uh um

seem like kind of the obvious like

milestones towards AGI. I think I think

we now think of them as kind of like

indicating like you know roughly how far

away we are in years. And so uh you know

if you look at a succession of of of of

milestones such as computers beating

humans at chess and then at go and then

uh you know computers being able to

speak in natural language and computers

being able to solve math problems right

I think well they clearly kind of get uh

closer together. Um

yeah, I I would say I think it's the AGI

term has become hugely overloaded and as

Jakob said, it'll be this process over a

number of years that we're in the middle

of. Uh but one of the reasons we wanted

to present what we did today is I think

it's much more useful to say our

intention our goal is by March of 2028

to have a true automated AI researcher

um and define what that means uh than it

is to sort of try to you know satisfy

everyone with a definition of AGI

>> and and maybe one other thing to mention

right like I think like one kind of

counterintuitive thing here is that

obviously we're working with like a

pretty complicated technology we're

trying to understand all these

algorithms and maybe initially we kind

imagined that like AGI is the moment

once you where you kind of have figured

out all the answers and and and and it's

kind of the the final thing and I think

now we increasingly realize that you

know there is kind of a some some curve

of intelligence maybe a

multi-dimensional one and you know

humans are somewhere on it and as you

scale deep learning as you develop these

new algorithms like eventually well you

kind of inch closer to that point and

eventually and eventually will surpass

it and you know already have surpassed

on multiple axis um and and and so and

that doesn't actually mean you have

solved all the problems around it which

is something we we need to seriously

think about.

>> Can you give us an idea of how far ahead

internal models are compared to deployed

ones?

>> Um

I think

um

we we have quite strong expectations for

our for for for our next models. Um so I

think I think we we expect quite rapid

progress over the next uh uh couple

months and a year. Um

yeah I think um but we

we haven't been like withholding

something something uh extremely crazy.

>> Yeah. One of the ways this uh kind of

often works in practice is there's like

a lot of pieces that we develop and that

you know they're all kind of hard one

victories and then we know that when we

put them together um we will have

something quite impressive and we're

able to predict that fairly well. Uh

part of our goal today is to say that um

we have a lot of those pieces. It's not

like we're kind of currently sitting on

this giant, you know, thing that we're

not showing to the world, but that we

expect by a year from now, certainly

with this September of 2026 goal

that we have a a like I mean not likely,

we have a realistic shot at a like

tremendously

important step forward in capability.

Um, what is OpenAI? Ronin asks, "What is

OpenAI's stance on partnering with labs

like Anthropic, Gemini, or XAI for joint

research, compute sharing, and safety

efforts?" Uh, we think this is going to

be increasingly important on the safety

front. Uh, labs will need to share

safety techniques, safety standards. Um,

you can imagine a time when the whole

world would say, uh, okay, before we hit

a recursive self-improvement phase, we

really need to all carefully study this

together. Um, we welcome that

collaboration. I think it'll be quite

important. Um one thing to mention on

the on the on chain of thought

faithfulness that I talked about earlier

um we actually have um started talking

about establishing industry arms and and

we started some some joint

investigations with researchers from

Google and entropic uh and and and some

other labs and uh yeah that's something

I'm very excited about and I think that

is an example of something where we can

really benefit from collaborating across

multiple labs.

Anonymous asks, "Will you ever open

source some of your old models like the

original GPT4?" Um, we might do those as

museum artifacts someday, but they're

not like GPT4

is not a particularly useful open source

model. It's big. It's not that good. Uh,

you know, we could probably make

something that is beyond the power of

GPT4 at a very tiny scale. Uh, that

actually would be useful to people. So,

for useful things, I expect more things

like that. For uh for fun museum

artifacts, yeah, someday, who who knows?

Like I think there could be a lot of

cool things like that. Another anonymous

or maybe the same one asks, "Will you

admit that your new model is inferior to

the previous one and that you're ruining

your company with your arrogance and

greed while ignoring users needs?" Um,

I believe that it is inferior to you for

your use case and we would like to build

models that are better for your use

case. on the whole uh we think for most

users it's a it's a better and more

capable model but we definitely have

learned things about the 40 to5 upgrade

and we will try to do much better in the

future both about better continuity and

about making sure that our model gets

better for most users not just sort of

people that are using AI for science or

coding or for whatever

uh Y asks will there ever be a version

of chatbt meant for personal connection

and reflection not only business or

education yeah for sure this is this

We think this is a wonderful use of AI.

Um, we're very touched by how much this

has meant to people's lives. We get all

of us get a ton of emails and outreach

from users about how Chacht has helped

people in difficult personal situations

or to live a better life. And like this

is what we're here for. I mean, this is

like as important as anything that we

do. Like we we love to hear about

scientific progress. We love to hear

about people that, you know, got

diagnosed with the disease and got

cured.

The personal stories are incredibly

important to us and we're thrilled about

that and we absolutely want to offer uh

such a service.

Your safety oh

uh okay two parts of this two questions

that are about tide uh from G and

anonymous. Your safety routing breaks

user trust and workflows by overriding

our choices. Will you commit to revoking

this paternalistic policy for all

consenting adult users and stop treating

us like children? Uh when do users get

control over routing? Where's the

transparency on safety and censorship?

Why can't adults pick their own models?

Yeah, I I don't think the way we handled

uh the model routing was our best thing

ever. There there are some real problems

with forro. Um and we have seen a

problem where people are forming people

that are in fragile psychiatric

situations using a model like 40 can get

into a worse one. um most adult users

can use those fine, but we do, as we've

mentioned, we have an obligation to

protect minor users. And we also um have

an obligation to protect adult users who

are not in a frame of mind where, you

know, we're reasonably likely that

they're choosing what they really want

and we're not causing them harm. as we

build age verification um in and as we

are able to differentiate users that are

having like a true um mental health

crisis from users who are not uh we of

course want to give people more user

freedom as we mentioned that's one of

our our platform principles um so yes

expect improvement there and I don't

think this was our best work and how we

communicated the previous roll out

how we strike the right balance between

protecting people and allowing adults to

speak about difficult things without

feeling policed

you want to say anything there?

Yeah. So, so,

so the definitely there is a problem

where um we we we

aim to lay out the kind of high high

level um

policies and and guidelines for the

model in in in the spec uh that we um

that we develop for for CH GBT. But um

the the the the

space of of of of of situations you can

find yourself in is is is is enormous

and and and um and at some point kind of

like establishing the the right

boundaries really becomes a a tough

intelligence problem. Um and so we are

seeing uh

improved results on this matrix from

reasoning models and from expanding uh

more more reasoning on on thinking about

this uh uh software uh software

questions and and and trade-offs. Um,

of course, this is like a bit more

difficult to train for also than uh math

problems, for example. And and so this

is something that we're uh researching

quite heavily.

Uh Kate says, "When in December will

adult mode come? Will it have more than

just NSFW? when writing even slight

conflict when writing triggers filters

on for uh I don't know exactly when in

December it will ship but yes the goal

is um when you are writing when you are

using open AAI to help chat to help you

with creative writing um you it should

be much more permissive in in many

categories than the previous uh models

are again we want this and we know users

want this too um if this is going to be

your personal tool it should help you

with what you're doing and every time

you hit a content filter for something

that

uh you know feels like it shouldn't. We

we understand how annoying that is. So

uh we are uh we're going to try to fix

that with adult mode. There may be new

problems that we face, but we want to

give people uh more flexibility.

Uh anonymous says, "Why does your idea

of safety require lying to users about

what model they're actually using?" Uh

again, I think we mis rolled this one

out, but the goal here was to let people

continue to use Forro, but in the

situations where Foro has behavior that

we think is actually really harmful

before we have all of the age gain in

that we'd like to kick it uh to to put

the user into a model where they are not

going to have some of the mental health

problems that we faced with for um FORO

was an interesting challenge. It's a

model that uh some users really love and

it was a model that was causing some

users harm that they really didn't want

and I don't think this is the last time

we'll face challenges like this with a

model. Um but we are trying to figure

out the right the right way to balance

that.

Um will we legacy models back for adults

without rewriting rewriting? Yes. Uh, y,

will the December update officially

clarify OpenAI's position on human AI

emotional bonds? Uh, or will

restrictions continue implicitly

defining such connections as harmful

worldwide?

I I don't know what it means to have an

official position like we build this

tool, you can use it the way you want.

If if you want to have like a small R

relationship and you're getting

something like empathy or friendship

that matters to you and your life out of

a model, like it's very important to us

that the model faithfully communicate

what it is and what it isn't. But if you

as the user are finding value in that

support, again, we think that's awesome.

Like we we are very touched by the

stories of people who find

value, utility, a better life in the

emotional support or other kinds of

support they get from these models.

Kylo says, "How is OpenAI increasingly

allowing so many features for the free

version users?" I can answer this from a

product and business perspective, but

Jakob, I think it might be useful for

you to just talk about the incredible

rate at which models are getting more

capable for lower prices and less

amounts of compute. Um yeah, we we are

seeing quite a lot of uh

um

ability as we get to like the new

frontiers of intelligence to to to

reduce the cost for that quite quite

quickly. Um

and so yeah, especially especially with

reasoning models, we've seen uh that

actually quite uh quite cheap models

when allowed some some additional test

time compute can become much more

capable. Um

yeah, and so this is uh this is

something that I expect will continue.

And so um yeah, as we talk kind of about

about um you know, getting to these um

new frontiers and automating research

and so forth, I I I expect that the cost

of of of a lot of that will will keep

falling quite a lot too. So yeah, our

our

we talk a lot about the increase in

model capability and for you know

pushing forward science that's that's

hugely important. One of the most

amazing things that I've observed about

AI is over the last few years, the sort

of price of a particular unit of

intelligence has fallen about 40x per

year for the last few years. Um, so when

we first,

you know, when we first had like GPT3,

we thought it was very cool and it was

at this cost that was kind of hard and

like GP3 scale models now basically run

for free like on a phone or something.

um the cost of a model that's as smart

as GPT4 at the time we launch it

relative now has fallen hugely and we

expect this trend to keep going. Now we

still think we need a ton of

infrastructure because what we continue

to find is the cheaper we can make it

the more people want to use it and I

expect that only to increase but our

goal is to drive the you know cost of

intelligence down and down and down and

have people use it for more and more

things that will allow us to continue to

offer lots of features for free. Um, but

that will also mean, I think, that

people who really want to spend a lot on

pushing AI to the limit to cure a

disease or figure out how to build a

better rocket or whatever will spend a

huge amount. Um, we are committed to

continuing to put the best technology we

can uh as long as we can make the

business model even sort of work into

the free tier and you should expect a

lot more from us there over time.

Uh, okay. Anonymous asks, "Will an age

will an age verification start that

allows users to opt out of the safety

route or a waiver that could be signed

releasing open air from any liability?"

Um, we're not going to like do again,

we're not going to do the equivalent of

selling heroin or whatever, uh, even if

you sign a liability. But yes, on the on

the principle of treat adult users like

adults, if you're age verified, you will

get quite a lot of of flexibility. We

think that's important and clearly it

resonates with the people asking these

questions. Um, anonymous also asks, "Is

chatbt the ask jeieves of AI?" We sure

hope not. We don't think it will be. Uh,

okay. Since we only have 10 minutes

left, we're going to um, and some of

these touch other things that we've uh,

we've already touched on. We're going to

skip down through some of the same

questions and try to get to more. Uh, in

future qu sessions, we can um, do more

of these if we don't get to everything

here. Um,

just as the Macintosh, no asks, just how

the Macintosh was the precursor to the

iPhone, do you see Chacht as the OpenAI

product or do you see it as a precursor

to something much greater that truly

reshapes the world?

Um,

so I would say like as a research lab,

um,

well, we haven't set out to build a a

chatbot originally, although I think

we've since come to appreciate how

aligned this this product is with with

our overall mission. Um, and we of

course expect Chad GPT to continue to

become better and and and and be this

way um for people to interact with

increasingly advanced AI. Um but we do

anticipate that uh eventually AI systems

will be capable of um

creating valuable artifacts of actually

pushing scientific progress forward as

we were discussing and um I believe that

will be the the real lasting legacy of

AI. I I think the chat interface is a

great interface. It won't be the only

interface, but the way that people use

these systems um will

change hugely over time. If you think

about what Yakob shared earlier of the

5-second, 5 minute, 5 hour tasks, um if

you think about a 5year or five century

task that would take something that

would take humans, it's hard to even

think about what that means. But

probably you want a different kind of of

product experience. I also think you

probably will want this to feel more

like a sort of ambient always present

companion. Like right now you can ask

Chacht something. It can do something

for you. But it'd be really nice to have

a service that was sort of just

observing your life and proactively

helping you when you needed it and you

know helping you come up with better

ideas and just I I think we can probably

push very hard in in that direction.

Um Neil asks, "I love GPT 4.5. It's by

far the best on the market for writing

and it's the main reason I pay for Pro.

Could we get some clarity on its future,

please?

Um, we think we're going to have models

that are much better than 4.5 very soon

uh and for writing much much better. We

plan to keep it around until we have a

model that is a huge step forward in uh

writing. But, you know, we'd like to we

don't think 4.5 is that good anymore.

We'd like to offer something much much

better.

>> But yeah, we are definitely not done

with that direction of research. Uh and

yeah, we we expect combining combining

that with with uh other things we're

working on, we'll get wheels are

dramatically better than 4.5 on all

axis. Do you have uh any sense of timing

to share about when you think we have a

model that is dramatically better than

4.5 on on this kind of task like writing

and also anything about like how far

that's going to go?

>> Um next year I think is definitely

what I expect.

Um, when is Chach Atlas for Windows

coming? asks Lars. Uh, I don't know an

exact time frame. Some number of months,

I would guess. Uh, it's it's definitely

something we want to do. And more

generally, this idea that we can build

experiences like browsers and new

devices that let you take AI with you

that get towards this sort of ambient

always helpful uh assistant rather than

something you just query in response. uh

this will be a very this will be a very

important direction for us to to push

more on.

Will you disclose the documents with the

opinions of the 170 experts so that

we'll have some transparency regarding

the new justifications for model

behavior? Um I will ask Fiji Simo uh how

she'd like to handle that, but I think

we could I don't know exactly what we'll

be able to share, but I think we should

do something there. And I I think more

um more transparency there is a good

thing. Anonymous says, "I've been a

prouser since month two. As a researcher

and fiction writer, I feel GPT helps me

think clearer. I lost the question." Um,

sorry, that was a really good question.

Let me try to find it again.

Uh, clearer but not freer. Has

imagination become an optimization

casualty?

What do you think?

Um

it is

I think it is definitely possible for

current systems that uh you know like I

think if you compare a model like 4.5 to

a model like 03 I would expect that like

there will be trade-offs there. I I

think there are definitely like

transitory as we like figure out our way

around these technologies and so again

like I expect this will this will get

better. Yeah, I

I I think there are going to be

population scale like one one of the

sort of strange things I've noticed is

people in real life talking in like

CHBTE

uh where they sort of use some of the

quirks of of things ChachiBT says and I

I think there will be other things like

this where there's like this

co-evolution of people and the

technology in ways we can't totally

predict but my expectation is over time

people are much more capable, much more

creative, think much more expansively

and much more broadly than they do

today. Um, and we start

we certainly see examples of this uh

where people are just like I never would

have been able to keep this in my head.

I never would have been able to have

this idea. And then we hear other

examples where people say, you know,

I've outsourced my thinking and I just

do what this thing tells me. And

obviously we're much more excited about

the former than the latter.

Um, can you help us understand why you

build emotionally intelligent models and

then criticize people for using using it

for accessibility reasons when it comes

to processing emotions and mental

health? Uh, again, we think that's a

good thing. We want that. Uh, we're

happy about that. There the same model

that can do that can also be used to

encourage delusions and mentally fragile

users. And what we want is people who

are using these models intentionally.

The model is not deceiving the user

about what it is and what it isn't. the

model's being helpful, the model's

helping a user accomplish their goals.

We want more of that and less of

anything that would feel like the model

tricking a user, for lack of a more

scientific word. Um, I totally get we

totally get the the frustration here

like to, you know, whenever you're

trying to stop something that is causing

harm, you stop some perfectly good use

as well. Um, but please understand the

place we're coming from here is trying

to

provide a service to adults that are

aware of it and that are getting real

value from it and not cause unintended

harm to people who who don't want that

along along the way. Um, all right.

Given that we have just a couple of

minutes left, let's see if there's any

questions in very other directions that

we should try to get to.

Okay. Uh, when do you think massive jobs

loss will happen due to AI from Razi?

So I think I think

already we are at a point where I think

a lot of the um

um the gap that stops that stops present

models from like being able to perform a

lot of um intellectual jobs. uh it's

more about um integrations and

interfaces

uh than maybe raw intellectual

capability. Um and so I think we

definitely have to

uh

think about that um uh think about

automation of of of a lot of jobs is

something that um will be happening over

the next years. And I think this is a

big uh

thing for for for for us to to

collectively think about like what are

uh

what are the jobs that will will replace

those and uh and what are the kind of

new pursuits that we'll that that we'll

all engage on.

>> This is a question from me not from uh

not from the live stream. What what do

you think meaning will look like? Uh

what do you think the jobs of the future

will look like? How do you think when AI

automates a lot of the current things

like how do you think we'll derive our

fulfillment and spend our time?

>> Um

I I expect um

yeah well I I think this is a this is a

quite um

philosophical question. I think I I

think can go in many directions but some

things I expect I think the high level

goal setting right like picking what

pursuits uh um

we're we're chasing that is something

will remain human and and

uh and I think that that is something

that a lot of people will derive meaning

from. Um, I think also just the ability

to

understand so much more about the world,

the wide the incredible variety of of of

new knowledge and new also

entertainment, but al also

uh

um just intelligence that that that that

that

will be in the world. I think I think

will uh will provide quite a lot of

meaning and fulfillment for people.

>> Okay, rapid fire, two minutes. Uh Shindi

says when GBT6

um

I think I think I think uh in some ways

maybe that's more of a question for you

uh in that like um I think I think with

with with GPD with GPD 5 uh we have uh

with previous models right like uh like

GPD4 GPD3 like we've kind of like kept

very tight uh connection of like how h

how we're training new models like what

are the products that we ship

And as I was just saying like I think

right now there's a lot to do on the

kind of integration side and on on the

uh so for example with GPT5 is the kind

of the the first time we really bring

reasoning models as kind of our our

flagship our main flagship model. Um and

so uh we're we're not coupling like

these releases and these products as

tightly to our research program anymore.

>> Yeah. Um I don't know either exactly

when we'll call it that but I think a

clear message from us is say 6 months

from now probably sooner we expect to

have huge steps forward in model

capability. Felix asks is an IPO still

planned and how would the structure then

look like? Are there rules in place for

increasing capital? Um we don't have

like specific plans or this is exactly

when it's going to happen but I think it

it's fair to say it is the most likely

path for us given the capital needs that

we'll have in sort of the size of the

company. Um, but you know that's not

like a top of- mind thing for us right

now. Alec asks, "You mentioned being

comfortable with the 1.4 trillion of

investment. What level of revenues would

you need to support this over time? What

will be the largest revenue driver? It

can't just be a per user subscription."

Um, you know, eventually we need to get

to hundreds of billions of a year in

revenue and and we're on a pretty steep

curve towards that. I expect enterprise

to be a huge revenue driver for us, but

I think consumer really will be too. And

it won't just be the subscript

subscription, but we'll have new

products, devices, tons of other things

there as well. Uh, and this says nothing

about like what it would really mean to

have AI discovery and science and all of

the revenue uh um possibilities that

that would unlock. And as we see more of

that, we will uh we will increase spend

on infrastructure. Okay. Um, we are out

of time. Thank you all very much for

joining us and the questions. and we

will try to learn from this format and

iterate on it and keep doing uh these

sorts of Q&A. Thank you very much.