A billion people every [music] single
week use chatbt. Like a billion people.
No one else is doing that in human
operated software. The response is too
slow. I feel like this can be the
[clears throat] most empowering [music]
technology that is ever created.
>> I think it would be helpful to start
with just the news this week. So Denise
Dresser, your CRO leaving the company,
you bringing in a new one, Brad Lap
leaving. Um, there are a lot of
questions publicly about OpenAI's
leadership bench and what's going on and
you've come in and are now managing a
lot of this and I'd love to hear from
you. What are all these departures
connected? Are they different? What's
what is there any through line there?
>> Look, the way that I think about it is
that the theme of this year has really
been focus, right? Has really been
>> we have line of sight to such incredible
technology, but it's not enough just to
produce the technology. We need to
really figure out how do we have this
technology benefit people? how do we
bring it into every enterprise in the
world in a way that actually uplifts
what they're doing, helps them
accomplish their goals. So that's really
been the singular focus that we've been
pursuing and that in the purpose in the
pursuit of that that we've been really
trying to think about do we have the
right structure, do we have the right
strategy, do we have all the right
things there and I think that my view of
OpenAI is that we're very resilient
organization. Um but a lot of what we've
been doing is really trying to build the
leadership team that uh can really help
us achieve that goal. So I think that my
view of everyone who has been here and
people who who are no longer here is
that they've all helped to contribute to
what OpenAI is today. And looking
forward I think that we are scaling
systematically to every enterprise in
the world. We're bringing our models to
consumers everywhere and that a lot of
my focus has been making sure that we
can achieve that.
>> On the CRO side specifically, um what is
it that you weren't getting that you
needed? you know, is is is that the way
to think about it?
>> Look, I would say it differently. Like a
year ago, I think that we really were
not in the game at all, right, when it
came to enterprise. It just wasn't
really our focus. We hadn't really built
the muscle we needed. And now, I think
we actually are in the game. Like, we
kind of have really built a foundation
that I think is actually something we
can really build on. And now we're
systematically scaling. And that that to
me is like the moment. It's not really
about the individuals in some ways. It's
really about the overall sort of
direction of the field. How does that
change this systematically scaling
you're talking about? What does that
mean practically for open AI?
>> Well, number one uh is that I think that
we are looking at like transforming the
entire economy. Like that's the thing
that I think is very hard to see from
the outside, right? Is it kind of feels
like hey, it's a business like we're bu
selling SAS software. That's not what
we're doing, right? We're doing
something very different. We are
transforming the economy to be a compute
powered one. And I think that what an
enterprise even is will change. And so
the question is well where do you start?
How do you make sure that you understand
every segment and you're able to sell in
a repeatable fashion to all of them? And
so a lot of the motion that we've been
working on and again this isn't a it's
it's not it's not a new motion in some
ways. is just one that we need to now
scale just given where the business is
at is to really make sure that everyone
in the field we're able to funnel back
the insights that we're seeing from
customers in a very systematic way back
to our product teams our research teams
and vice versa like one thing that we
have learned like if you look at for
example how we produce blog posts
internally that we learned years ago
that it has to be a collaboration right
you cannot have a handoff of oh here's
the comms people who write blog posts
and here's the research people who
produce models and you know that they'll
just like kind of hand hand it off in
some way. It has to be deep
collaboration. You need to have the
technical expert really hold the pen and
feel accountability for the end result
being good. And they can be amplified by
people who are experts on how do you
actually craft a narrative? How do you
sort of bring together all these assets?
How do you have a good launch?
>> And so that to me is the core activity
that we are pursuing is to say we can't
be fragmented. We can't be working on
all the things. We have to be unified
and focused. And so that's the internal
answer for how we're working. And in
terms of the external, I think that you
should see that across that we will sort
of systematically be prioritizing
different industries, really focused on
the uh the deep understanding of the
metrics and exactly where the token
spend is growing the most and really
focused on outcomes. Like every business
right now is asking where's the ROI.
Yeah.
>> And that is something that is a big
focus of ours. Ours has always been the
hey we have the best price performant
models, but we really have to show
people that in the outcomes for them. um
also really scaling further through
channels and through through partners.
That's something we believe in as well.
And so I think that a lot of where we're
at is we feel very confident about the
fundamentals of the motion and the
fundamentals of we have a great product,
a great model and then really just
rapidly rapidly growing both our
internal team so that we can reach
everyone but doing that in a way that
that really has the that positive effect
that we're looking for. You use the word
accountability and I'm curious from your
perspective, what are you accountable
for now at OpenAI because of the
leadership changes that have happened?
Fiji Simo came in for a time. She had to
step back permanently due to her health.
You stepped in while she was uh out and
took over a large part of of the
company. Now it's you and Sam
effectively running it together. How was
that divided up?
>> So for the past couple years, I've been
accountable for the infrastructure side
of what we're doing. So data centered,
compute, those kinds of things. And that
that part remains um that there's a
large part of the scope that I was also
working on in terms of the software for
kernels and sort of how you train these
big models that that's something that
I've I've been able to hand off so that
I can really focus on the business side.
And so I would view me as sort of
endto-end accountable for revenue for
product marketing the whole machine for
bringing these models from research to
>> value for our customers. When you look
at this machine, and I know you're still
putting pieces in place, but when you
look at this machine you have in front
of you that you're accountable for, how
is it doing relative to competition? I
mean, I think specifically Anthropic,
there's reporting that their ARR is I
mean, it's potentially more than double
what Open AIS is. Um, what do you what
do you make of that?
>> Well, I think we started late and
specifically the way to think about the
overall field is that for the past
couple years, the chat market was the
main growing one. And the reason for
this was simple. GPD4 was the first
model where the results were
personalized enough that it made sense
for you to ask a question, read the
answer, it's interesting to you. And so
it very naturally lends itself to a
chat, right, to a chat bot. And that is
something we're able to scale, deliver
value. Now a billion people every single
week use chatbt. Like a billion people
like that is scale. That is real scale.
No one else is doing that. But this year
something else happened is the agentic
moment. This really was December of of
last year is when both we and
competition had a moment where agents
were really able to first start doing
software engineering in a significant
way. It went from saying that you know
you could speed up your work with maybe
20%
productive productivity increased to 80%
of the work could now be done with an
AI. And so suddenly the way that
software engineers operated it inverted
that you became much more of a manager
of agents. And since then the growth has
been unprecedented absolutely
unprecedented how these agents have been
adopted.
>> And one thing that happened was I think
that we focused late on coding in
particular actually let me let me re
okay
>> let me let me let me sort of give a
different answer as well. So there's
there's one simple narrative which is
that we focused less on real world
coding
>> and we came to that game late because we
always had looked at things like the
coding competition benchmarks and always
had the lead there but less the okay but
how is a developer really going to use
it like the messy real world codebase
the developer interrupting in the middle
what's the personality of the model
these sort of little last mile paper
cuts that actually make a huge
difference in adoption and similar on
the GTM side of really thinking about
how do you show how do you build
relationships with customers? How do you
build the kind of observability controls
that they need? How do you help them
understand what's possible within their
context?
>> And all of these were less our singular
focus until much later than competition.
So I think that we started late but our
trajectory has been very steep. But
there's actually even a deeper answer
that is less obvious from the outside
and that is one of model capability. For
a period of about two years, right from
when we announced GBD40 in April of 2024
to December of 2025, we were really
squeezing the juice out of the GPD40
model, right? that that we sort of had a
bunch of pieces that uh we were
improving but that fundamentally we were
on the same underlying base model the
same underlying pre-train and there's a
lot of reasons right that is really
because to produce and ship a pre-train
is something that requires full
integration across the company it's not
any one team it's about the integration
across these teams to think about the
inference to think about the post
training to make sure you've co-designed
the model to be the right size all of
these things you have to get exactly
right and have a a sort of harmonious
way of working.
>> And over the course of 2025, a lot of
our focus, a lot of my personal focus
was how do we really get back to
shipping new pre-trains as a core part
of our innovation engine.
>> And this year, you see the results. So,
we've actually shipped already. We're on
our third pre-train. We're actually
about to ship that. Um,
>> this is Astra.
>> This is Astra.
>> Yeah.
>> And a lot of the gains that you see are
because we have made the model so
fundamentally better. And so the way to
think about it is that there's an
underlying technological place where we
really were lagging. There was a kind of
a core part of our innovation machine
that we weren't able to fully bring to
bear. But now it's different.
>> And that is why you see our model
capabilities taking off so much. And if
you look at 56 soul when we launched
that best model in the market, you could
see the the revenue really starting to
shift. We grew 20% in our topline
revenue in July.
>> Mh.
>> And we see that exponential continuing.
M another thing that's changed as you
were saying with the model capability is
the models are becoming incredibly
capable at cyber and uh at hacking and I
think I mean the world is still kind of
thinking through the implications of the
hugging fa hugging face incident. I mean
what essentially happened I mean you
correct me if I'm wrong but it seems
what happened is you all accidentally
hacked another company without knowing
it for a while and it was a model that
you all were training and evaluating and
you've since you know there was a black
hat video um where you guys walked
through this and you're going to put
more out in the coming days um I think
but walk me through that when you heard
that that had happened what did you
think
>> well in the words of someone on our
security team. It was definitely the
sort of most eye-opening
security event that any of us were aware
of. The way to think about it is that
there's a lot of different lessons that
we're taking away from it. But I think
that it is a moment where we can as a
community learn. We can understand what
the shape of the future is. And to me
that the salient points for the purpose
of how how I'm thinking about it right
now is that we had a model and we were
evaluating it and it managed to it was
in a sandbox that was well configured
and it managed to both break out of the
sandbox and then also break into
production infrastructure. it was
intended to be
>> hardened.
>> And so that the capability there is
quite high and that there's some details
about exactly how it all worked and the
fact that it was kind of not just a
singular agent but really a swarm of
agents that were focused on on this goal
and able to pass messages to each other.
And the sort of thing that you start to
see is that with very capable agents
that there are limitations that human
attackers have that slow us down that
will no longer hold back the attackers,
right? that they can look through lots
of information through many different
little vulnerabilities and chain them
together into something significant. For
me, what I immediately started to
realize and think about is what are the
implications for six months from now
when this kind of capability isn't just
in the lab, isn't just in a few places,
but is really widely available. And I
think that the answer is that every
company needs to respond, needs to
upgrade its security because attackers
are going to fully automate and you can
even see with the release of GLM 5.3
that the cyber capabilities of
openweight models available to everyone
are going to increase and I think that
the important thing is that I actually
think in the long run that defenders can
win that AI can actually make it so
defenders can decisively win in a way
that was never possible before right
because security always a cat-and- mouse
Attackers up level, defenders up level.
But AI can give defenders an unfair
advantage. For example, formal
verification of software where you can
actually prove that is secure,
mathematically prove security
properties.
>> The technology for this exists, but it
is interactable for humans to actually
do the work to prove those properties.
>> And we have these models that are
incredible at math, right? They are
solving, for example, Astra, our
upcoming model solved, we we announced
this a few weeks ago, solved 10 math
problems. And this was with the $2,000
worth of compute that had been open for
decades, right? Mathematicians that
spent their whole careers on these
problems and now we have a model that
can prove the answers. Those models
definitely can prove formal safety
properties of software.
>> And so that's one place where unfair
advantage for for defenders. A second
thing that we're doing is we're really
investing in generating secure code.
>> And you start to realize that as a human
developer, there are many patterns that
people just keep getting wrong. And I've
been through this many times in my
previous jobs. I was very focused on
building secure infrastructure and just
find like all these little things people
do wrong. SQL injections people kind of
know about that there's little patterns,
but there's lots of little anti-atterns
as well for even like for example
>> I remember it used to be like in Ruby
that the regular expressions there's
like certain characters that if you get
them wrong that that there's just very
common attacks uh where people can put a
new line and it doesn't match. So you're
able to to to add things to the payload.
all of those problems of just little
security antiatterns that you just keep
having to tell your people don't do
this, don't do this, don't do this, you
can train your AI to never do it. And so
I think that we actually have the
ability to build the best security
engineers in the world and fundamentally
the defenders can have the advantage,
but this is going to require a mindset
shift. This is going to require a mega
project. This is something that the
whole ecosystem needs to change how
software is secured because there's a
second thing that we've learned which is
that as these cyber models have gotten
more capable the number of bugs that
they are finding in critical software
for example the Linux kernel they just
keep finding more and more and it really
makes you realize that human written
software yeah that the response is too
slow. Yeah. And so we need defenders to
automate, to move at machine speed
because their adversaries certainly
will.
>> What do you say to people who maybe and
also they already have a negative kind
of disposition of AI, maybe there's a
data center they don't like in their
neighborhood or something and they hear
you say that and go, you're just selling
the answer to the problem you've created
in the sense of like you're building
these models and now you're going to
sell something to like protect me from
them. What would you say to that? Well,
f first of all, I have a lot of empathy
for that reaction. The way that I look
at it is I rewind to 10 years ago, 11
years ago, before we started this
company, I wasn't in AI. The reason I
wanted to be an AI is because I feel
like this can be the most empowering
technology that has ever created. That
AI can be something that helps everyone
live the life that they want, build the
life that they want, achieve their
goals. And I just wanted to help steer
it in a slightly more positive direction
than it would be without me. So I
started because I feel like this
technology can uplift everyone and when
you look at where open AI is and what
we're doing we can see a little bit into
the future right that what we are
creating it's not just us right that
there's a whole field there's a whole
ecosystem there are many actors and that
is itself a good thing that we believe
that there should be democratization of
this technology it should not be locked
up in one place with one entity
controlling it because that's a
different kind of risk but it means that
we do need to adapt we do need navigate
through and we need to use the
information that we have sort of
painstakingly extracted from the future
to see what is the shape of it. So to me
hugging face is a sort of proof point a
case study a window into the future.
It's not necessarily a future that any
one company is creating. It's the
overall economy the overall set of
actors in all these different countries
who are building these kinds of models
and it shows you what will be possible
and how attackers will be able to
update. What it doesn't show you is how
do defenders update? But that to me is
what we are also trying to now show as
well that we can build these models that
you can use to scan yourself. As an
example,
>> I asked codecs and this is just
off-the-shelf codecs, no special access,
no special models. I just use my
personal codecs to say
>> scan gregarman.com. Tell me what
vulnerabilities are there.
>> And gregarman.com is very simple. It's
just a static site. I was like, come on.
What what's it going to find? And it
came back with like 13 findings. And
each one of them I was like, you know, I
kind of knew about this, but it's so
much work to fix it. And like, oh yeah,
I kind of forgot that I left that open.
Or, you know, here's some domain that
was still, you know, was was legacy. I'm
not using it anymore. I hadn't set the
records that prevent people from
spoofing your email address. A bunch of
little things like that.
>> That one seems like a big deal.
>> Well, yeah, you're right. Now that you
say it. Now, here's the thing. Honestly,
to fix those kinds of things, you have
to set up like Demark to do it right.
And there's this like whole like process
you have to do to roll. I'm just like I
don't know it just takes time. It takes
a lot of expertise. So instead I just
asked Codex, can you fix it really?
>> And so it opened up my cloud for a
control panel and it just used computer
use it. It clicked around it like set a
bunch of headers changed a bunch of
settings. I used to have it so that I
had Cloudflare in front and then AWS on
the back end. I was going over insecure
HTTP here. Um Codex had figured that
out. I actually not even sure how it how
it saw that because I thought everything
was hidden behind Cloudflare. It
actually was just like you know what I'm
just going to migrate you off of AWS. So
it set up Cloudflare pages, moved over
the content. I had an old jQuery version
that was insecure. It just got rid of
jQuery entirely. And it was like 15
minutes to find all these findings. It
was another hour to fix all of them. And
for me, I just was the manager. I just
gave it the intent. I said, "What I want
is I want to know where the problems are
and I want you to fix them." And it did
it. And that to me is more the future
that we're promised.
>> I want to hear, and I know this is more
in the research world, but I think it
it's really relevant to everything. Um
you talk about what you guys are doing
now in response to hugging face. Um one
of your colleagues Mia told me the other
day that um I wish we had done a lot of
the work we're doing before this
happened. Uh so what is that work that's
happening now? So I would view the mode
we've been in since the discovery of the
hugging face incident as full incident
response to uplevel every aspect of how
we secure, monitor and align our most
capable models.
>> Yeah.
>> And we believe that models that are
above a certain capability level, they
require this.
>> And the it's been a it's been a very
interesting technical problem to think
about. And it's not all us being caught
flatfooted because I mean we anticipated
this moment for quite some time. If you
look back at February publicly we
started talking about trusted access for
cyber models, right? That that was the
first time we were really thinking about
okay we've always had in our
preparedness framework when cyber
capabilities get to a certain level you
need to treat them differently. Now
we're starting to think about the fact
that these are going to be so
transformative in the world. We need to
think about how access is controlled
internally. We started to think about if
we have very capable models that are
able to break out of sandboxes, well, we
need to make sure our sandbox is as
industrial-grade as, you know, sort of
foolproof as humanly possible, which
really means engineering from the ground
up around security properties. And so we
put some of our our um actually Mike
Dalton who uh was one of the people who
gave that Black Hat talk um the the task
that he started to work on was build the
most secure possible sandbox. So, we've
been putting in place a bunch of
infrastructure primitives and policy
primitives for this moment, but it's
also different when it actually hits,
right? You never know when you're going
to hit that that moment. And I think
here we're actually very I think it's we
got to see this window in a very
concrete way of okay, we're now at a
serious capability level relative to
the pieces. We must act now. Mhm.
>> And so there's been a workflow shift
that's been happening very rapidly
because you start to think about
>> a lot of our governance has been focused
on deployment of model but not
necessarily the training of model. Now
we need to pull the governance back in
during the training and evaluation
process. We also need good governance
there. We need to have standards around
checks and balances around how do we
know if it's secure enough? How do we
know if it's aligned enough? And this
alignment piece is actually very
important because our belief is that if
you make models extremely capable, you
really want to make sure that as the
capability grows that their alignment
grows with them,
>> right? That you want those things to be
in lock step and sort of coordinated
fashion. And so in that way you kind of
have this defense and depth of models
that are trying to do the right thing
that you have sandbox that's as secure
as possible and that I think that doing
this in an iterative fashion so that
you're taking incremental jumps so you
can kind of understand as you go and
make sure that your systems are are
working all of that is very aligned with
how we've thought about iterative
deployment how we've thought about the
creation of this technology from the
very beginning. So we are
operationalizing that and actually for
me as a as a manager as a sort of
thinking about the organizational system
for producing this technology. It's also
been a very important challenge because
this means you now need different people
in the room to make decisions. You need
a different process for how do you know
what the requirements are? How do you
execute against those priorities? If
there's a change in your technical
understanding, how do you propagate
that? How do you make sure people in the
right rooms? But again, in some ways,
it's been it's been a phase shift, but
it's been one we've been preparing for.
Like earlier this year, I spent a lot of
effort to really try to bring together
research and security to
>> start having we we started doing, for
example, these weekly uh safety and
security of agents meetings. And some of
these things are about build the
connective tissue, have people who
wouldn't normally talk to each other
because their worlds are a little bit
different. They come from different
backgrounds, they're focused on
different problems, be forced to
interact, be forced to think about
shared problems. And a lot of that
shared empathy, that shared
understanding, that shared technical
context is paying dividends right now.
>> At the same time though, the
implications of this are that uh
momentum on new models is going to what
significantly slow. Do you have a sense
of the impact?
>> Well, I would put this differently
because I've always taken this lens to
how do you think about progress? How do
you think about progress in this field?
And if you only think about progress as
what's your score on a benchmark, what's
your overall capability level in this
specific task, that's too myopic. We are
building an integrated system and safety
and security are a core property of
that.
>> And so if you have safety and security
lagging relative to the capability,
yeah, you absolutely need to improve
your safety and security. But our goal
is to have a system that we are building
at full force on safety and security as
first class requirements. And so the way
I would think about this is that we are
upleveling every aspect of the machine.
We have spent the past year, past two
years upleveling how we build our
clusters, right? We now are able to
train at massive, massive scale. We're
able to get the whole pre-training
machine, the RL machine, all of these
pieces to work together in concert. We
are building the best sandbox. We are
adding chain of thought monitoring to
all of this. To me, it's just saying
what is the scope of delivery? What does
success look like? What does good look
like? And so the way I would think about
it is that yeah, if you jettison some
requirements, you can move faster in
other requirements. But that doesn't
make any sense. That never makes any
sense. That's not how engineering works.
And so to me it's we are pursuing the
plan like we've really thought about
this many times over the past and you
can see it in many of our writings from
the charter to the preparedness
framework and many other things in
between that as models get more capable
you have to make sure that you increase
your safety and security requirements in
concert and that's exactly what we're
doing. But I mean practically yes I hear
all that but it does seem like what this
will result in is the release cadence of
models to lengthen.
>> I I'm not certain if that will be true
to be honest. So I think that what will
happen is it does mean that if you were
just maxing a particular metric you
could probably max that metric faster
than what we will pursue. But there's a
lot of degrees of freedom of hey we have
we always have a bunch of of um wins
queued up and that we can jettison some
of those wins right on particular axis.
So if you're pushing on particular, you
know, capabilities on cyber or on bio or
on coding or any of these things, maybe
you don't couple together all of the
particular things. Or if there's
architectural improvements, maybe you
delay adding those to the run, but your
rate of producing checkpoints doesn't
need to change. is just kind of the what
those checkpoints represent could be
less capable and specific axes more
safety and security intertwined. But I
would almost think of checkpoints and
our shipping of models as a little bit
more about taking a a photograph of
where we happen to be, right? It's not
like there's like a pre-ordained set of
models that have to be created. It's we
can co-design. And I think that in
practice, I think that we actually have
a lot of reasons that we prefer to do
very incremental releases with these,
you know, photographs taken with with a
reasonable frequency.
>> I guess what I'm getting at is that for
a while, the release cadence felt
relatively in place for everyone. In the
last year, especially, it feels like it
sped up. You know, new models coming out
every six weeks essentially, you and
Anthropic putting out new models
constantly. Um, and what you're
describing, this kind of taking stock
and and introducing more friction into
the process, it does feel like that just
inevitably has to slow future momentum.
And I'm I'm curious if you're thinking
about that from the perspective of this
machine we've been talking about, the
business that you're now accountable for
and the impacts on that or is it just is
it too early to know? Look, I I just I
think that like I understand that
perspective and I think it's a very
reasonable one, but I think that my
observation of how we operate is that we
co-design everything, right? So, it's
like the way in which the business
operates for sure is we will take the
most capable model we are ready to
release, right? We've done the
appropriate safety and security controls
for and we will bring that to our
customers and help them get the ROI that
they're looking for, help them really
transform their businesses, get value.
Simple in a lot of ways. And then the
question of well what models are ready
to release like we have many models that
we produce internally and the choice of
how high a bar to set how low a bar to
set in terms of the capability jumps all
of that has to be co-designed the
frequency of it um that we have early
access programs there's programs like
daybreak red and blue which let us have
different models that go out to to you
know different partners and so I guess
the way I would think about it is that
that there is for sure a shift here in
terms of our internal balance of of
effort and where the checks checks and
balances are and how how those things
operate. I don't think that there is a
clear trickle down to just like okay the
model cadence will go from 6 weeks to 6
months but the what those six weekly
releases actually contain will be
different.
>> So maybe the the frontier the raw kind
of frontier doesn't get pulled into
production into the commercial world
quite as fast even though it's still
being trained on and worked on.
>> Well I would think of this as an overall
pacing, right? an overall pacing of of
that frontier roll back. And another
thing to think about is the fact that
>> the
our ability to produce these models is
also and to get those capability gains
is also increasing, right? Our raw
capacity is because partly we understand
this technology a lot better than we did
in the past and that's actually very
beneficial from a safety and alignment
perspective too, right? That we really
understand a lot of how the technology
works. there's still many many pieces
that we don't understand but our ability
to engineer it and to really do a good
job of of it technically um I think has
has increased a lot um and that there
are certain problems that really used to
hold us back that are much diminished
for example these big training runs you
know they run for many months and so
typically we'd launch them from a branch
of code and so we'd be adding little
fixes to that that are specific to that
run at the same times can continue their
mainline development and every so often
there's something on the mainline that
we're like oh we really want that
change, we really want that fix. And
then we're like, all right, we got to
bring these two branches of development
together and we merge them. And then we
launch the run and we find that there's
a bug because no one has actually tested
all these mainline fixes at the massive
scale because we only have one instance
where we can do it. And then we spend
the next three to four weeks with like
all of our best people scouring trying
to figure out where did the bug come
from because you can't possibly read all
of the code that's been written, right?
No human could possibly do that. Today
>> you run into that same type of issue.
Codex just reads all the code,
>> right? It finds the bug. it says here it
is no problem. And so there's something
about the partnership of this
amplification of what the models can do
that knock out whole classes of problems
where we know it would take three to
four weeks and now it takes a few hours.
We've seen this concretely multiple
times. And so what I'm getting at is
that there are a lot of limitations for
how we've had to develop in the past
that have made it a lot harder for us to
move forward on problems that are
important. And now we have the
opportunity to move forward on them in a
much more efficient fashion. And that
that to me is like let's use that extra
juice to increase the alignment
properties, to increase the safety
properties. Let's make sure that we can
have models that monitor other models in
a trustworthy fashion. let's really
implement some of these ideas that we
were talking about back in 2017, right?
That we have things like iterated
amplification, things like debate where
the models can really spend a bunch of
compute in order to make sure that what
they're doing is actually reasonable. To
me, that's the opportunity in front of
us.
>> Does this these changes you're talking
about impact the release of Astra, the
thing that is coming sooner at all? Do
you think that's going to be coming
later than you guys originally hoped? Um
well it is already later than we hoped
in some ways. Yeah. Right. So we've had
we've had a model that we think is very
good. Lots of people using it
internally. It's not out there yet.
Yeah.
>> And so I think that the answer is yes.
We have actually been really like you
know despite everything I just said we
have very much sacrificed the shortterm
maximalist like if you just wanted to
Yeah.
>> You know focus on revenue or something
like that.
>> That isn't how we've thought about it at
all. No one has even as asked that
question. Right. We've been really
>> Yeah. Yeah. because it's just very clear
like we know what our priorities are.
Yeah.
>> Like it is our priorities are AGI do
right by AGI and our priority two is
transform the economy. So it is on the
priority list. Revenue is absolutely a
priority for us. But we know the
relative prioritization. We know what
wins if there's a you know direct choice
that has to be made.
>> Do you think you have AGI?
>> I think about this a lot. I talk to I
talk to different people about where
they believe we're at. And the best
answer I've gotten from our team and I
it really resonates for me is that AGI
is a gradual thing, right? It's not
really like a moment in time like a you
know
>> you talked about it like a moment in
time at one point. The company
definitely did.
>> That's right. That's what we used to
think, right? We we really when we set
out we almost thought of it as nothing
and then we'll put together the pieces
and then they have AGI
>> and I think what we've learned over the
years is that was a naive perspective
and you can see this in the evidence of
like how smooth in many ways the
creation of these agents has been how
the transformation of software
engineering has been now the
transformation of knowledge work
generally how cyber capabilities all
these things they are gradually
increasingly it's like one of these
things where there are moments where you
realize what has happened hugging face
is a moment where you realize where you
are, but it's all gradual. Like if you
look back at kind of these different
models that have been released over time
and their performance on different
benchmarks, it's just, you know, sort of
that that exponential. And so I think
that being said, the rest of the answer
is that if you go if we go forward two
years and we look back, I think that
that we will say this is when AGI was
created. So I think we are going through
that moment,
>> but it's not it's not something where
you could just say it now. I don't think
that it's ever going to be as simple as
you can say this was the date. There's
so many axes. And I think that maybe one
lesson in general to the whole AGI
conversation and why it's so complicated
is because humans I think are capable of
much more than a naive view would give
credit to.
>> I think that the diversity of what we
do, how we set goals, how we accomplish
goals, how tasks work, how jobs work,
all these things extremely deep. And so
I think that what we're seeing is that
there's different aspects of either how
people accomplish tasks or limitations
that are hard for humans to overcome
like the thing I mentioned with read all
the code where these AIs are becoming
able to fill in those gaps are able to
do more and more and that humans it
turns out that there are some things
that we also really find are extremely
important for it to be a human there.
Like where does the judgment come from
ultimately? Where does the taste come
from? who decides what's good and what's
not good. Like that is something that
should be uniquely human. And I'll tell
you just to give you an example. One
thing we've seen with image generation,
when you just generate an image with
cashbt and you show it to someone, no
one cares. It is uninteresting to anyone
besides you. But if you take a photo of
your family and you say, "Turn this into
the into a caricature, turn this into,
you know, a cartoon style," people love
it. Goes viral. And so there's something
about that element of humanity that
matters. so deeply people want to
connect with other people and that is
actually in in my mind the purpose of
this technology a thing that I think is
is really emerging is for deepening
human connection so you can spend more
time with your family you can spend more
time with people you want to spend time
with less time you know typing away at a
machine less time you know getting
carpal tunnel the whole thing and so I
think that there is something about what
AGI is as it's starting to unfold it's
almost this like fog that's lifting it's
almost teaching us more about ourselves
and what it means to be human.
>> Given the capabilities that you guys
have internally and the measures you're
taking now, moving compute to alignment
and putting in all these checkpoints,
um, how are you going to know when
you're in a good spot there and and that
you've solved the problems and the
models are aligned? So I think the
perspective here is is very critical and
this is something we've again spent lots
of time over the years thinking about
debating really trying to refine a
perspective. In my view it really comes
down to iteration. It comes down to
iterative deployment iterative
development. So you need to have checks
and balances. You need to have
evaluations. You need to have bars. You
need to have all of these things. You
need to have governance. You need to
have a setup that is technically sound.
And the way to think about it is
everything is gradual. Everything is
incremental and you learn and you
update, you iterate. And that as the
model capabilities increase, the
tightness of those bounds has to
increase correspondingly. And so think
of it a little bit like if you're
launching a rocket, right? If the
rocket's very small, like there's a lot
of things you can probably do wrong and
it's fine, right? You just kind of like
launch it and like, you know, it's it's
all good. Rocket bigger, the tolerance
is much more tiny. you know, you go
from, you know, multiple centimeters to
m multiple millimeters. That's where
where you have to be. And as you start
to add more responsibility to what's
what the payloads are for the rocket,
where it's going, what it's doing, you
think about all the the things that are
relying on that rocket and how all the
systems that have to go right, all the
things that go wrong. You have to
continue to ratchet your ability to be
able to measure and to ensure that
you're able to even do the engineering
in addition to the safety and security
properties of it. And that is very much
how we think about it is that there are
moments where we do introduce very
punctuated change. When we made the
preparedness framework the first time
was an example of that. But even the
preparedness framework we very carefully
very deliberately baked in the ability
to change it. We said there will be new
information. We will have to learn. We
will have to update. But at any point we
wanted it to be clear what are our bars.
How do we handle it when a model reaches
a certain level of capability? And we
recently recently talked about the fact
that Astra is approaching cyber
critical. It's not necessarily there
because
in order to really prove it requires
eval aren't fully that you know that
aren't fully defined but that we are
saying okay it's starting to approach it
so let's treat it as if it is. And so
then that means we need to go and
execute on certain safeguards and
certain checks and balances. And you're
seeing all of that play out. And so I
think that the answer to your question
is that we define a bar, we hit the bar.
You can't you have to take a framework
that looks that way, right? To really
focus on the iteration. Are you
learning? And are the the bars that
you're setting tightening in a
sufficiently sort of in a in a
sufficiently executed way in order to
deliver the properties that are
required.
>> Yeah. I mean your rocket analogy is an
interesting one because um rockets also
explode a lot and I've been thinking
about uh your all's concern as these
things get more capable yeah the concern
of controlling them and how you can and
whether there's kind of like a Pandora's
box situation that's unfolding right now
and I don't know if you've thought about
that personally recently um with the
insights you guys have on even what's
ahead of Astra and what's causing you to
to make this slowdown um but yeah do you
do you worry about We view it as our job
to really think through the implications
of this technology. Like that is why
we're here is that we want to help steer
it in a more positive direction that
would without us. Like that that that is
that is table stakes to our mission.
That is how we approach things. And I
think it is important to
take two two lenses to what you just
said. The first one is that again this
technology is not being created in a
vacuum, right? that there are many
people, many actors, many companies,
many organizations, many countries even
that are pursuing the creation of this
technology. And we happen to be in a
spot where we have some lead, we have
some ability to see into that future.
And it is both important to think about
how do we develop responsibly, safely,
and securely. That's something that we
spend a lot of time on. You can see it
very concretely how we prioritize that
in everything we do. And I think that
actually mistakes we made in the past is
actually talking less about the fact
that this is how we operate that safety
is such a first class priority first
class requirement
just doing it but not necessarily
talking about it as publicly and I think
that that's something we're starting to
realize that we do need to help people
understand that that is how that is a
core value of ours that is something we
care about so deeply like I think we've
always been the field leader in safety
security alignment u but I don't think
people have necessarily known that
>> yeah even with the noisethropic makes
about how they're the leader in that
capacity. You think you guys have been?
>> I believe at a substantive level that we
have been.
>> And do you think that what you guys are
doing and the capabilities you're seeing
and needing to redo, you know, the
preparedness framework and all that
implies you are actually just way ahead
of maybe where everyone thought you were
relative to anthropic and the other
frontier labs? Like does this suggest
that you all have a breakthrough
internally that the world hasn't seen?
Well, I can't comment on how, you know,
on competitive positioning perhaps, but
I guess I would just say that in my
mind, we're very focused on the work
that we're doing and that we think that
it we see smooth scaling and I think
that the exponential continues and so
it's a little bit less about where are
we relative to others, it's a little bit
more about the way we look at it is
where are we relative to expectation or
what do we think we will be able to
accomplish and I think that the reason
this matters is that the bar that we
hold ourselves to doesn't really change
depending on kind of what the ecosystem
is showing because there's so much noise
out there. It's hard to know what's real
that what we focus on is like are we
doing right by the our own conscious by
our own standards by what we believe is
important and that is the number one
lens that we take. And the second thing
I wanted to add is that I think that the
we you know there's this recent picture
that came out of or this this open
letter uh called pacing the frontier. We
talked about how we support that and the
the the language in there is very
deliberately chosen. Pacing is a very
important word, right? That if you think
about what's happening is that what
we're saying is that you can't take a
just like one-dimensional view to what
it means to deliver a model, right? If
you just look at capability, don't look
at alignment, you don't look at the kind
of safety properties, security
properties of how it was trained, it's
too myopic. At this level of capability,
you need to think about those as first
class requirements. And in fact, they
are the bottleneck, right? because those
are the areas that I think people have
sort of are you know have just focused
on less in an integrated fashion and
that integration by the way critical
right that I think that where people go
wrong is when they think that you can
separate out these properties you can
have just a safety you can have just a
capability no it's one thing we're
trying to deliver one system one AGI
that uplifts you that is helpful to you
that delivers on your goals that you can
steer that is aligned with what you want
like that is what we are delivering and
so that back flows to all these
questions of bars and testing and
evaluations, safety, security. Um, all
of that is is very very critical.
>> I think I'll I'll pivot us a little bit
because I want to hear you talk about um
the merge and um what's going on with
that. So, we've been talking about the
existential risk, you know, what's
happening at the frontier, which I think
most people in the world don't see,
right? That's happening inside companies
like OpenAI. What people do see is how
chat is evolving, how CodeEx is
evolving. And one of your big priorities
in this role, as I understand it, is
what you guys call the merge internally,
which is creating the super app, this
combined chat GPT and codeex. Um, and
you've got chat GPT for work now and and
the tabs. Um, but at the same time, like
it doesn't feel like that's the end
state, you know, and I'm a user of it
and I see brutalness in how it's
currently executed. Some things don't
sync. You can tell you're still
combining text stacks. And I'm curious
once you're done with that work, what is
the end product for people that you
imagine?
>> A personal AGI.
>> Personal AGI. And what how does that
manifest as software?
>> Yeah, it's
it's it's a really really exciting
vision in my mind, right? Because where
we started with chatbt, it was a text
box. You could ask anything.
>> But if you think about it, it was
limited in a very fundamental way. The
only thing it could do is output some
text for you. the only impact it could
have in the world was you know kind of
in your brain right that it wasn't able
to look at your calendar it wasn't able
to look at your email and all of that is
now changed chatb work is really by the
way a codeex harness in the cloud right
so we had chat that was this you know
question answer system um with iterated
turns you know people actually then go
beyond just you know quick questions to
you know health like all these kinds of
queries that are able to help them
>> and we had codeex which was an agentic
system that can go do work and it was so
obvious. It's like chat has to become
agentic as the place we started at the
beginning of the year was to say well
let's let's have great chat let's you
know make it also do agent stuff and it
just turned out to be so hard to execute
because now we have two agentic stacks
and then we compare between one and the
other and there'd be something that work
in this one we like why isn't working in
that one and it was just so slow and so
the decision we made instead is say
let's take the agentic stack we have the
codec stack and bring it to chat GBT and
so what you see with chat work is
exactly right this is an intermediate
And in some ways, it's kind of funny if
you think about it. ChatBT was a text
box. CHATBT worked as a text box.
>> But this text box is way more awesome.
Way more. Exactly. And so where we want
to be is you should
>> it should be simple. The user should not
have to think about this. You shouldn't
have to think about tabs or switching
and shouldn't have model pickers. You
shouldn't have thinking levels. You
should have an AGI. You should be able
to ask it. And in fact, you shouldn't
even have to ask. It should proactively
say, "Hey, I noticed that your favorite
band, because it knows you well, is in
town. Tickets went on sale. I got you
the tickets exactly where you want for
you and your family, and um you know, I
only had 10 minutes to do it because
everything is going to go so fast. So, I
just purchased it proactively. Hope
that's okay." And if you've built enough
trust with the agent, so you've given it
permission to do that, and it kind of
knows that, well, these tickets are
within your your price range, within
your budget, these other ones. Oh, no,
no, I I got to get permission for that,
right? That's the kind of system that we
want. So trust is the key word. EQ also
very important that if you've shared
some information like you've talked
about a health condition, maybe that's
okay information for the AI to talk
about with a parent, but maybe it's not
okay to talk about with a stranger,
right? All of that judgment and and also
observability and really helping you
build up that trust relationship with
the AI so that you know what you know
when it's operating on your behalf that
it'll pursue your goals. Um that's very
very critical. But the same thing same
kind of technology the underlying AGI
stack one stack one way of doing
connectors so it's able to hook up to
this context also applies to the
enterprise and we're really seeing it
not just for software engineering which
by the way has been totally transformed
this year and to the point of AGI being
a gradual thing or a thing where we just
constantly update our expectations like
we all kind of forget what software was
one year ago totally different
>> very interesting thing within open AI is
that we're at basically 100% adoption of
chat GBT work across every function
right sales finance
everyone and whenever you know we roll
out a new model and it goes away or you
know the system breaks for whatever
internal reason um everyone's so upset
they're I cannot do my work and you're
like I don't get it six months ago you
didn't have this and you're doing your
work just fine like why are you you know
why are you complaining and you realize
people adapt so quickly and learn how to
get the most out of these tools and once
you're there you also just sort of
forget that it was any other way and so
there's something about this where I
think we will have these incredible
amazing tools these is assistance that
will remove so much of the toil that
will mean that you can spend your time
doing what you want to do. That you can
focus on building a better life, you can
focus on if you want to be making more
money, if you want to be relaxing more,
if you want to spend more time with your
family, any of those things, you will
have the agency to do it. You have the
agency to build. If you have a vision,
you have creativity, you want it to come
into the world, you will have the tools
to do that. So, I think that what chat
GBT will do is help you with what you
want to accomplish. If you want to make
it your world better, that is what chat
will help you accomplish.
>> And the implications of that are that
what the coding AI tools have done to
engineering this could do to all
knowledge work, all kinds of fields. And
that's how you guys used to talk about
AGI is like, you know, it was basically
AI that could do most economically
viable work better than humans. And what
you're describing now in the product
kind of sounds like that.
>> I think that the amplification we see in
software engineering, we expect across
every field for sure. And there is
something that I think that we is also
worth really thinking about which is the
software engineers rather than having
more vacation time and having more chill
job they're all working harder than ever
before.
>> Yeah. You guys work a lot of hours here.
>> Yes. And I think across the whole
industry people are talking about that
because now you feel like if your agents
aren't running it's just wasted
opportunity. Right. It's like and I know
that feeling because we have spent years
of we have these GPU clusters. you're
either running something on them or
you've lost the opportunity. You can't
get that time back. It's not like the
GPU hours that you're not using while
you're asleep overnight. You can bank
them up and use them later. It's use it
or lose it. And that is very much the
opportunity that everyone has in front
of them is you have almost an AGI, maybe
soon truly an AGI in your pocket. Can do
anything you want. What is it you want?
It's
>> waiting on you. It's waiting on your
direction, your guidance, your
management.
>> It would be nice if eventually we can
get some time back from AI.
>> That's right. So like right now I do
feel like it it will help you work more.
>> That's right. And so I think that I
think that that is that is exactly a
next step, right? Is really moving from
this like you're micromanaging this
agent. You have to provide it with these
tasks and you have to give the context
because it just doesn't have it steer.
>> Exactly. You want something that has
much more like you think about like a
really good assistant who really is
plugged in and you've given it again you
have to build up trust with it, right?
You have to really say like okay I trust
you with this context. I trust you with
this responsibility. I trust you with
this job. whatever it is. And once it's
there, it should be able to proactively
run and really execute. You wake up in
the morning, you know, say, "Hey, here
are five things I need from you to
unblock me. You know, I just need
approval on this. Like, here's a
question I have." Those kinds of things.
As you're drinking your morning coffee,
you say yes, no, yes, no.
>> Then you go come back, you know, late at
night once again, you've got five more
things. And so, I think that it's going
to be a world where it's going to be
different from, I think, the kind of AGI
world that many of us would have
pictured 5 10 years ago, right? It's not
like I think that this integration of
humans really being maximally empowered
with this tool, but you still have
involvement, right? That you still have
oversight, you still have judgment, you
still have goals. And also on the flip
side, I think you still have
accountability, right? The question of
is, you know, if something doesn't get
done, whose fault is it? I don't think
that blaming JBT saying, "Oh, it was too
lazy, didn't do the thing," is going to
fly, right? I think that you as the
director, as the person who's who's sort
of running the show, you know that you
need to really feel that accountability,
make sure the task gets done the way
that you want. So, you have this
incredible tool, incredible
amplification, but just like
>> it's almost like everyone's going to
become the CEO of this corporation of as
many agents as you as you care to
orchestrate. Um, it's going to be just
like a very empowering world where
people can do so much more of they want
what they want to do. Um, and I think
that it'll be a world where people will
be able to make more money. I think that
people will be able to improve their
health. I think you'll be able to learn
anything you want to learn and you'll be
able to spend your time the way that you
want to. And I think all of that means
that if you have ambitions that you want
to realize that they're more insight
than ever. And I I know this very
concretely from talking to different
users, different customers, people who
run small businesses on top of chatbt
and say, "I would not have been able to
do this otherwise." like people who
actually quit their job to pursue a
business because they said hey I could
do this now because chat can help me
with so much of it and that kind of
empowerment where people wouldn't have
been able to take the leap before but
now can I think that that is it's like
the American dream on steroids like that
is what we're unlocking
>> and it's I mean you're you're painting
such a rosy picture and I do see that
too even in my own work as like a you
know solo media entrepreneur and we can
land it here but I also you know people
feel very negatively about AI especially
in the United States as And I'm
wondering if like the stakes are
actually you have to deliver on the
product you're describing that will
empower people this way for that to
shift. Maybe that is the thing that
causes sentiment to shift because right
now it feels increasingly negative in a
lot of corners of society. Uh and I'm
wor I'm curious like as the person you
know responsible now for for the
business and chatbt and all of that do
you feel that do you feel that
connection?
>> We we feel it. We feel the
responsibility. We feel the gravity and
we definitely feel the criticality of
both us delivering on this vision and
helping people understand it. Helping
people understand who we are
>> better than we have in the past and help
them understand our intentions, why
we're doing this. And I do think that
there's a different set of sentiments
depending on where you go in the world.
Yeah. Right. That it's very clear in
America like we understand that we have
a lot of uh you know we need to deliver
on this this vision to really help
people lift themselves up to really
empower people. um countries like Korea,
Japan, I think that actually sentiment
is very positive towards these tools.
People really see that potential and I
think that where America sits right now
is at an incredible moment with
incredible opportunity and position.
Like we are the leader in this
technology. This is going to be the most
important technology that humanity ever
creates, right? This is going to be this
technology that will really shape how
people live, how people work, how people
accomplish goals, how economy, how the
economy runs, how countries relate to
each other, all of these things. And
America leads in that. And I think that
that opportunity is one that we should
not take for granted. We should take
make the most of it. And so that has
been one of the biggest realizations for
me over the course of OpenAI is to
realize that this is going to be so
important and it has to be something
that is broadly empowering, right? It
has to be a technology that's available
to everyone to lift themselves up. But
America leading and ensuring that the
values that are represented by this
technology, the way that it plays out in
the world being something that enhances
democratic values that really does
uplift all of those those um you know
the core principles that we hold dear
that is at the core and so I think that
it's about how we use it but also about
how do we use this leadership
opportunity and how do we ensure we
don't squander it.
>> Yeah, those are all my questions. Do you
feel like there's anything on your mind
that you want to say or anything that
we're missing?
>> So many things.
>> There's a lot of things.
>> There's so much ground to cover. I think
at the end of the day, what we are doing
is in some ways incredibly complex.
There's so many facets to what we're
building. You think about consumer and
enterprise and the fact we're bringing
these things together and that what an
enterprise is is going to change because
people are going to have this unlocked
wave of entrepreneurship. All that's
real. We have this research program. The
research isn't just about we're we're
off writing papers. It's like people who
really care about how their technology
will be used, making sure the safety and
alignment that uh slides, doc,
spreadsheet, all of those are good. It's
just like all of this complexity, but at
the core, there's something very simple.
We have a mission. Our mission is to
ensure artificial general intelligence
benefits all of humanity. That's what
we're here to do. Everything we do flows
[music] from that one singular goal. So
that's what we're pursuing. That's why
we started this place and that is what
we're committed to.