We at OpenAI noticed that like virtually everyone at the company is using the Codex app and Germany is actually one of the top five markets in the
world for Codex Momentum. Germany. Yeah, >> those models are so good that we're still too shy with them. >> We're still not using them to the maximum
of their abilities. It's very likely it could work for like one day, 2 days, maybe 3 days or more completely uninterrupted and come back to you with
like I'm done. Everything is up to date and I tested everything.
The codeex harnesses agent harnesses
here by AI to the DNA league
users. for
deep dive. Open AI codeex by AI to the DNA. I ask you to say the typical sentence even if I'm the host because this is what everyone knows from you
from YouTube. What's what you usually say? I usually say something like, "Hey everyone, I'm Roman >> from OpenAI." >> From Open AI indeed. >> Roman,
welcome to the podcast. We're in Paris at your at your office and you're not only uh from OpenAI,
you're one of the key people behind Codex >> and it's what I want to chat with you about today. >> Yeah, I'm excited. Thank you, Kristoff, for having
me and thank you for making the trip to Paris, too. >> Very, very welcome. We had a very nice sightseeing tour last night. I I took the team and said
like, "Okay, we at least have to do some walk around the city." So, it's it's very beautiful and uh
it was a good start. >> The goal for this podcast for me is always to like take people along and explain them that we don't want to be tool tourists
jumping from tool to tool, but more becoming like really more like a Champions League player when it comes to tools. And I want to challenge you right
away with that. Um, I made the first video about Codex last May and that was when Codex was not an
app but like a tool you were using internally and I tried it and I was like didn't work the way I wanted it. Then I opened when you released the app
the Codex app again used it with the model I think 52 and then 53 CEX. I was like okay this is working now. That was kind of the same moment I had
with Claude in December, November. And then I realized with GPD55, holy smokes, this is really working
now. How it is that Codex is not just a better IDE for developers, but now changing into a new layer of work. Is that planned or did it just happen?
Well, I think we uh it's a great uh great observation by the way. We we thought about like coding for a long time, right? But I think even the name
Codex at OpenAI goes back like six years now with the when the first like models were able to output
code and when we started like this new generation of like Codex product a year ago. We always had this vision of like agentic coding taking place like
in the cloud like agentic delegation with agents being able to write code. The reality is that the models were not quite good enough at the time. You
could still do like pair programming which was very popular in 2025. the idea of having an AI that
you can like talk to for like architectural questions, maybe like writing some files or some functions, but you could not really delegate complex like
features to build from scratch or like let alone a complex refactoring of a codebase. And now we're there. So we um we really made like strong
progress on our stack of models. And as you as you pointed out, I think 5.2 two and 53 codecs were the
models that had the turning point were like >> I could even see the very best engineers at OpenAI late last year were still writing some code >> and
as of 5.2 2 and NGBT53 codecs. This is the moment they realized that the models were so good that they did not have to write any code by hand either.
And so that's when we really started to manifest this agentic coding. The model being able to take on
very long tasks. It could be like a five minute task, but it could be like a 5 hour task or even like days of work uninterrupted. And as such, we also
realized that well the interface we've been using to write code is no longer the correct one, right? Because if you're pair programming in an IDE, it
makes sense because you're talking to the model, you're outputting a function and you're kind of
like in the flow. But if you start delegating tasks to a model that you really trust to do the right thing for like many hours at once, you know, like
what we saw last year was like terminalbased interfaces where you have like 10 or 12 tabs in your terminal and each of those agents is trying to get
your attention at some point. We knew we had to build something better for this new world of like
delegating to agents and that's how the codeex app came to be. It was really this realization that like the models had become so good that we needed
to have like a command center to be able to like not just delegate tasks to agent but also give them the right permissions, review their work and
really like be in the seat of a manager of agents if you will. >> The command center is a much better
word. I used the word agent inbox last year to explain it and it was very early days and a lot of people in my company were so afraid of using the
terminal because like we have psychologists at blackboard uh we have consultants enablement people who do enablement and they were like do I need to
write code now so what if you need to describe in a typical let's say German uh corporate company or
middleand um codeex to someone who knows chip pt and says well I I have JGP. I ask JGBT. What should I do with codeex? Because I don't write code and
you're not allowed to to use the word code. You have to describe it in a work way to >> Sure. Yeah. Um I think like if they've been used to chat GPT,
the great news is that they've already like already mastered the first step of like, oh, I can talk
to an AI and this model is very smart and it's able to actually have a very personalized response for me. and maybe if they've used strategy before,
they may have connected their calendar or they may have connected with their Gmail, you know, so they've had that personalization and that experience
so far. What is very magical with the Codex app for literally anyone now is that you have not just
access to all of the services that you have online, but also uh with the right safeguards of course and approvals, you also have access to your
computer. M >> and what that means now is that any kind of everyday task that you do on a computer, the codeex app is able to do it for you. Uh so for
some people that's writing code and that's where we started. But for some others that's like well I have
presentations to make every time I meet a new prospect or a customer or maybe it's like I work in finance and I'm manipulating Excel sheets every day.
Or maybe it's like, you know, I'm I'm working in recruiting and I need to have like a pipeline of candidates. That's triage. Like all of these things
are actually tasks that we all do on a computer. And most of them take time. They usually connect
data sources from multiple uh places. They try to visualize things. There are actions to click on. And then you realize that like something like
codeex which was initially made for coding uh can be unlocking all of that and really give you like more power and like higher productivity for what
you're trying to accomplish. >> I'm always trying to not to overhype apps unless I have a strong feeling
and I I'm a heavy user I would say but you can challenge me on that one what a heavy user defines. But um I I observed throughout the last weeks
before the podcast. Usually the first app I open in the morning is codeex. Um also to prepare the podcast. Um so I have a very neat presentation ready
on the on the phone. It's made on in the on the blackboard design. So people listening to the podcast
now they cannot see it. But um where is it? I show it later on the uh on YouTube. So it even uses the blackboat design that we have even day mode
night mode we can you can switch and do all the stuff and to me the question is since you are so strong when it comes to product I mean the job that
you had before open AI was more developer related using the API but >> my feeling is and help me to
understand it better is that you're a very deep product person at the same time you're a very deep engineering person and this is how you bring this
product to life is this the way you work on that on the app or is it more testing experimenting how do you work on that >> yeah I think it's a great
question I think like the codex team is very much like iterating quickly like part of the culture of
that team like um you know building quickly but also iterating quickly based on user feedback so you might have seen many of us on X you know every
single day trying to engage with developers with builders of any kind. And um by the way, I think that's what I love about Codex. It's the idea that
like everyone is a builder and maybe that's another way to like you know have a shorter description
your previous question which is like how would you describe Codex to anyone where now anyone can build and build can mean many things. For >>
engineers it means it means like building apps or products but for everyone else it could be like you're building something that relates to your to
your uh to your work that's very exciting. um things that were usually limited. So like, oh, I need to have
access to an engineer to do this thing. Well, now you can. And so the way we iterate on the Codex app is very much like on one side benefiting from
all of this feedback from the users that we talk to every day, but also on the other side trying to manifest the capabilities of the models that are
changing so quickly, right? It used to take us more than a year to build a new model at OpenAI and
you've seen the pace of change recently, right? Like now on average I believe since GPT5 and 5.1 I think it's been like a model every 5 weeks that's
the pace now and those models you've noticed yourself like when we went from 5.4 4 to 5.5 in 5 weeks. Also, these are not just like incremental tiny
changes. Like for some people, they've seen like, >> wow, I I've I've noticed like glimpses of AGI
there. Like I'm I'm noticing something that no other model was able to do before >> and so we're trying to see like how can we manifest these features
into codeex while at the same time talking to uh talking to users on the other side. And by the way, the reason why we try to make it so much better
on everyday work is because we at OpenAI noticed that like virtually everyone at the company is
using the Codex app day in day out. it went quite quickly far beyond the engineering team and so we noticed there was something there right when our
CFO for instance Sar Frier is using codeex non-stop and I'm like she's mastering like our you know how we do finances and Excel and all of that like
on codeex and I'm like there's some there's surely something here for others to also benefit from. >>
Do you train other teams or is it just happening one-on-one in the kitchen? I think there's a bit of both like we do some enablement but I think it's
like the natural curiosity of course like we have an edge where like when you when you have like Codex teammates at the company uh we often like refer
to uh each other as like we're codex build you know like how passionate we are about um it's very
often if you walk around like open HQ that every few minutes you would have you would hear something like just ask codeex if someone has like a crazy
idea of a very complex task like are you ask Codex. Yeah, let me ask Codex. And and you would overhear that quite often. And so I think u uh no matter
which team you're on, we try to kind of blend people as much as we can. That's why we have a strong
like inerson uh culture at the office. >> So um we're naturally well equipped to get uh everyone on board, you know. I when I when I have the feeling
for a new software or tool like back in the day I mentioned Salesforce for example we had the same in the company I always said like look take a look
in Salesforce back is there something so we don't need to talk or then later with Slack and you see
that people change the behavior what I see now as a challenge and I want to disconnect it a little bit from the tool is that the way we used to work
from the chat GBT moment in terms of chat answer chat answer chat that answer to um delegating jobs to agents and then to coordinate an agent harness
is by far more complex and more challenging. What would you explain to someone who says like help me
to understand the difference between a pure chatbot and an harness >> no matter if it's codex or any other harness and why it's so important to have a
close look into these harnesses also for companies to be prepared for the next years coming >> right I think if you rewind like a couple years right
when people were starting to deploy JPT and chatbots at scale if you remember the conversations at
the time they were like oh does the model hallucinate, does it have like does it have enough in the training data to help my company or I should
probably do like rag in order to like get my documents and my context uh into the flow of the conversations. >> Um these were kind of the main
questions at the time. Uh what's very fascinating now is that thanks to reasoning progress first, the model has
been able to call tools. And so when it's able to call tools, it's able to like look up something on the internet. It's able to look up something in
one of your data like sources within a company and all of that as we mastered and polished it over the over time, we refer to it as the kind of
harness of the model. So it's the idea that like maybe two years ago the battle was on the model layer and
we would just like get higher and higher intelligence at every step of the way. Now when you look at a model like GPT 5.5 the intelligence is
extremely high. I would even say that like GPT 5.5 is already smarter than me and like most of us at >> but what makes the difference between GPT 5.5
being a good employee uh versus like someone else within the company. It's really like how much context does
it have on how to do the job correctly? How much access does it have to the context of all the documents we've written, the conversation with
customers, the codebase we have, the conventions we have in place, things like that. And so all of that refers to the harness. It's the idea that like
okay, there's the model layer, but on top of that, there is the environment that the model lives in. the
list of tools it's able to call to whether it's like read a file write a file whether it's like connecting to like uh the web for search connecting to
my data sets and all of that harness I think is what matters when you want to really manifest this like agentic capability of the model not just doing
a task >> writing code or writing a document but the ability to cross check with multiple sources
and check its own work so for instance in the case of writing code what makes the model so good. It's like, okay, I'm gonna I'm gonna write some code,
but I'm going to test that code. In fact, better yet, I'm going to open it in the browser. I'm going to take a screenshot of what I've done. I'm going
to make sure it complies with the design, for instance. Well, no, not quite yet. Okay, let me go
back to it. Let me edit this piece of code. Let me recheck again. All of that we essentially refer to the harness, the ability to have a model, its
environment, its tools, and all of that. And so to me that's where really we um we push the frontier with a model like GPT 5.5. And that's why um when
it comes to uh when it comes to models and codecs that's why we also wanted to like have a great
foundation and make sure like Chad GPT itself for instance is now benefiting from that CEX harness already. The new agents we're building within Chad
GPT are on top of the Codex harness. And what I find very compelling as a developer is that we open source all of this. So the Codex CLI, so the
interface in the terminal, that's open source, but this Codex harness that I that I refer to, that's also
open source. >> Um, and the Codex app server also connecting that harness to our API is also open source. And I think that's quite unique and I think
frankly we don't talk like enough about this. I I shared it a couple of times, but I I I think many people don't realize that most of the other big
players are closed source >> unless something leaks, then it's also open source, but well, that's not
planned then. And and you open source it. You also shared last year um the model instructions in parts of it. I think not everything because when you
refer to the harness coordinating stuff, I think many of the things in the background are related to simple prompts getting more complex if it's many
prompts, right? >> Yep. >> Is it something you can also like get a glimpse into or is this uh pure
secret how you do it? >> No, I think like what's very interesting now that you have this harness layer and the models be becoming like so smart. So we
we bring now back into the training the harness as well. So like now that we have this like open source harness that people can introspect and they
can see like what kind of tools do we have in there, how do the things connect together. >> Um we
also bring this back into the training the model it's even better when we have this like reinforcement learning techniques um where the model gets a
reward for achieving a task. Um >> we also uh do so with the right like harness in place. So we gain even better gains. But to your point on prompts,
I think what I find fascinating now is how much >> even just a year ago we were still talking so much
about prompting techniques or prompt engineering guidance, you know, uh if you wanted to get the best out of the model, usually you had to kind of
prompt the model a certain way. Um I remember even in the early days of coding we would even say like oh you are a staff senior engineer very high IQ
and as such you may notice differences in how the code is written. All of that is really gone. What
what I love about the models to date is that there's even less to learn to use them versus a year ago >> and yet they are so much more powerful. When
I tell when people ask me like oh can I do this thing with codecs? I'm like, "Sure can like just talk to Codex the exact same way you just talked to
me. All of the same words. You don't even have to have your prompt perfectly written cuz the models
are so smart. We don't have to give them u you know as much like care in the prompting, you know. So uh so yeah, I think that's that's very uh that's
very cool. And I think we have other parts of the codeex like suite where we sometimes have very specific prompt. So for instance when we put codeex
in the form of code review code review is really powerful with codeex because it's able to find bugs
or vulnerabilities and things like that. Well for this we give codeex a very specific prompt like your job is to go after like what could potentially
be a bug and how to stack rank them and how to give feedback on these bugs. And so by giving the model this different personality it can behave
differently. But as you talk to Codex for your everyday work, what I find magical is like I barely type
anymore. I just dictate to Codex and that's my prompt. And uh often times people are surprised by this, especially if they're still in the kind of
chat GPT mode of thinking, but like when I start talking to Codex, I don't even know where I'm going to end up. Sometimes I'm just brainstorming out
loud. I might change my mind a couple times over the course of that minute uh where I'm still talking to
Codeex for what I'm trying to accomplish and yet I trust Codex to figure it out. So, and it's how would you describe the percentage of is it 30% or
40% model intelligence and then 60% the smart ways of routting through different prompts in the system prompt or how would you >> I think it's like
part model intelligence part context. I think like it's really the context of like um being able to
access these tools, right? Like if it's a let's say if it's a coding task and if I'm asking like codeex to like build a new feature or like completely
change the behavior of an existing uh part of my app like the intelligence of the model is pretty high of course but I think what matters is like all
right it the user is asking me something here let me look at the entire codebase let me piece
things together okay I understand how the whole system works the user is asking me to build this new feature but I saw over there that they had a
design system. So sure enough, I can probably reuse these components over here and bring them over there. Like >> that understanding of the whole
picture by being able to read your files, connect your context. Same thing if it's like a if it's a kind of
like a knowledge task almost. Um if I'm asking like, hey, you're going to be my chief of staff every morning at 7 a.m. I want to have a report of
everything that matters to me. all the emails I should be answering to first thing in the morning. I also want you to draft the response because you
know my style of writing. You've seen me write before. You know what matters to me. We've talked enough
about these documents and what matters for this coming quarter. These are my priorities. >> And based on all of that, you know, the model will come up
with something quite magical. And of course, it's part intelligence because the model is smart. But it really comes down to this connect to this
context, right? like the ability to connect to all of these tools to know what really matters to
Kristoff versus like the raw IQ of the model. And this is something that I that I'm observing a lot and I was asking I think pretty much since the
video last year like how would we organize the data internally and I really brainstormed a lot about it and I think in March um I stumbled upon
something I never realized although I'm so long into tech that Google all of Google is basically one
repository. So for people who are not technical, repositories technically folders and text files where all the code is I think 2 billion lines of code
and then I also stumbled upon that you guys uh in one podcast someone mentioned that you are also just you have one major repository at OpenAI. >>
So this is a very I thought like that must be a very smart way to organize it. So what we did at
Blackboard now is we have a huge repository called the Blackboard operating system um where apps are running, internal apps are running, the external
website is running like many things are running. Is this the way to organize your I mean I I I think of it in layers. I think in like one box is the
repository where all the logic is. The second would be the layer of um work conditions meaning
databases, clients and stuff that you don't have in the repository. And then you have the agent layer working on that or the human layer. >> Y >>
because the artifact in the end, the fourth layer is just the PowerPoint presentation, the website, the code, the whatever it is >> and you can then
recollect it, reconnect it to learn. So to me, this is the very simple way to think of it. Is it that
clear or is it more messy? What is your best practice on that for a company who thinks like okay h how do I need to change? Yeah. No, I think you're
you're pointing like all the right things and I think like monor repos have actually become quite popular now for many years at tech companies. Uh
even before AI was so good and I think there are many properties that are like elegant about monor repo.
um you oftentimes have like many teams working in like different products especially like at large companies like Google that you mentioned and ours
now but many others and I think oftent times like you ship a new feature of an app but it also requires an API update on a different service and
everything and like the orchestration across like dozens and dozens of repos is hard. So I think like the
monor repo has this like nice property where like all the engineers can work collectively on one codebase even though they might actually touch like
only a fraction of it. >> So interestingly this was already becoming popular before AI and AI and an army of agents are actually benefiting from that
as well because they have the complete picture of the whole codebase. Right? If there's one API
chance change to make and one front end change to make it's all in one place so they can see the whole picture. Um but I think that's right. I think
it's the idea that like you have now empowered um engineers but also any kind of employees to do more by delegating to agents and it's really this
idea of like okay like what is the best way to structure the company so that uh these agents can do
their best work. One thing for instance I love is that like our company relies a lot on Slack and I'm always telling my teammates like try to always
avoid like private channels unless something is sensitive of course and has to be discussed in private because everything else like let's say a new
idea for a project or a new prototype we're building >> if it's all in public channels I can simply ask
Codex hey catch me up on that project. I've not had the chance in the past week to like figure out what was happening. Sure enough, like dozens of
people have chimed in and like started to create uh documents or maybe some artifacts of code. Catch me up. What's the latest? What decisions have
been made? What are the decisions still to be made? And I think that's cool because Codex can access all
of these contacts too. >> Is it that you have certain rules saying like, okay, this was Ro the human and this was Ro the agent or how do you do it? >>
You mean like within Slack? >> Yeah. like is there something I'm like I give you an example I prepared um I prepared an enablement internally that we
will have within the next 2 weeks >> and within one prompt I prompted Codex and said hey well I'm
thinking of doing an enablement for the whole team building agent harnesses understanding in each domain how it works I want you to inform my
colleague from HR my colleague from uh training uh write them the email in my style prepare the presentation make it in HTML upload it to our internal
sites tool just share it with the two of them and just make them a note in the end that this whole thing
was even the even the find the dates in the calendar set everything and then quick note it was done by Christoph the agent. >> Yeah. Yeah. >> So that
they realize it >> exactly so on Slack is already what's happening. So for instance last night my I received a slack from my colleague Jason who was
working on something over the weekend u and I saw this was like a note sent by his agent because
that's already part of the signature on Slack. I was like all right. So, he had been working on this thing and I just noticed that like his agent or
Jason prompted his agent to give me an update about it. >> Um, but I think what's even more powerful is not so much the writing because that's usually
like on Slack in a conversation manner. It's pretty rare that we use um like an agent for this
except for like a small status update. It's more for the context gathering, right? like if I want to do my very best work, I got to build a very
complex understanding of what's happening and keep the latest in my mind. And I think like Codex can really be magical for this. >> A lot of people
pointed to AI and did that and saying, okay, we become more efficient and behind closed doors. I then get
the question from top level years ago, who can we let go if we do that with AI? And I said like I think it's the wrong question to ask because you
usually don't have one task connected to one FTE. That's a very >> mathematical way of thinking and the world is not like that. I said work is way
more messy. Um I refer to it and say how about you leverage you truly leverage the work and I have the
feeling with agent harnesses exploding in the last month we now have the tool to do that. But we what we also need on the other side is once you give
all the context on the stuff is trust and security built around um and also like because I mean in the end it's trust that people buy when they when
they relate on a company. Um tell us a little bit about how you build it into the product because I
see these little nuances saying okay I can use the auto mode and I know Codex ask for this and that. I for example um also use of course I know cursor
quite well I know claw code quite well um pi um and so on even openclaw and her like agents >> and we experiment a lot on what is the model or the
harness asking back if it comes to trust. um how do you build it in the product because you guys go
quite far um with computer use or what I read last night even uh locked out close and locked out Mac that the model can use but that's pretty new I
think at least I saw on X >> uh working on a Mac >> oh when someone was talking to us from a phone >> yeah like from the phone and then still use my
Mac even if it's at home stuff like that >> you need a lot of builtin trust and security how does it
work >> I think there's a few ways that we uh we think about this well first of all like we care about safety like you know it's it's paramount to us
like we want to make sure that everything we do and ship has to be like extremely safe for the users um that's why we also want to put like the
technology in the hands of people all the time as early and often as we can but in the case of like codeex
first it starts with like the mole intelligence right like with a mole as smart as GPZ 5.5 by default the model kind of like is already very rigorous
in its thinking and it's train of thought in terms of like using the right tools the right time but on top of that you mentioned like um auto review I
think that's what I recommend to most people to use actually by default where like the model when
you are in this mode will always operate in the sandbox right and it will always like execute these commands in a sandbox that are kind of like
constraints to your >> sandbox to someone who's not a developer >> so it's the idea that like instead of having like complete access to your entire OS
and file system and thing it will work on a limited scope by default and only if and when you the model
or the agent might need more it would ask for your permission right and that's very important to us >> um by the way I think another interesting note
on that on that point is that you know the way we uh designed this sandboxing is using a lot of the primitives that exist today on Mac OS right
because it's a Linux-based OS and has like um a lot of the nice properties >> but People were asking us a
couple months ago like what is taking you so long to ship the Codex app on Windows and now we did like the Codex app works on Windows but it took us a
few weeks even though it's the same codebase as the Mac app. Well, the one reason the one and only reason we took longer >> was because we wanted to
get Windows sandboxing right. the ability for the Windows users to also have the same properties and
be able to have an AI agent that runs on their machine in a very safe way. And it turns out the Windows operating system does not have as many core
primitives. And so we really wanted to get it right. Very very complex architecture, but we wanted to spend the time there to make sure that people
are like oh I installed the Codex app and now my hard drive is like everything has been deleted. That
would be a terrible experience, right? So we wanted to get this sandboxing right and so that's the that's the sandbox uh >> keyword right there and
you also mentioned like computer use so that's coming soon uh to Europe it's something I'm the most impressed by in recent weeks uh I'll show you a
demo later if you have not seen it yet. >> Yeah for the ones listening now we will do the demo and show
it on YouTube so that's where you can find it then because I'm really excited about that. >> Yeah I think what is very uh magical about computer use
and why does it even exist? Why do you need such thing? Well, we talked about the fact that you have like dozens and dozens of plugins where you can
connect your sources online. We've talked about the fact that it can access with this like safe guard
rails the files on your computer whether it's writing code or an Excel sheet or something. >> But what if you would like your agent to complete a task
for you and there's no like application for it and there's no API for it. But all there is to do frankly is like hundreds of copy and paste around
different applications that live on your laptop. >> Well, that's exactly where computer use come into
play. You can actually ask your agent and say, "Hey, could you open my iOS simulator for instance? Can you like test my iPhone app on every possible
screen size? Please feel free to click around and test screenshots all of this cuz I have no time to do it by myself." But you know go ahead and try
if there's any bugs we should fix them and things like that. So it's very powerful because you can do
QA you can do all of these tasks or even the toil that has to be part of your days. Well once again if you're asking such thing computers would ask
you are you okay with me taking control of your iOS simulator right now and you'll be like yes just for this time yes always because that's an app
that I don't really have any worries about. So at every step of the way, whether it's like connecting to
a data source online, reading your files locally or accessing apps on your computer, we try to have all of the right defaults and the right guardrails
so that the user is always in control to say yes, I'm approving this approving this agent to use this app of read my files. >> What are the heavy
users like when you talk about them doing at the moment? what did you see and where are they like give
us a little bit view into also like numbers of growth where you see codex at the moment I have a I have a strong gut feeling I also shared with the
company because um we of course we are independent from from all the players but we have a very close look at all of them and I'm currently pushing
the company a lot into heavy users and see that more and more of the people using it a lot use codex but
like what do you observe when you see heavy users and where are they typically >> yeah I The it's a great question. I mean like we've been very
fortunate to see the growth like since the beginning of the year. I think on average like since the beginning of this year like the Codex growth has
been multiplied by eight and Germany is actually one of the top five markets in the world for codex
momentum. Germany. Yeah, Germany has done 10x uh on codeex momentum. It's growing really fast. >> Um I think auto scout 24 for instance like we're
talking to them like a week ago. Uh they've deployed now like Codex to many parts of their company. um more than a thousand people use it um and great
results. I think like um what's very interesting around the momentum is to see like on one side some
of the very best engineers in the world, the most prolific engineers in the world now kind of getting into codeex or getting into GPT 5.5 to achieve
very complex things. you know, of course, like Peter Steinberger from OpenClaw, uh, he's one of the big Codex users, but I've seen many others like
Michel Hashimoto from Ashi Corp and Ghosty also moving to Codex, many like top leading open source
developers moving to Codex. >> But then at the same time, you also see the momentum on companies starting to adopt it very fast, right? Like it took
us only two weeks to go from like 3 million users to 4 million users on Codeex and we can see the adoption of these >> four million users weekly.
Yeah. In Codex now and when you look at GPT 5.5, the model alone like was the fastest to adopt of any
model launch like you know it was like 3x uh more used in the first week that any of our past models. So we can really see >> that we now have a
combination of a product like codeex and a model like GPT 5.5 that is delivering real value for people. Um and we've seen of course like engineers are
using it to like you know delegate complex complete tasks to codeex like more and more complex tasks
even like they completely rely on it. But I think what I find very exciting too is this like you know uh number of people flocking to codeex and
starting to use it for non-coding tasks and though so we've made the onboarding now much easier as a result of that because of course if you're
someone who's working on you know documents and spreadsheets and presentations you don't really want to see
like a git diff or a pull request you're like I don't even know what this word means I'm just trying to do work and so we've completely reimagined and
the onboarding so that you have a very elegant interface toggle and say don't show me the code. >> Yeah. As pure as a chip interface >> but what I
find very exciting is that the codeex app is able to morph itself depending on what you're trying to
accomplish. Right? If you're generating an image and you're like hey I would like to have a new like image for the podcast. I would like hey I would
like to um rethink the design. It should be something very visual and by the time you love this design, maybe you're like, "All right, GBD 5.5, take
this design and make it happen." And boom, now it's like a coding task, >> but maybe you're going to
be like, "Hey, I'd like to see the stats of viewership of my podcast." And all of the sudden should be a graph, you know, like. And so we're trying to
make sure that like the artifacts morph themselves depending on what you're trying to accomplish. That to me is like the real exciting part of Codex.
It's an app that's going to like universal for everyone. >> Yeah. I like I I I um saw it when I
prepared the video that you can even choose your own colors now. I'm like, "Oh, I remember that moment with Slack back in the day when that comes." >>
So, that really gave me the hint that you kind of prepare it to be the app. Are there any plans or anything you can share in order to make the
experience like more one again or be more clear and say like people are say because you have agents now in
CHP a little more playful. Not my personal style and preference I have to say but my people use it quite often. And then you have this more
professional app and when I explain it like I kind of do you know the German company Bosch >> of course they have the green Bosch for at home the
small tiny one. I said this is a chip PT it looks nice you can use it's easy and then there's this blue line more
powerful you can carry around more professional >> and this is how I explain these two. Um is there a plan to like more merge it or anything you can
share? Yeah, I think like um what I can share is like you you now see kind of like on one side we have the great fortune to have like 900 million
active users of chat GPT and it's the brand that most people have um >> you know basically equalized to
AI when they think about like a chat interface to like help them in their daily lives, personal lives or work lives. um on a product that's like very
powerful but still essentially like connecting to um your services online and for which you have a chat interface and on the other side you now have
like a very fast growing product with codeex that has like now more than 4 million users on a weekly
basis >> to do like any kind of work but really this idea of like delegating task. Codex is like my teammate doing my work. I can like control all of
these agents in one place and you see that like on chat GPT you mentioned the agents these are like agents you can share across a team for now codeex
is still mostly single player also you know you can create your agents you can create your sub
agents inside codeex but you're not quite easily able to share them with your team >> yeah we have them in the repo and then access to that >> we
should we should uh do something easier there. So obviously like we want to make sure that there's a path on onto which like users could benefit from
both sides right and I think there are still many questions in flight uh but uh we want to make sure
that like when people are at work or you know doing anything on a computer um the power of the codeex app and the momentum we're having with it that's
so amazing right now should benefit to many more users and how do we progressively reveal some of that power to the people who are using Chad GPT. So
it started already behind the scenes with a few steps right like um >> the like of course the
models are now shared by the way which was not the case even two months ago right like two months ago we had coding specific lines of models we called
it like GPT3 codeex right meanwhile on chat GPT you were maybe on GPT 5.2 so this was like a different lines of model now we've completely reunified
that as of GPT 5.4 before the same exact models that are like are in both products. >> So we won't
see a dedicated 55 codecs we >> no we're now like on the main line where like all of the reinforcement learning techniques we've done for coding and
it turns out they're also benefiting any kind of task >> are also in the same model where we're also making the personality improvements and things
like that as a kind of conversational partner. Now the second thing we've also made behind the scenes
is that most of the new features even all new features now in Chad GPT are built on top of this codeex harness I mentioned. So not just the model
itself but like access to all of these tools. >> And so we now have this great foundation where like two products are sharing the same models and the
same harness. And I think the ultimate um goal is to is to think through like what is the very best most
powerful experience we can give to any user at any given time and how do we make sure that like someone in Chad GPT >> like who's trying to accomplish
some task could have the power that we today only have in the codeex app right because it has access to their files and things like this. So we're
very much like thinking through that. I think there were a few steps along the way, but we're very
excited to see the momentum on Codeex and we're very much thinking about how do we bring that power to many more people. >> I'm we I'm definitely part
of that as I'm a little bit but it's really good. It's really good. I have to admit I I wouldn't say it in a not sponsored podcast and video because
you made a lot of effort. >> I appreciate that. >> I have one last nerd topic I want to cover. I'm
thinking a lot about um security within the companies and to the data nerds there is this topic of SQL injection when it comes to databases saying you
separate um the data you put into the rows and baskets from the ones that's structuring the database um the term prompt injection which was coined um
by who was it actually um Simon Wilson I think >> I believe I think so he did >> he talked about it
he talked about very And it means or explaining to the people listening when the model receives the message it's not just the user message but it
always receives the system prompt and the user message as a so-called payload together. >> Mhm. >> And this is by design I think and it's very hard to
change that. Um, meaning is the only way to have guard rails in a company that you build for
yourself the way to put extra prompts in or are there other ways you put into place to solve that? Is there anything that's known from the research
side to work on that? >> Yeah, I think there's a few things that happened, right? I think like before we had reasoning and the ability for the model
to have a chain of thought, it meant that like everything inside the prompt from the system prompt to
the user message all the way down >> um was going to lead to the model creating tokens right as an output and every token was committed to. >> So
there was no going back like this was the response of the model and you were kind of hoping for the best and >> you could maybe inject something in
there. What has really changed with the introduction of reasoning? Yeah. So starting with like our first
series with 01 03 and then ultimately this entire GP5 family is that the models have a chain of thought and they're able to kind of like understand
the query um before committing to an output and that has dra drastically lowered already uh the ability to do like injections and things like this
because the model is like wait a second why am I trying to do this thing? Is the user really asking for
this because this seems like not the right thing to accomplish. So all of that has already changed like a lot. Um so that's one and I think on top of
that I think when you're kind of like managing a product uh a code base or maybe like just like any kind of task at a company um we uh we have like
multiple products now that can help right like I think for instance GPT 5.5 is outstanding at code
reviews. So I really recommend companies who have not tried CodeEx yet or maybe they're using different kind of products already today. >> I think I
usually tell them like turn on code review. Don't even change all of your tools just yet if you're not convinced, but turn on code review and see what
happens. More often than not, they're like shocked by what they're discovering. >> You posted that
on X, right? That you can now also like put it in your cloud code. >> Yeah. I mean like why not, right? Like that's my team who did this. We thought
well you know we in this fortunate position. where like we live in this industry where the tools are so great >> and of course codeex is an amazing
tool but there are some other great tools out there and so we thought like there's no point in trying
to convince someone who likes their tool today to like force them to change what I thought was more interesting is the ability for them to have a
quick taste of what's on the other side right so if you're using another tool today but if you want to have a taste of codeex could it find some bugs
could it find anything compelling in your codebase that gives you a taste of what the model can do,
right? Um, especially something as powerful as GPT 5.5. Uh, we've had companies like Data Dog, for instance, that kind of actually replayed all of
their past incidents with code review and they were able to notice like that a pretty big fraction of these incidents could have been avoided by
simply turning on code review. Codex would just have found these bugs. And I remember I think it was like
one incident that would have four would have been uh avoided. But this was like three months ago. So three generations of GPT5 models ago, I'm pretty
sure by now like 5.5 most likely would have caught more than half, you And so I think that there are these things um and you know we can also do like
um with GPT 5.5 we also like encourage people to use it for code review but also for scanning for
vulnerabilities in their applications like making sure that like everything is safe and robust. And for the very most sophisticated actors, maybe
let's say like in the banking system, if you really want to battle test uh your um your codebase and your products and if for some reason GBT 5.5 who
sometimes refuse to like red team really hard something because you're literally putting yourself in the
shoes of a hacker trying to hack your own product. We also have GP 5.5 cyber uh which lets you actually red team. So you have to be a trusted actor on
this one. >> It's a special product made for that. >> Yeah. It's not so much that the capabilities are vastly different. That's not the case. It's
more about like the refusals and what can it let you go to bypass some safeguards because we trusted
you as a safe actor. You're actually trying to red team your own financial system for instance. So like that's that's that's thing and so we have this
like now umbrella that you might have seen a week ago like daybreak. So this kind of like collectively is our suite of like um security product. So
broadly speaking on the theme of security, I think like if we were to summarize this like the malls
are far more intelligent that they used to be and I think as such like um their chain of thought and the reasoning capabilities make it much less
likely uh to be like um you know uh injecting things in prompts like leading to the right uh to the wrong outcome. But also when you start turning on
a lot of these like security features and really leveraging the models capabilities for your own good,
protecting your products, your features, your users, your database, um you can really go a long way to have something that's really robust. >> Nice.
Very nice. Is there anything that comes to your mind where I have a blind spot that I forgot you to ask and I should foresee it when I think of agents
throughout the next year? Wow. Um, I mean, we could talk for hours about this topic, so it's hard to
find one. Um, maybe one thing that has been my kind of like sparkle of like filling the AGI in the past couple weeks that um we could uh close on
given how big of a deal that is, but I think people have not really used it yet. It's this command that we introduce in codec called /go >> and I find
it quite compelling because it's the simplest thing you could imagine for people listening at the core.
The idea is to give codeex with this command goal >> a goal. So what does that mean? Well, you could have an extremely ambitious goal and it could be
like hey I have these like 10 millions of lines of code that have not been touched in the past 5 years. I have no idea, but I'm pretty sure
dependencies are out of date. I'm pretty sure there are some security vulnerabilities in there and I have no
engineer to work on that, but I'm pretty worried. Can you go ahead and migrate this whole thing end to end and make sure it's completely polished?
That's the goal. And what would happen next is that GPT 5.5 would relentlessly pursue that goal until it's complete and would not be like interrupted
along the way. It's very likely it could work for like one day, two days, maybe 3 days or more
completely uninterrupted and come back to you with like I'm done. Everything is up to date and I tested everything. >> I switched it on a couple of
days ago. It's for people listening to me it was a little bit still a hack turning going into the codec cli right >> switching on /experimental and
then I could choose goals >> right >> and now I could tell it in the UI so I I'm I'm I'm pitching for
you to have a special uh toggle there switching to goals and see that because what I find interesting this is something that brings together
psychology of humans being able to set goals which is not an easy thing >> to an intelligence that also does not know like I know when the goal is
reached it needs to be like kind of measured right and I guess the model will then figure with some agents am I
there yet am I there yet am I there yet >> and once again cooling their tools to make sure the work is complete correctly if not they'll come back and
try again >> okay >> um and so we are now making this like goal obviously very elegant in the codeex app as well you'll be able to like use goal like
without leaving the Codex app. It already works today, but like we're going to make the UI even
better. >> Um, but what I find fascinating about this is that for people listening, I think it's to me like one of these frontiers were like >> you're
really starting to rethink what these models are able to accomplish. Like how many companies, I'm sure many of those listening, maybe all of them have
a backlog of bugs that like you have, you know, you have to >> work that you never touch because
it's just nobody's there can do it. Imagine there's like 50 bugs sitting in your backlog that you know you have to fix, but you know they're not that
big a deal just yet. So no one was able to prioritize these bugs. Well, you can also connect as we said Codex your tools. So maybe it's linear, maybe
it's Jira, maybe it's something else. But what if you say one evening /gall codeex can you pick up
all of these bugs that we triaged this week and they're all flagged with this particular label of I don't know UI polish or something like that crush
them all. I'm going to bed. I want to make I want to wake up tomorrow morning with a buck free application. I mean, chances are maybe it's not going
to be flawless yet, but chances are Codex will be able to solve completely a very large majority of
these bugs. >> And I'm like, wow, this is quite incredible. Like with the ability in just like one command, which is like >> take this up that like
Codex will gather the context, traverse the codebase multiple times, checks its own work, call some tools, and actually deliver that. I tested it on
my Apple's reminders app because I have a list of ideas for later that's beyond 100 points and I gave
it like five of these saying okay goal can you kind of do something with that and uh it worked quite well I was quite amazed that it like took the
Apple reminders and like pull it in and I like figure a plan and um but I only tested this for a couple of days. I'm not I'm not there yet to say like
really deep dive into that but that's a good point. I almost forgot about that. No. And I think like
uh one thing that um I noticed for myself and I think that I'm sure is the case for most listeners is that those models are so good that we're still
too shy with them. >> We're still not using them to the maximum of their abilities, right? And I think to me like setting an ambitious goal is a
reminder like damn yes that's true. We can actually send these very complex goals and these very complex
tasks. And when you start seeing codeex and GPT 5.5 do so well with those, you're like, okay, I can trust this model for much more than I thought. I
can actually delegate now most of the toil in my day-to-day preparing for meetings, catching me up. I can actually prepare customer conversations
because I can let codeex do like trustworthy research on each of them. I trust the code being written
more and more. And like now it frees up time for like okay, I want to be more creative. I want to think about like more problems my customer have.
Well, like what else could I solve for them? And so it's really like freeing up some of your time by building that trust so you can do more on the
side which I find compelling. >> So that was I I could go on for hours as you said um with you and then
continue that conversation or we do a second round next year and say where we are there and then we compare. I would say that would be very
interesting. Yeah, I'm sure it might have to be like sooner than a year at the pace of everything changing this fast, you know, like uh I'm sure
things are going to get very exciting in a few months. >> Very happy to come to Paris again and uh >> next time
I'll meet you. >> I'm overdue to go to Germany. Um I lost most of my German that I practiced in school, but maybe >> then let's let's get back to
that. So, thank you very much for giving us that insight and taking the time. >> Of course. Yeah, thanks Kristoff. Amazing conversation. And for
everyone now who's still excited and listening, we are now switching to a demo on YouTube and showing how Ram
himself is using Codex. And uh you can switch to YouTube for that. Thank you.