- 5 hours ago
Join Head of Developer Experience at OpenAI Romain Huet for a live look at how Codex can help teams bring ideas to life faster. Through a series of demos, he’ll show how Codex can help turn an idea into a working app, supporting teams as they move from problem-framing to prototyping, testing, and launch.
Category
🤖
TechTranscript
00:00Romain Reviewer 2D
00:19All right. Good afternoon, everyone.
00:22It's always a pleasure to be back here in Paris at VivoTech.
00:26My name is Romain. I lead Developer Experience at Oppo.AI,
00:29and I'm kind of acting as a bridge with my team
00:31between the research, product, and engineering functions
00:34within the company and millions of builders
00:36who are outside building with our tools and our models.
00:41It's really an exciting time. Founders are telling me
00:43that they're building faster than ever.
00:45Meanwhile, the larger companies actually do more than ever.
00:48Last year, they were talking to me about proof of concepts,
00:50and they're now going live in production.
00:53They're even empowering their entire teams using codecs and agents.
00:58So today, I'm going to show you a brief presentation
01:00about how we got here,
01:02kind of the progress we've made recently,
01:04but I want to save most of the time for some live demos
01:07to really see what we've built in action.
01:11So with that, for a long time, software gave us tools, right?
01:15It gave us something that we could interface with,
01:18but AI made all of this conversational.
01:21And for those of you who were in the room last year for my talk,
01:24this was actually my slide last year,
01:26and I was kind of explaining the progress we had made with reasoning
01:29and kind of this new era of agents.
01:32Well, the exciting thing is that fast forward to today,
01:35we are entering another shift where AI can do work for you,
01:39this era of AI coworkers, if you will.
01:42And that's the arc I want to focus on today
01:44and how Codex is one of the clearest examples of that shift.
01:48But first, when you look at this graph,
01:51the reason the shift is happening
01:53is that our frontier model has crossed the threshold.
01:56The pace of innovation is absolutely incredible.
01:59Like, I work at upon AI and I have a hard time myself
02:02keeping up with what the research team is really cooking.
02:05When you look at how fast the progress have made,
02:08we've come from a world where we took about 18 months
02:10to build a new model to now roughly six weeks.
02:13And those models can now reason across tools, files.
02:17They can work for much longer than even just a year ago.
02:21But capability itself is not enough.
02:23You also need a way to bring these two products
02:26so that people can actually use these model capabilities
02:29a lot more in their daily lives.
02:32And that really changes the way we interface with the software,
02:35right?
02:36We used to have these old patterns of chatbots
02:39where we kind of thought about them as conversations
02:42and we kind of had these turns to decide what to do next.
02:45But the new pattern is really around goals,
02:48the idea that you can give an outcome to the model
02:50and it can actually drive to these results,
02:54figuring out what it needs to get there
02:55and figuring out the steps on the way.
03:00And Codex is one of the best examples of this new paradigm.
03:03We built Codex so people could work with agents
03:06in the best way possible.
03:08And last year, we saw Codex kind of take off for coding.
03:12And this year, with the new models,
03:14it's really taken a new pace with the Codex app
03:18that you can see on the screen.
03:20And the really interesting thing now
03:22is that people actually using this Codex app
03:25are more and more non-coders.
03:27They're also knowledge workers doing any kind of task
03:30on their computer with it.
03:32So Codex is really having a moment.
03:34We're now up to 5 million users using it every week.
03:36It's like up 400% since the beginning of this year alone.
03:40And the broader point here that I want you to kind of think about
03:42is that it's kind of the first surface where, like,
03:45we can see the agentic way of working becoming real for people.
03:50You know, we talked about agents last year,
03:52but they were not quite that working all the time in production.
03:55And now, it's quite different.
03:57Maybe a quick show of hands.
03:58Like, who in the room has heard of Codex before?
04:02Amazing.
04:02Like, most of you here.
04:05That's a great crowd.
04:05Well, all of this was made possible
04:09thanks to the progress in our models, right?
04:11Like, GPT 5.5, the most recent one,
04:14is the most intelligent, intuitive-to-use model we've built.
04:16It's strong at reasoning.
04:18It's also great at working with tools.
04:21But there's one thing that GPT 5.5 actually hates doing,
04:25and that's wasting tokens.
04:27We want GPT 5.5 and our models
04:29to be the most token-efficient possible.
04:31We want to pack as much intelligence in fewer tokens
04:35at the frontier.
04:37And to double down on that, actually,
04:38when you look at these benchmarks
04:39and the progress we've made,
04:41quite often you get the same quality of output
04:44for now a third of the tokens
04:46compared to, like, just six weeks prior to GPT 5.5.
04:50So the goal for us is to bring world-class intelligence
04:53with also world-class token efficiency.
04:57And because we build Codex with Codex,
05:00we've been able to ship an epic series of features this year.
05:03And that is not even everything on the screen.
05:06I could not fit it all.
05:08But I know many of you have been waiting for some features,
05:11especially here in Europe.
05:13And I'm super excited to announce that as of yesterday,
05:16all of you here in this room can use these four features as well.
05:19The first is computer use,
05:21the ability to let Codex have its own cursor
05:23and actually manipulate the apps that you have on Windows or Mac
05:27to complete tasks on your behalf.
05:30The second is the Chrome extension,
05:32that you can actually let Codex drive work inside your browser,
05:37complete form or fill out tedious steps.
05:40I'm going to show you some of the live demos of this.
05:43And two new ways of also working with Codex on your computer
05:46with personalized memory and Chronicle.
05:49Chronicle is a research preview
05:51where we can actually gather more context
05:53from what you've seen on your screen
05:55so Codex can work even better for you and build more memory.
06:00So we're building Codex into a powerful agent
06:02that can take on work, any kind of work, in fact,
06:05that you do on a computer.
06:07And the real thing that's very interesting
06:10is how do we go beyond engineers?
06:13Maybe a quick poll again.
06:14Raise your hand if you're actually writing code every day
06:17as your primary thing.
06:20Well, that's awesome because very few of you.
06:22And you'll see that that's great
06:24because for all of you that are not raising your hands,
06:27I have great news.
06:29You may be wondering,
06:30well, what's in it for me really?
06:32Because Codex sounds like a developer product,
06:34so how come a developer product can be useful to me
06:36and non-engineers?
06:38Well, I want to double click on that
06:40by telling you how we got here specifically with software engineering.
06:45And a useful way to think about this for engineers
06:47has been thinking through the throughput problem
06:51of engineering in the era of AI coding.
06:54Well, for decades, we used to think of engineers
06:57as the scarce resource,
06:59and so because we had this one assumption
07:02that writing codes was hard,
07:04we had to carefully plan everything.
07:06We had to really think through what we wanted to build,
07:09and beyond planning,
07:12we had to carefully review every line of code,
07:14and we built the systems around this idea
07:17that writing code is hard.
07:19But agentic coding completely changed that assumption
07:21because now writing code is much easier and much faster.
07:25It dramatically widens the pipe around the build step,
07:28but the challenge is that you kind of have to think
07:30about all the other steps.
07:31Otherwise, if the pipe stays narrow,
07:34building and writing more code
07:35does not make more sense to ship features to your customers.
07:39And so with Codex, what we've done
07:41is we wanted to build an agent
07:42that would expand that capacity
07:44around the completeness of the pipe,
07:46which means we want to help engineers to plan,
07:49you know, before they actually build something.
07:51We also want to help engineers review the code
07:54and make sure there's no security issues in there.
07:56We want to help them inspect the changes,
07:58we want to help them surface issues
08:00when they're happy in production,
08:01or even get support with human review.
08:04We also want to help with Codex,
08:06the engineers kind of like do deployment operations,
08:09making sure that like the issues
08:10that surface in production
08:11can actually be taken care of with Codex 2.
08:14And the goal here is not to remove the humans
08:17of the process, it's quite the opposite.
08:19We just want to make sure that at every stage
08:21that you see on the screen here,
08:23the code output can have like a higher value.
08:26We want to make sure that like we empower humans
08:29to ship high quality features.
08:31What's interesting now is that everything I just mentioned,
08:34before and after build,
08:35has nothing to do with writing code, in fact.
08:39And between this new surface of the Codex app
08:42and the model layer,
08:44you have this one more critical piece
08:46that we developed called the harness.
08:48And the harness is actually what makes
08:50long-running agents possible.
08:52We want to make sure that you have memory,
08:54you have the ability to kind of manage context
08:57over a long time,
08:58which is why we have capabilities like context compaction
09:01in order for the models
09:02to actually retain information
09:04over a long horizon of a task.
09:07And we want to make sure the agents
09:08can keep track of their work
09:10and verify their work along the way.
09:12So what is a harness exactly?
09:14Well, you can think of the harness
09:15as the environment
09:16where the model can actually turn intelligence
09:20into repeatable work.
09:21What that means is that the model needs
09:23to have the context of the task
09:25that you want to accomplish.
09:26It has to have some tools to call
09:29to deliver on that task,
09:31a place to execute,
09:32like a sandbox that's like very safe,
09:34a way to validate the work,
09:36and finally a review path back to the human
09:39to make sure the task is actually completed correctly.
09:43And what's special about Codex here
09:45is that first of all, it's open source.
09:47And that's really cool
09:47because I don't think we quite talk about that enough,
09:50but the Codex harness that I just mentioned
09:52is fully open source.
09:53And we put it back in the training of the model
09:55so the capabilities increase with it.
09:58We also have like the CLI in the terminal
10:00and kind of the app server that the Codex app talks to.
10:04All that is open source.
10:05The second thing is we want to foster an open ecosystem.
10:08So you can actually use your ChatGPT subscription with Codex,
10:12but you can use it with open code,
10:14maybe your open claw for the open claw fans in the room,
10:17or you can use it with Pi on some other things.
10:19And conversely, you can actually use Codex
10:22with open source models too.
10:23So if that's your choice, you can do that too.
10:25And since we care a lot about safety and security,
10:29we took a few extra steps.
10:31Like for instance, when we launched the Codex app on Windows,
10:34it took us a few more weeks
10:35because we wanted to make sure
10:36that the sandboxing was gotten right.
10:39There's not as many primitives for sandboxing on Windows,
10:42but we want to make sure that the agents work precisely
10:45the way you intend,
10:46and they don't go ahead and delete files on your computer
10:49without asking for your permissions.
10:52And it turns out all of this work we've done
10:55on the harness and the tools and the environment,
10:57that's all applicable to any kind of work you do on a computer.
11:01And so today, Codex is the strongest
11:04where the work has the context, the tools, the reviews that you need,
11:08but those conditions are not unique to engineers.
11:11Those conditions show up in research, analysis,
11:14maybe writing documents, maybe like dealing with operations,
11:18and a lot of the work that we all do on a computer every day.
11:21And that's a bigger product shift, right?
11:23Because it's not about turning everyone into an engineer,
11:26but it's about taking all of this work
11:29and all of these product surfaces
11:30and giving everyone and every team at the company
11:33the ability to have an agent,
11:35and an agent that can understand context,
11:37that can create artifacts, that can test some ideas,
11:40or maybe prototype some new features,
11:42and help any kind of work move forward.
11:46All right.
11:50So, what's very interesting here is that,
11:54maybe going back one slide, actually,
11:56I think I went too fast right here.
11:59We also have the ability to show all of these abilities
12:05from the harness into all of these capabilities
12:08that you see on the right side.
12:09So, the Codex app has an in-app browser,
12:11it has the ability to annotate,
12:13it has the ability to conduct research and display artifacts.
12:16I'm going to show you all of this in the demo in a moment.
12:19And at OpenAI, every single team is actually using Codex.
12:23What that means is that our sales team
12:25is actually preparing decks with Codex.
12:27Our marketing team is testing ideas for campaigns with Codex.
12:31Even our CFO, Sarah Fryer, and the finance team,
12:33they're actually using Codex.
12:34In the latest fundraise,
12:36they were keeping tabs on investor relations using Codex.
12:39So, I'll show you some of that in a moment.
12:41And that is why we're actually bringing Codex and ChatGPT together.
12:45So, everyone will have access to the most powerful agents
12:48right in the kind of familiar product that they use every day.
12:52And so, it's that simple.
12:53The idea is to have everyone having access to the most powerful agents
12:57at work and in daily lives.
13:00And to make this even more useful,
13:02we want to make sure that people have the ability to use these agents
13:06from everywhere they work or everywhere work happens, right?
13:09Like, it can be on your laptop,
13:10it can be on your phone as you're on the go.
13:12You can talk to your Codex agents from your iPhone or Android phone.
13:15It works in your browser and in any tools that you use every day,
13:18from PowerPoint to Excel and Slack.
13:20And so, we believe that people should not have to choose
13:23kind of which surface they should use for accessing AI,
13:28but the ability to have agents that work all the time for them.
13:31And so, the idea is really, like, you shouldn't be working 24-7,
13:34but your agents in the cloud should be using and working for you 24-7.
13:39And that's the vision for bringing Codex into ChatGPT.
13:42Really having agents everywhere you need them
13:45and helping every team without forcing people
13:48into, like, changing the way they work,
13:50but bringing all of this together.
13:53All right.
13:53So, with that, enough slides for today.
13:56I'm going to move to my demo laptop
13:58and show you a few things of how maybe you can use also the Codex app
14:02and how using agents has really changed the way I work personally.
14:06So, if we can switch to my demo laptop right here.
14:09So, to kick us off, the interface should look familiar to all of you.
14:13On the left side, I have, like, my projects.
14:15And on the right side, I can just start a conversation.
14:18So, here, I'm going to try to start and dictate.
14:22Can you take a look at my calendar?
14:24I'm speaking on stage right now.
14:27I'm not even typing on my computer these days.
14:29I'm mostly dictating because that's much faster, especially here.
14:33I'm just going to send this one prompt.
14:34And, you know, it sounds like a very simple prompt,
14:37but, like, I've given no context to Codex here.
14:40It's going to have to pull out details from my calendar.
14:42I just said it was on stage.
14:44So, you will have to look at this.
14:46It's figuring out the date right now.
14:48It sounds like it's, you know, it's finding now in my calendar
14:51that I have from idea to launch.
14:54And sure enough, it found it.
14:56And there's a game tonight.
14:57And the reason why this is possible is because we have plugins, right?
15:01You can connect Codex to all of the tools and the services you use every day
15:05from your Slack, your Gmail, your Google Drive, and all of this.
15:08We even have plugins that are dedicated to some roles.
15:12So, if you work in sales or if you work, like, in marketing,
15:14you can actually just use these plugins right away.
15:17But what's even cooler is not to ask Codex for things.
15:20It's also to create things that happen automatically for you,
15:23like having Codex being proactive.
15:24So, here, for instance, I said, hey, like, I want you to be my chief of staff.
15:29So, I want you that every morning you should send me what's mattering to me
15:32and what's going on about my day.
15:34So, this is what it came up with for my day-to-day.
15:36So, you know, rate stage is the main moment.
15:38Here we are right now.
15:40I had some interesting meetings in my calendar over here.
15:42And, you know, this is obviously a demo because my Slack usually is not that quiet.
15:48And a few things about what to watch.
15:50And from here, I can actually change this automation anytime.
15:53I can actually tell Codex, the World Cup is happening right now.
15:56So, make sure to put the friends team on top every morning.
16:00And just like that, I'm going to send this prompt.
16:02And starting tomorrow morning, I will have an updated way for the chief of staff to work.
16:07Okay.
16:07Now, I want to bring one concrete example about what that means for all of you in this room
16:11in different fields.
16:12So, I'm going to go over one example app.
16:14Let's call it Hudson Bank.
16:16Imagine we're building a new bank together.
16:18I'm going to look at four different roles, starting with sales.
16:21So, imagine you're working in sales at this company like Hudson Bank.
16:26And you have a customer meeting about to happen tomorrow.
16:29Well, what you can say is simply, hey, I'm meeting this customer,
16:33pull the sales template from Box, fetch the latest on Salesforce and Slack.
16:37We had some conversations with them before.
16:39And also, with that template, make a new deck.
16:42Sure enough, in 11 minutes, this is what it came up with.
16:44I can visualize it here.
16:46And you can see it actually creates a custom deck for that customer,
16:49based on the template, and also using our image gen model
16:52to actually generate images that are tailored to them.
16:55So, it's pretty amazing.
16:57And of course, it's not going to be perfect in the first go.
16:59So, you can go ahead and edit anything that you want to edit.
17:02So, that's like one idea of how, as a salesperson, you can use Codex.
17:06But moving on maybe to marketing.
17:07Let's say we're about to launch our new mobile app.
17:10And let's say we want to refresh the website for it.
17:13So, sure enough, I gave it a few screenshots of the mobile app,
17:15and I asked it to do like a dozen paths.
17:18And this is kind of like some of them,
17:20completely using our new image gen 2 model.
17:23And it's pretty magical how every text is so crisp,
17:27how the UI of the mobile app is perfectly reflected.
17:30I can now explore some ideas for how the website could look like.
17:34But they're all pretty good directions,
17:36and this would have taken my designers so much time
17:38to even explore different ideas.
17:40So, now we can kind of reflect and then dabble down on one of them.
17:44But what about beyond websites?
17:46What if we want to do a marketing campaign for this new app?
17:49Well, sure enough, I can ask this like new creative production plugin
17:53that we have packaged and say,
17:55hey, this is the logo of the bank, and this is the screens of the app.
17:58Could you imagine what a takeover of Paris, London, New York could look like?
18:03And in 11 minutes, this is what it came up with.
18:05So, I can visualize everything directly inside the app.
18:09And these are, once again, like images created by the model.
18:13Pretty magical to see how you can see like the kind of brand and like mobile app taking off in
18:19real life.
18:20So, this is what it made for the Louvre, for instance.
18:22Interesting concepts by the river.
18:24Hudson Bank, that makes sense.
18:26And so, you can really explore multiple ideas at the same time for different cities,
18:30which is really cool.
18:32So, that's kind of how marketing can be empowered.
18:35And of course, let's say you're preparing a meeting to review it.
18:38Sure enough, I can ask to make a deck out of this
18:40and review all of these ideas with the team.
18:43Now, switching gears one more time.
18:44Imagine you work in product or data science,
18:46and we actually just launched a mobile app.
18:48So, you have launched a mobile app.
18:50I'm sure the first question in anyone's mind is like, how is the launch going?
18:53Are we actually like doing so well with this thing?
18:56And so, you can see the problem that I just sent here is very simple.
19:00It's like, can you take a look at the launch metrics and make me a dashboard by country?
19:04For some of you, that might be Stripe or Shopify as the metrics.
19:07So, it could be your product analytics.
19:08It could be anything else.
19:10Sounds like here, I need to refresh.
19:12There we go.
19:14And that's like in a couple minutes actually, what's crunching the numbers.
19:17And these are the kind of dashboard it made for me.
19:20And sure enough, I can see the onboarding funnel on the mobile app.
19:23It's pretty cool.
19:23And here, I can see that like something is kind of slightly going off in Mexico
19:28because it's really like not performing as well in activations.
19:31But sure enough, because we have intelligence at our fingertips, Codex itself actually flagged this.
19:36On the first pass, it's noticed that like Mexico seems to be a pretty interesting outlier
19:42because the phone number format seems to be accounting for some issues.
19:46And that's quite interesting.
19:47And again, you have the visuals in a couple minutes, but you also have how to interpret these numbers.
19:54Well, I want to move on to the last kind of persona for this Hudson Bank demo.
19:59So what about the engineer now?
20:00Like what is happening in Mexico?
20:02Because we know from the data that something's off.
20:04But we're not sure what the customer on the other side is actually experiencing.
20:08So here, if I bring the mobile app on the screen in the iOS simulator, I can actually tell Codex
20:14right here by taking an app shot.
20:20So an app shot is kind of like a smaller screenshot.
20:22I'll give you the explanation in a moment.
20:24But let me dictate first, and I'll say, can you review the onboarding flow in Mexico, then in France,
20:31and click around the main screens of the app to make sure everything works well?
20:36So you can think of an...
20:38It didn't catch the beginning.
20:40I'll just say, can you review?
20:41Boom.
20:42So an app shot is like a smaller screenshot.
20:44So what I did by pressing Command Command is like taking what's on my screen,
20:48giving that context to Codex, and asking it to actually be able to drive my computer to take a look.
20:55And so because I approved the simulator already to Codex, it's not going to ask me for my permission.
21:00But check this out.
21:01Now, Codex is now clicking around the different screens of the onboarding without having me explaining.
21:07And this is not my cursor.
21:08My cursor is here.
21:09So Codex is having its own cursor over here.
21:12And sure enough, it's now clicking and trying to find out, like, okay, sure,
21:16I have to change the country to Mexico.
21:18Now it's probably going to find out, like, what's the right number in Mexico to actually test this onboarding.
21:23And you can see the thinking of Codex along the way.
21:26It's like, all right, sure enough, I have to now integrate a number.
21:29And that's pretty magical because what matters here is, like, it's background computer use.
21:33So as you can tell, it just found out the issue.
21:36It sounds like despite having a valid number, the customers in Mexico are blocked.
21:41So now, as I said, it's going to try for France.
21:44But when you take a look at this, I can keep using my computer and doing something else.
21:48But in the meantime, what's very cool is that Codex can do this in the background.
21:52So I can actually keep doing something else, which is very, very magical.
21:57And so that's background computer use now available in all of Europe as of yesterday.
22:02Same thing for browser use.
22:03Any kind of tasks that you have running, you can do so.
22:06So one way to think about it is that you have all of your files on your computer that Codex
22:11can safely use.
22:12You have all of your plugins in the cloud that Codex can also connect to.
22:16So let's say your Slack or your Google Drive, like we've seen.
22:19But for any other task that might take place on your computer only, Codex is able to drive these apps.
22:25And see, I said click around the apps to make sure everything works well.
22:28So it's now going ahead and clicking around the app.
22:31So it's pretty magical to see that for the first time.
22:33And even though I should be putting it in the background, I find myself often blown away and just watching
22:38it do the work.
22:39So we'll let it run in the background.
22:41It's now clicking on the screens.
22:43Cool.
22:44But when it's doing this, imagine that I now want to share the work that we've done with my team
22:50from the dashboard you saw earlier.
22:52Imagine I build this admin dashboard.
22:55And I just simply ask Codex to pull all of the metrics that it found with the data analytics into
23:01this dashboard.
23:03Another really cool thing of Codex here is that I can actually visualize everything I'm doing in this in-app
23:08browser.
23:09But I can also annotate and pair with the model.
23:12So here, for instance, I can say, add the date of today.
23:18And I can just comment inline and say what I want to change.
23:23So here I'll say, oops.
23:25All right.
23:26And maybe I want to change some other things.
23:28Maybe this, I can actually make it bigger.
23:30So I can just do that myself by tweaking things.
23:33Press enter.
23:34And boom.
23:35The date of today just appeared.
23:36And once I'm done and I'm happy with the result, I can actually tell Codex, add sites, and say, oh,
23:44I can't apparently use dictation anymore.
23:46I'll just type for a second.
23:49Deploy this and share it on Slack.
23:54And so what's pretty magical here is that it's a new feature of Codex where, like, as soon as it's
23:57done with the changes,
23:58it's going to be able to deploy this app automatically.
24:02Like, it's going to have built-in authentication just for the company or for the team.
24:06And it's going to be able to post the link on Slack.
24:08So you can really start thinking about how now, as a PM or data scientist, I can actually start building
24:14dashboard in just a few minutes.
24:16And I can even tell Codex, oh, and update the dashboard every morning.
24:22And sure enough, if I do that, it's going to create an automation.
24:25That's it.
24:26That's all it takes to really work with Codex.
24:30And now, going back to the flow here with the mobile app, we can tell that, like, the simulator probably
24:36finished what it had to do.
24:39And now, switching back gears to the engineering team, we clearly have this bug to fix.
24:44So I can just tell Codex to fix this bug.
24:46But better yet, I can actually tell Codex to achieve a bigger goal.
24:50And so, with the goal command, you can give a larger objective to the agent.
24:55And the agent will be able to pursue relentlessly until it can verify it has completed the work.
25:01So here, I could tell, like, prepare a PR to fix all the bugs and make sure the tests pass.
25:11And by doing that, as a goal, it's going to pursue this goal.
25:15It might actually take two, three, four hours to run.
25:17We've seen some very complex tasks run multiple days.
25:20But that's great.
25:22Agenting delegation.
25:22I can just give this task to the agent.
25:25And here, in this case, if I remember correctly, I think we have 27 bugs to fix.
25:29And Codex is going to go through all of these, find them on linear, and then crush all of these
25:34issues by preparing 27 PRs.
25:36So it's pretty magical when you start thinking about, how ambitious can I be?
25:40What kind of ambitious task can I give an agent to actually pursue this work on my behalf?
25:46So that kind of wraps the portion of the demo about this Hudson Bank.
25:50And I hope this inspires you with sales, marketing, data science, engineering, and how you can apply that to your
25:57own services.
25:59Now, to have a little bit more fun, I wanted to show you a couple more examples because it's the
26:02World Cup.
26:03So for the World Cup here, I made a new app.
26:06And the theme of this talk is to turn any idea into launching them.
26:11And so, sure enough, in just a couple of prompts, I asked Codex, can you make an app to share
26:16predictions with my team?
26:17And this is what it built for the next French game.
26:19And it's pretty cool.
26:20And once again, I can annotate, keep reading on it, deploying it to sites and having them shared on Slack
26:28or Microsoft Teams if you're using that.
26:30And what's very cool about this as well is that once you share these apps, they're actually very easy to
26:38be updated by anyone on the team.
26:41So here, for example, if I want to interact with this app, what should I do for France Arak?
26:46Any particular prediction I should enter for this one?
26:51How much?
26:5210-0.
26:5310-0?
26:53That's very ambitious.
26:55I'm happy with it.
26:56You know what?
26:56I'm actually going to tell Codex to make this prediction for me.
27:00So I'll just tell it 10-0.
27:02And so, sure enough, what Codex is going to do in this case is that it's actually going to use
27:06browser use to do it on my behalf.
27:09And so what it should be doing is actually taking a look at this screen over here.
27:13And it's probably going to try to click around and submit the 10-0 on my behalf behind the scenes.
27:18It's now looking at the different things over here.
27:22It sounds like maybe it's struggling a little bit with the screen sharing now.
27:25But you can tell that, like, we also have browser use so it can click around and look at Chrome.
27:30We'll give it a moment to deal with that.
27:33Another interesting thing that I built, and I wanted to show you what ImageGen can do, is this ability to
27:38kind of turn any concepts into reality by using the new ImageGen model.
27:43In this case, I was like, hey, could you build a game that, like, simulates, like, penalty kicks?
27:48And in 30 minutes, it actually was able to generate images that are kind of beautiful, frankly.
27:53Like, a concept for the game, some assets.
27:56But then I gave it, like, hey, this is the official jersey for France and the World Cup.
27:59And then, sure enough, it started to, like, build assets.
28:02And I kind of iterated a few more turns with the model.
28:05It's like, it's pretty cool what it can generate with this consistency across the board.
28:11And so here, for instance, I'm going to give it another prompt and say, can you add an Xbox controller?
28:20And sure enough, in a few seconds, it's probably going to be able to, like, add the Xbox controller to
28:25the thing.
28:25So I'm going to turn this into full screen right there.
28:28And I think the hardest thing will be to connect this Xbox controller.
28:32It's probably going to take more time to connect it to a Bluetooth than to actually deliver the task.
28:36It's still working on it.
28:37Okay, the Bluetooth controller is connected.
28:40Let's take a look.
28:41And there we go.
28:42I can start now manipulating the game with an Xbox controller, just because I had one sitting around.
28:48It's pretty magical when you think about all the ideas that you can pursue with Codex and just, like, keep
28:54on iterating on them.
28:56It sounds like over here.
28:59It submitted the prediction.
29:00And it's now about to, like, you know, validate.
29:04And I can upload a picture just to have one more example of what Imogen can do.
29:09I actually update that and try to create my supporter card.
29:12And by creating my supporter card, it should show up.
29:15There we go.
29:15So I just gave it my picture, and it made, with Imogen, a pretty cool supporter card.
29:19So I think I'm ready for the next game.
29:21All right.
29:22So that's Codex.
29:24And there's one more thing I wanted to show you before we wrap.
29:27And it's kind of outside of Codex, but just to show you the progress we've made, not just on coding
29:32and on knowledge work tasks,
29:34and not just also on Imogen, but on another modality, which is the real-time speech-to-speech.
29:40And we recently shipped three new models called GPT Real-Time.
29:43We have GPT Real-Time 2, Translate, and Whisper.
29:46And I wanted to show you a quick preview for those of you who have not seen it before about
29:50the ability to translate anything in real-time.
29:53So here, what I'm going to do is I'm going to start a session.
29:56I'm going to start speaking in French.
29:58And as I do so, I'm going to ask the crew backstage to kind of lower the volume of my
30:02French mic,
30:03and so you can hear both at the same time.
30:06So we'll give this a go and see if it works.
30:12Bonjour à tous.
30:14Hello, everyone.
30:15Ravi d'être ici à VivaTech.
30:18We're delighted to be here at VivaTech.
30:23What is really interesting about this model is its ability not to wait for the end of the sentence to
30:30start translating,
30:31but simply to wait for the key word, often the verb, like a real translator would.
30:39So it allows us to translate now in real-time into 70 languages.
30:44So for those of you who maybe have content to create and distribute all over the world,
30:51we now have a model that lets us translate all of that in real-time,
30:54and it's pretty impressive to see how it works.
30:58Voilà.
31:05Thank you so, so much, and I can't wait to see what you build with Codex and Agents.
Comments