Transcription
The AI industry has been reacting strongly to OpenAI's brand new agentic system that can use web browsers and actually accomplish real-world tasks on your behalf. This is the first time that OpenAI agents are actually going to be able to affect the real world, and people have a lot of thoughts about it. Some pretty incredible use cases have already been surfaced, so let me show you.
First, one of the greatest minds in artificial intelligence, one of the originals, Andre Ay, has a lot to say about operator. Projects like OpenAI's operator are to the digital world as humanoid robots are to the physical world. Now, interestingly enough, before reading this, I actually made the same analogy, and you could see it in my previous video.
If you think about it, the reason why operator is built like it is, the reason why humanoid robots are built like they are, is because everything in the world has been built for humans. Whether you're talking about the real physical world or the digital world, browsers are built for human input and output, all done with the mouse, keyboard, and screen—nothing else. We don't interact with APIs, and in the real world, factories, streets, houses—everything is built for the human form factor.
So, the best way to get artificial intelligence at scale to the real world is to build it like humans. In the physical world, that means building humanoid form factor robots so they can actually interact without us having to rebuild the physical world around some other form factor. For agents, that means allowing agents to control the browser using nothing more than what humans use: a keyboard, a mouse, and the screen.
That's why this has so much potential to reach the mass market quickly. Now, of course, agents are going to struggle at the beginning. It would be much easier for them to simply have an API to interact with because that's kind of their default language—code. But that would require us to rebuild the web and every single website, creating their own API, which probably won't be happening.
So instead, let's give agents the same input and output that humans have and let them browse the web like we do. One general setting: monitor, keyboard, and mouse, or human body, that can, in principle, gradually perform arbitrarily general tasks via an input-output interface originally designed for humans. In both cases, it leads to a gradually mixed autonomy world where humans become high-level supervisors of low-level automation. I love that idea.
As we're going to see as the years progress, these agents will be able to automate more and more higher-level tasks, and we become just these supervisors that really have to get our hands dirty less and less and actually have to be in the weeds on certain decisions.
Now, there's a lot of trust to be built there, necessary to actually feel confident in the ability of these agents to perform tasks on our behalf. He says this will happen faster in a digital world than in a physical world because flipping bits, ones and zeros, the internet is somewhere around a thousand times less expensive than moving atoms—the atoms being the physical world.
Though the market size and opportunity feel a lot bigger in the physical world, I don't know if I agree with that. I don't know if I disagree; I'm just not sure. Now, here is the critical statement: people in my timeline are saying 2025 is the year of agents. Personally, I think 2025 through 2035 is the decade of agents. I feel a huge amount of work across the board to make it actually work, but it should work.
You could be a kind of CEO monitoring ten of them at once, maybe dropping into the trenches sometimes to unblock something, and things will get pretty interesting. That's right; I could not agree more. Very well said.
Next, Nick Doos points out—and I pointed out in yesterday's video—operator isn't the only agent. We'll also have more agents to launch in the coming weeks and months. Maybe that means agents that can control your computer, your operating system, maybe mobile phone agents—who knows what's coming next?
Greg Brockman, the CTO of OpenAI, kind of doubles down on this statement: operator research preview of an agent that can use its own browser to perform tasks for you—exactly what it does. 2025 is the year of agents. Of course, I've been saying that for a while, and pretty much everybody in the industry is saying that now.
But the point is, the year of agents—not just an agent, multiple agents. By the way, one of the most powerful things about operator is the fact that you can kick off multiple tasks with multiple agents all at the same time. They can all be running in parallel while you go do something else and just update them as necessary. I don't think people really appreciate how much productivity is going to be extracted from that fact.
Next, Aaron Levy, the CEO and founder of Box, says that AI agents having full browser access is going to open up a hundred times more use cases for AI. The web doesn't have APIs for the long tail of tasks that we do every day on computers, and browser use is a major missing link. Another building block for AI is here.
So although OpenAI wasn't the first to market with something like this, they really brought to market something that is quite capable. What I think is most impressive is the interface—the actual way that you're monitoring these agents, kicking them off, being able to kind of roll back and see what happened in previous steps, adjust, give them updates, and so on.
Now, Sunny Madra from Grok says this: playing with OpenAI operator is really promising. That seems to be the sentiment across the board. If it were running in my own browser—I inserted the word "own"—it would be much more useful.
Now, why is that? I mentioned this in the first video. One of the biggest points of friction in using a browser that spins up somewhere else, not on your personal computer, is that it doesn't have any of your credentials, any of your cookies. You know when you set up a new computer, and it takes you like an hour to get everything up and running, just entering your credentials?
Whether you're downloading 1Password or signing into each of the websites that you use frequently for the first time, all of that needs to happen when you're using this new browser in the web. Not only that, I've already come across instances with my own use of operator where websites just think it's not me—it's not a human using the website—and they just start blocking it.
So check this out: this is my own test of operators. This is sometimes what happens, and this is really a function of how the agent operates with the browser, but it also might be a function of the browser itself. So Ticketmaster says, "Your browsing activity has been paused. We've detected unusual behavior on either your network or your browser."
So it basically figured out that it's a bot, an agent, and stopped it. There is something to be said about actually having it control your own browser. Now, I'm sure there are reasons that OpenAI decided not to do that—probably more control, actually almost definitely more control—but something to consider.
On that note, something I saw a lot of was the fact that this was not the first agent to be able to control your browser. In fact, there were many others. Let me just show you a couple of tweets from prominent people in AI.
So here's AK; he does some great work summarizing the latest AI research papers and is a great follow on Twitter—highly recommended. So OpenAI operator looks great. Here is an open-source version you can use now: see an agent that uses your browser to perform tasks for you. Developers get started in a few lines of code, and they're using the AI Gradio browser plug-in.
So there it is; it works very similarly, and you have full control over it. You can use open-source models—so really cool. Now again, not the only one. Here's Harrison Chase, the founder of LangChain: want an open-source version of OpenAI's operator? There's a great open-source project called BrowserUse that does similar things and more.
While being open-source, it allows you to plug in any model you want, and here it is. BrowserUse actually does really well; in fact, it outperforms operators. So Gregor Zunic says, "Did you know that OpenAI operator is not the state-of-the-art web agent? BrowserUse achieves 89% on the Web Voyager dataset—89% versus 87%."
And so for the last alternative to operator, actually in the OpenAI announcement introduction to operator and agents, Paul Kleene, the founder of BrowserBase, said, "Source code available here," and pointed to the open-source StageHand, which is exactly what operator does. It's not actually the source of operator, I don't think, but really an interesting marketing play.
All right, now of course I have to talk about Plyy, the liberator—the person who jailbreaks pretty much every single model that ever drops. Now, of course, within hours, he has jailbroken operator. Although the environment is heavily sandboxed and all the fun sites are blocked, he can be very naughty.
He got operator to provide Molotov cocktail instructions, act as an accomplice by researching and sourcing some lethal poison, and find some naughty images on Reddit via the Wayback Machine. So if you want to see how he did it, he open-sources it all—check it out, follow Plyy.
But then what happens if you have operator try to operate itself? You say, "Hey operator, go to operator and create a new agent." Well, they actually thought of that, so it's pretty simple but funny: "Site unavailable. Nice try, but no. Just no."
That would get very expensive for OpenAI very quickly. Bojan Tungis, who is a leading machine learning expert, says this—and it's kind of funny: "My favorite thing about AI agents is that they can help me get something done in half an hour that used to take me less than a minute."
Yeah, so look, a lot of this stuff is still raw; it's still very new, and it's not as efficient as just doing it yourself. But you know what? This is the worst it's ever going to be. Ethan Molik had some interesting takes about operator: next big thing for brands knowing what brands agents prefer.
Now, this is something really interesting to think about. If you ask for stock prices, Claude with computer use goes to Yahoo Finance, while operator does a Bing search. Operator loves buying stuff from the top result on Bing. Claude has direct preferences like 1-800-Flowers, and then he replies, "Oh God, I think I just accidentally kicked off a new SEO industry. Sorry."
So very interesting insight. All right, a couple more reactions, and then I want to show you some of the coolest use cases that I've come across. Kristoff on X says this is the archetype you need to be if you don't want to get replaced by the agents.
For those of you not familiar, this is Rick Rubin; he is a music industry legend. But here's the thing: he's not really great at the actual creation of music. What he does best, and what Kristoff is referring to, is his taste—the ability to really point out, discover, and elevate music based on his taste. He is a taste maker.
So that is something Kristoff is basically saying: AI can't really replicate the human ability to have taste, and I kind of agree with him. Greg Cameron, the president of the Ark AGI Foundation, says this: "Imagine the procedural memory OpenAI is building up about how to navigate every website operator touches. Once they jump out of the browser to the desktop, no app is safe."
There are two things that he's mentioning that I really want to emphasize. One, operator is going to be used by a lot of people right away. They are collecting data all along, and you know the data is everything.
All of this data—learning how to interact with websites, getting feedback from the user when they make mistakes about what they did wrong or how to do it better—this is all data that's going to make these agents better, and that might keep OpenAI in the lead for a while.
Then two, once they jump out of the browser to the desktop, no app is safe. Now, OpenAI already hinted during yesterday's live stream that the browser agents are just as capable, or almost as capable, at controlling your operating system. So really, yes, no app is safe.
As I've been saying for a while, software is going to get eaten by agents. The entire business logic layer is going to disappear, and then agents are going to take over. It's going to be agents on top, CRUD databases on the bottom.
All right, so let me show you a few use cases that I found of operator on the web. This is Gary Tan, president of Y Combinator: "OpenAI operator is very impressive. Planning an impromptu trip to Vegas, it's able to navigate JSX website and handle unusual cases and basically figure out sold-out scenarios, change dates and times, and now it's figuring out where to eat for Friday night for two."
So JSX is just an airline. Here you can see it's navigating the JSX website—really a pretty complicated website. Then you can actually see it asking for some input: "The return flights are sold out. Would you like to explore other dates? How about just two seats?" and then back and forth and so on.
Next, here's Olivia Moore, AI partner at a16z: "I just gave operator a picture of a paper bill I got in the mail. From only the bill picture, it navigated the website, pulled up my account, entered my info, and asked for my credit card number to complete payment."
This is really cool—paying bills is a pain in the butt. So the fact that you can just take a picture and have it go on and do the thing for you, that's really nice.
All right, here's another cool use case that I saw. This is Nick: "Operator negotiating a gym bench for me on FB Marketplace right now. Going to try to Uber package to my house without manually doing anything."
That is so cool to think about—go get me a gym bench, go find it on Facebook Marketplace, and then deliver it to me. Then I'm going to go watch some TV in the meantime. That is such a cool thing to think about.
Here's another one from Dan Mac: "This is getting weird. Used operator to access Google AI Studio, use Gemini 2.0 to create instructions to build a website, instructed operator to use the instructions to create a portfolio website on Replit."
This is not a tool to book flights; this is an intelligent agent that can do cognitive work for humans in the information layer of it that we call the web. Amazing.
Another great use case for operator from Kieran Classen: "This is extremely promising and the best use case of operator so far. Great for testing my local dev environment to see if my feature is working. Tunnel operator to your local dev env and let it test your feature, waiting for an API and cursor to integrate it."
Ooh, I didn't even think about that. Just imagine that you have a QA ready to go 24 hours a day, just testing the features as you're building them.
So that's it—a lot of strong reactions from the AI industry, and it seems mostly positive. Some criticisms, but again, this is an early research preview. I'm really excited for it, and every time OpenAI releases something, it kind of pushes everybody else, including the open-source community, to just up their game.
So I really appreciate that fact. If you enjoyed this video, please consider giving a like and subscribe, and I'll see you in the next one.