Open transcript
The fastest, most accurate computer using and browser using agent just launched. This is Sergil Ozair, hopefully I'm pronouncing your name right, founder and CEO of General Agents Co., previously autopiloted Tesla, deep learning at Google DeepMind, and PhD. He and his team just launched ACE. This is the fastest agent I have ever seen. One of the biggest points of friction with using computer use agents like operator from OpenAI is the speed. It is painstakingly slow and it makes a lot of mistakes. And now with ACE, it can actually run faster than what a human can do. Let's take a look. So remember, this is 1x speed. Copy a cute dog image from Google Chrome and share it with my boss on iMessage. Boom, opens it up. Cute dog images. This is 1x speed. It looks fast forwarded. Opens iMessage, shares the dog. And it can not only use the browser, it can also control your computer. This is macOS. Zip these files and upload the zip to Google Drive in Chrome. Here we go. Highlighting the files, compressing, adding tags, opening Google Chrome, going to Drive. I mean, I can't even describe what's happening fast enough. Upload files. There's archive.zip and uploaded. Highly accurate. Very, very fast. And once again, sending the Drive link to somebody on iMessage. Now, here's one I know producer Alex will really either love or hate. Here it is controlling Premiere. So slow the clip down to 50%. Okay. There it is. Speed duration, 50%. And just like that, split the clip once with Cut Tool. There it is. And split. Now export the clip to MP4. We go to the export tab. We select the specific format. And we start exporting. There it is. Now, obviously, you can do a lot of this stuff yourself. But the point is, you describe a long horizon task. And this computer use agent can do it for you. And much faster than anything else we've seen on the market. Let's look at another common example. Flight research. So find the cheapest flight from SF to NY on May 4th. Opens Chrome. Goes to Google Flights. And again, remember, look at the bottom. 1x speed. Does all the searches. Clicks all the right dates. Search. This is faster than what I would be able to do. And it can also use frontier reasoning models for more complex tasks. Let's watch this. So add these guys to my HubSpot 1x1 no import. And it's basically just an Excel file. And he just says, add these guys. Not even any more description needed than that. So we have the Excel file. Let's see what happens. Adding contacts. Okay. So we open HubSpot. Adding it one by one. Going through each of the input fields. Boom. Done. Look at how fast this is running. Now, this, as you can see, is 2x speed. But still very, very fast. And it's just doing it one by one. And it's not only fast. It is highly accurate. So as we can see here, here are some benchmarks. We have operator, which is OpenAI's essentially same version. And we can see it right there. And Ace Control small and Ace Control medium outperforming. And this is an important distinction. This is click accuracy. So we're seeing nearly 80% click accuracy with Ace Control medium. These are models. These are models. Obviously, there's scaffolding around the model, but this is a model trained in a very unique way, which I'll tell you about in a moment. But we can also see Malmo 72b, Cloud 3.7 Sonnet, UITARS, OmniParser, 2.0 Flash from Gemini, Quen 2.5 VL. But here is the key. It is so fast. And we can see operator here. And this is an action prediction latency. So 6.3 seconds for operator versus a fraction of a second for Ace Control small and Ace Control medium. And as he says right here, we're just getting started. Ace uses a new behavioral training paradigm, which I'll detail in a second. As we scale up training resources, Ace will grow more intelligent and capable. Yes, of course, the scaling laws. So this is normalized training resources at 1x, 2x, 4x, 8x, 16x. It just continues to improve. And so if you want to give it a try, go to generalagents.com slash Ace. Sign up for the research preview. I don't have access yet, but as soon as I do, you know I'm going to put together a test video. So a little bit more about the underlying model. Unlike language and vision models, which are trained on text and images, Ace is trained on behavior, the process that generates text, images, and other work outputs. If you remember, this is kind of what the promise of the rabbit device was, but this actually seems like it might be working out of the box. But of course, I'm going to come to it with a little bit of skepticism until I play with it myself. Training on behavior generalizes better as corroborated by the use of step-by-step reasoning in training frontier language models. Creating behavior data is also more natural for domain experts who simply need to record themselves performing a task using the tools they are already familiar with. They don't need to learn new tools or new processes. Ace is able to use the screen recordings, mouse, and keyboard logs to learn how to perform similar tasks. This is an untapped amount of data. And honestly, Apple and Microsoft should be taking the same approach. Apple has all of this data. I don't know if they're actually storing it, but they probably should or at least allow people to opt in. Give them the usage data. Here's where I'm clicking. Here's how I do this task. And then they can train their own models. But it seems general agents has done it. And Ace is only trained using a small fraction of the computing data used by frontier language models. As we scale up, Ace will become better. Now, let's look at some more examples. This is Book Me an Airbnb in Cairo on May 1st. These are examples we've seen before, but not at this speed. I'm going to play this. This is 1X. Book Me an Airbnb in Cairo on May 1st. Chrome. Okay, go to airbnb.com. Where? Cairo, Egypt. Change the check-in dates. Boom, find it. Select the first one. Looking around. Scroll down. Confirm and pay. And of course, it can be used as a developer as well. So let's see it work. Model not found error. Select the traceback and error message. I can't even read that fast. Okay. So we're selecting the error. Opening Chrome. Going to GitHub. We're going to a specific GitHub repo. Finding an issue. Creating an issue. There we go. Very quick. Very easy. And of course, it can control your computer. Let's look at that. So switch macOS to dark mode. And then enable Game Center. So system settings. Appearance. Dark mode. Easy. No problem. Game Center. Turn it on. Use this account. Just very easy. And so this definitely looks cool. I just wanted to give you a quick preview of what they launched. And I cannot wait to try it out. Hopefully, I get access pretty soon. I already asked the CEO on Twitter. Hopefully, he sees my message. So if you enjoyed this video, please consider giving a like and subscribe. And I'll see you in the next one.
