Google DeepMind工程师实测Gemini 4

After I received early access to Gemini 4 Argon, I've spun up more changes into Chromium in the last two weeks than I would typically do in a matter of months. Chromium is an open source repository that powers web browsers and it's also used to power a lot of other electronics like here also here, here too, here, here, and even here. Chromium has been around for two decades. It's got about 40 million lines of code and like any longunning open- source project.

It's really complex. There's the core engine [music] itself and that's a very solid piece of technology. Then there's a set of features on the top. Uh think the bookmarks bar, the tab strip, the downloads [music] bar, right? There's hundreds of features that have been added by lots of different teams and these features have been built up incrementally and as a result they often have unexpected interactions with each other.

Allow me to demonstrate. Imagine this is the codebase and each of these is a feature. [music] I want to make a change to the bookmarks bar. I should be able to make one small change. Things that seem unrelated [music] can be connected underneath. The goal of our team is to create modular components so that we can independently develop and test our features. This is something that would have taken years of time and therefore we've tried a lot of different models in attempt to kind of speed up some of this work.

[music]

But the Chromium repository is different because of its size and that's really hard for AI agents to navigate effectively. [music] Quite often the model would kind of go off rails. It would generate code that was not the level of quality that we were expecting. There was a lot of handholding involved. Gemini for Argon seems to have surpassed these limitations. [music] I just tell it, hey, we're trying to detangle this spaghetti.

No skills, no harness. Here's [music] a memory file for, you know, just how to work with some of the existing tools. And it spins up 20 changes at a time. We typically run a couple hundred thousand tests [music] on every change. Then we have two humans that review the changes. We then run a couple million tests on both new and old devices. Gemini 4 [music] can kind of stand up 20 of these changes at a time that pass all the tests.

That's kind of one of the clearest signs of the quality of the model itself. That a typical engineer might put out one change a day. [music] 20 changes is 20 days of work. We can now think about where do we want to take the Chromium repository. Historically, we were limited by our ability to do the [music] work. Now, we're only limited by our own imagination in terms of what we can accomplish and how we can improve [music] the experience for our users.