Making a research assistant using Antigravity

Transcript

Cloud code has taken the world of agentic coding by storm. It has become the de facto tool for it. However, there is an underrated counterpart that is looked over. I'm talking about anti-gravity from a small company known as Google. It is a code editor that offers incredible value in my view.

You can essentially use Gemini 3 Pro for free. Well, there might be some limit on the usage, but for all practical purposes, it is as good as free. Um, let's talk a bit about the state of AI coding agents as it stands in early 2026. Cloud code is clearly the leader. Terminal user interf interfaces or TUIs have like gained significant popularity over the past year.

Opens codeex with GPT 5.2 to high is among the mix. But if you want to still look at your code and want to GUI curs cursor is perhaps your best option. I really like using cursor with composer one. In my experience, I've had to switch to office 4.5 or Sonic 4.5 even only a handful of times. But today we'll test and see if anti-gravity is any good.

Um anti-gravity will help me make something I have been meaning to make for a while now. It is an AI research assistant, not in the deep research way, but something that explains specific parts of the papers to me and provides the prerequisites and makes learning or reading the paper more comfortable. I feel a Chrome extension could be the best form factor for a product like this, but I'm open to reconsidering it. Let's get started. All right.

So, I'm in anti-gravity. I will just open a new window. It really looks like VS Code. Um, so let's me just create a new folder. I'm just I'm going to call this research assistant and I'll open that folder here.

I was making a Flappy Bird game and the game was immaculate, but there was an issue with the browser and the game kept crashing even though the graphics were spectacular and the gameplay was also pretty good other than the other than the fact that it kept crashing. Um, I know Flappy Bird isn't the benchmark in 2026, but it still got the job done in one shot. All right, I have the terminal open. Uh I have the agent on the right side. I will select planning mode and I would use the highest possible thinking mode pos here a available here which is the Gemini 3 Pro high.

Um the idea is to get this done in the shortest amount of time with fewest um interruptions from my side or any form of steering other than the initial prompt I'm going to give it. All right. So let me describe what I'm looking for. Okay, there it can be a Chrome extension should be able to use it with PDFs in the browser which is obviously going be Chrome. Uh should be able to select and highlight images, text, and charts.

And the research assistant should explain the selected content. There should be a mode to get the prerequisites. I'm butchering the spelling, but I'm sure it's going to understand it. of the highlighted piece and also an explain mode where the user just gets the explanation for the high lighted text. Um, now comes the question of which AI model I'm going to use.

Um, let's go with uh Gemini 2.5 flash. If it's not good, we can switch over. But I think it's cost efficient, fast, and quite smart. So, my first attempt would be with Gemin Flash. All right.

Uh, that looks good enough to get started. Let's see what it's going to come up with. Uh, anti-gravity has been in preview mode and obviously I don't think it's going to be in preview mode for long. Well, Google might just decide to kill it off altogether. Or they might realize that it has a big enough user base that they can offer it under that Google AI plan.

Either way, it's not going to let it's not going to last in this price tier, which is free for long. Um, I would suggest make the most use of it while you can. What's interesting here is that you get the voice mode as well. I haven't seen voice mode in Kursa yet. Maybe it's in the works or maybe they have already shipped it.

Who knows? They ship everything um in real quick time and they have many updates notorious for that. Maybe they have it in the mix, but I haven't seen it on cursor yet. So, um I'm going to test the voice mode as well and see if it's any good. I'm guessing it's going to be pretty useful because people like to speak and they can talk at a much faster rate than they can type. All right.

Uh looks like it has generated an icon for me which is quite good. Um, and the interesting bit is it doesn't have that Gemini watermark that you see on images generated by Nano Banana Pro and the like. All right, it wants me to accept. Uh, okay. I think this is just for resizing the icons because Chrome extensions allow you to take 16x6, 48x 48, and 128x 128 sized images.

I'm going to say accept. I like my AI agents to go rogue. Give them complete access to my codebase. Obviously, if I don't intend to keep it usable in the long run, otherwise I have to review it. They're still not competent enough to be left unsupervised.

Uh you can see the train of thought focusing on the core. I just hope the UI is good. Uh well, I don't think this is a very complex project. I intend to make it more complex by having to share the snippets with others, save it to your uh bookmarks folder or something. Maybe have a standalone website where you could do that.

So, I probably should have mentioned this in the initial prompt, but I think it's going to be uh quite clunky if I had done that. So, I could ask for a new feature. All right. What is it doing? Yeah, looks like we have content.js. Uh so much code has been commoditized and anyone can generate any code they want in the blink of an eye.

Well, not blink blink of an eye but several blinks of an eye. Uh it has implementing background and content scripts with Gemini integration. Okay. Okay. So, it supposed to Oh.

Oh, you can see this. This is pretty bad. It thinks that 2.5 is not a public stable version. Uh that's why it's using 1.5. I've seen this issue with composite one as well.

It feels it it it uses 1.5 for some reason. Gemini 1.5 even though we have three Gemini 3 available publicly. But I didn't expect this to be the case with um with Gemini 3 itself. Okay, I think it's done coding. Uh complement I have completed the implementation.

Okay. Okay. Let's see the walk through. Uh, all right. It's the usual stuff.

Okay, let's open Chrome. Oh, yeah. This is this is the awful part. It just changes my entire IDE. So now I don't have any of my previously saved Chrome tabs or any history.

Uh it rewrites my original Chrome. It happened with me uh once earlier and I somehow fixed it. So let me just enable developer mode. Load unpacked. Uh, I guess I have to go to dev research assistant.

Okay. Um, strange that it doesn't have the um icon that it set. All right. Uh, all right. Where do I set my API key? Okay.

Settings. Is it this one? I don't think so. Okay, let me just ask it. Uh, let's try the voice mode. Hello allow.

Where do I enter the Gemini API key? Oh, okay. Got it. Uh, let's go to hugging face. Um, okay. Hogging face papers page.

We're going to go to the daily papers page. Let's look at this one. Um, number two, paper of the day be PDF. All right. I'm going to pin the agent the research assistant.

I mean, okay. Um, why can't I use it? Let's see. So, clearly I can't use it. Um I'm unable to access the extension after unpacking. Let me see if there is any errors here console.

Uh think I could see the errors here. Click this. Do I see anything? Not quite. Okay. Details.

All right. Let's just ask it to fix it. If I had to pinpoint the worst aspect of anti-gravity is the browser, which is quite ironic because Google owns Chrome updated extension to ensure. Okay, I'm going to go here. Click refresh.

See if it's accessible now. All right, it's accessible now. Um, I'm going to get my Gemini API key. And I think I have it here. Was this another research assistant that I was making? Uh, it should be in cursor.

Yeah, research agent. Go to the NV file. Get the API key. Paste. Save.

All right. Close the dev tool. All right. I have no idea what this paper is about. And again, I'm not able to click the select area to explain button.

And I'm going to ask gravity to fix it. So, I would I would say it's not a good start. I expected more from Gemini 3 Pro, honestly. Uh, and given the given the uh given the task and how easy it is, it should have been able to oneshot at least parts of this. So you can see the errors over here.

Anyway, I'll just click refresh. Go here. Nope. Still unable to do it. Receiving end does not exist.

So looks pretty awful. I'm going to go to settings. Save. Let me see which motor it is using. All right, it's using Gemini 2.0 flash experiment.

All right, go here. Clear all back. Refresh. Hopefully it works now. Okay.

Okay. I'm just going to take a screenshot of the server and hand it to Gemini to fix it. The UI also looks quite minimal. Again, I expected something better. Claude is obviously much better than this, but I'm hopeful that anti-gravity is going to come of age at some point.

All right, refresh. Go here. Click this. Select. Why? Why? It's not even a special Chrome page.

Now I got this. Huh? It says, "Please reload the extension one last time and try select area again. Check if you're on a restricted page. It works best on regular websites or PDFs, which is exactly what I was doing. At times I wonder how did we manage to write so much code by hand?" Well, I didn't Not not so much at least because the onset of LLM coding was upon us by the time I got serious about writing code.

I started somewhere in 2017 18. Uh but I seriously started writing code for a living in 22 and we had the launch of GP 3.5 in late 22, November 2022 if I'm not mistaken. All right, looks like it has fixed it. Let's try again. And it works finally.

The abstract highlights that current multimodal LLMs struggle with basic visual tasks that even young children can easily perform. Despite excelling in knowledge heavy evaluations, to address this, the authors introduce baby vision, a new benchmark specifically designed to assess core visual abilities independently of linguistic knowledge. Baby vution consists of 388 items across 22 subclasses and four main categories. Results show that leading MLMs score significantly below human baselines on this benchmark. The research also explores using generation models for visual reasoning with baby vision gen and provides an automatic evaluation toolkit.

The code and benchmark data are available for pre-production. Okay, that looks pretty good, I would say. I'm looking for something math heavy, but unfortunately I don't think it's there in this paper. I'm really taking a glance. All right, let's go to some other paper.

I think this this one reinforce reinforcement learning paper. Um, not sure which one is it. Okay, is it this one? I just want a paper with some math in it so that I can see if it works on that or not. Okay, this is quite minimal and trivial. Where are the math heavy papers when you need them? Uh people are publishing papers left and right these days which is good.

Obviously we we want more people to be doing science. Okay. Can we use this? Um, let's see if he has something better. Nope. All right.

Yeah, let's try to understand the inference pipeline here. Uh I'm going to click on the extension select area to explain analyzing the final answer is obtained when there is only one reducing trajectory. The parkour. Is it supposed to sound like parkour? I don't think so. But all right.

Um the model takes the problem instance X and the previous method set M R minus one as context and generates multiple reasoning trajectories omega subscript R in parallel. Okay, that seems to be reasonable. Oh, nice. So it also explains you it also generates the prerequisites tab content in parallel when you select area when you select area to explain any piece of content each turn. Okay.

All right. I think that's a good enough start. And uh obviously it's not a complete product by any means, but you did see uh how it's how anti-gravity is useful for all of us. Um it's obviously quite rough around the edges and I expect that to change over time. Not anytime soon, but at some point Google is going to come up with better agent in coding for sure.

And maybe it's going to be anti-gravity over cursor as the def de facto choice for many people because it's going to come bundled with many more Google products. But who knows? Anyway, um, thanks for watching and I would try to make this a polished product with anti-gravity and then publish it to the Chrome web store.