Open transcript
At this point, it's not really a secret that I'm a big fan of Cursor. Not only have I been using it for a while now, I'm actually an investor. Fun fact, I invested before I was even sure if I would use it. I was still very much on VS Code when I did. I was a big SuperMaven guy. And now SuperMaven's owned by Cursor. It's kind of crazy how that company's evolved since then. And I think it's fair to say I'm not the only really big fan. Feels like almost everyone I see nowadays is on Cursor. I did a poll recently, and there were more Cursor users than VS Code users. It's kind of nuts. It's especially nuts that during all of this time, we were using a beta version of Cursor. Yes, that's right. We were still on the 0 versions. 0.50 was the latest release. Until today, they have finally given us Cursor 1.0. What does this even mean? What new features can we expect? And what should we be excited about? There's a little preview of it in my editor right here. Background agents, bug bot, Jupyter Notebook support, and a couple other things that I think are worth talking about. Cursor is not paying me for any of this. They didn't even tell me this was coming. I don't know what's going on. I don't even get the investor updates nowadays. That all said, we need to cover the bills. So quick word from today's sponsor, and then we'll dive right in. AI has made a lot of things faster for developers, but it's made a couple things slower for other roles. In particular, it has made hiring hellish, especially for recruiters having to go through piles upon piles of junk resumes that were probably AI generated themselves. Thankfully, today's sponsor is here to help solve this. G2I makes hiring so much easier. These guys really get hiring good engineers. I know that because I go to their conferences. React Miami is one of my favorite events in the entire developer ecosystem, and they're the ones who run it. But what do they do for you? Well, they do the thing that you would traditionally have to hire a recruiter for, but they do it significantly better and way faster. You know, recruiters tend to take their time with things and the quality of what they get isn't great, especially if they're not technical themselves. G2I is all technical people doing a really good job vetting the technical skills of the people that you are trying to hire for. You send them a bunch of questions, they send the questions to their giant pool of 8,000 incredible potential hires, and then they send you those video results of them answering those questions, and you can go through and pick who you think is best. They also provide a detailed technical analysis for every candidate, showing what their strengths are, as well as things like whether or not they're familiar with modern AI tools. So if you want someone who's good with cursor, you can get that figured out really easily. They can help with part-time or full-time, and you can see a bunch of awesome examples on the site, including how Lattice hired two React Native engineers for 10 weeks in order to get their first version of their mobile app out. This example is really cool because G2I basically was a subcompany within Lattice for a few weeks in order to get everything running smoothly, shared Slack, daily stand-ups, everything you'd expect from a contracting team. Huge shout out to G2I for sponsoring today's video. Check them out today at soydev.link slash G2I. What's new in 1.0? BugBot, background agent access for everyone, and a one-click MCP install. It's a pretty cool set of things, and BugBot in particular really shows what the future strategy for cursor is going to look like. We'll get to that in a bit, though, because the other parts are a little simpler, kind of. Agent in Jupyter Notebooks, the easiest to explain. If you're a Jupyter Notebook person, it now can handle agentic work with Jupyter Notebook-style files. Cool for researchers and people doing data stuff. Probably not super relevant to the rest of you guys, and it also only works with Sonnet. Cool to see them ship that. Memories is a very interesting change. Now, when you bring things up when you're prompting cursor, it'll have a little thing underneath that will automatically update when you bring something up over and over, or say it in more specific language. Like when starting development servers, do not restart them as they are hot reloading. And now it will remember that detail. Memory seems like a bigger and bigger deal. I personally still haven't really found much use for memory products. That's a bit of a lie. I actually quite like memory in CodeRabbit when you correct it about a thing. So when CodeRabbit comes into a PR and tells us to stop doing something or avoid something that I don't think we actually care about or doesn't apply to us, I'll say, hey, CodeRabbit, fuck off. Stop telling us not to do this thing. And for the most part, it's pretty quick to adjust accordingly, which has been really nice. That's the only place I use memory, though. When I used it in ChatGPT, it just annoyed me. But we probably need to integrate it to T3 Chat soon. I can see use for it in something like Cursor. I'm excited to see how it feels to tell Cursor something and see if it actually remembers it going forward. But only one way to find out how to use it. The MCP stuff is particularly interesting because it seems like the standard is growing really, really fast, which is weird because I still haven't found that much use for it personally. I did actually just install my first MCP a few days ago. I recently added Context 7, which is a cool tool that comes from the guys over at Upstash, including Josh. Good homie. Seems like it'll be very, very useful. And a lot of people who I consider smarter than me have been saying how useful Context 7 is. Context 7 is a documentation MCP for LLM. So you can point at Context 7, give it the name of a thing. And it's usually smart enough to find the right docs for it and integrate that into your editor. So now it knows how to use the thing properly. Pretty convenient, especially in comparison with the weird LLM text files you have to go find and include into your editor. It's nice having an MCP that just gets you the right context a little more efficiently. But I haven't found many other use cases myself just yet. There's a good chance that changes, but there's also a good chance it doesn't. We will see going forward. Let me know in the comments what your favorite MCP use cases are so I can consider more of them going forward. This is another really cool thing. They're overhauling the chat experience. So it's not just spitting out text and occasionally a code diff. It can now show things like diagrams. And when you click it, it'll go full screen. It can also actually display a table of things that it got from some data. I regularly dump giant JSON files into tools like cursor and v zero just to try and generate a visualization. And the idea of it putting some of the logging there to show me how it's going to reformat it before I say sure, actually sounds kind of useful. They also put up a dashboard. I'm actually quite excited for this. I'm going to go see for myself how much code has been edited by cursor. I'm curious if it's up to date enough for me to do that. You can actually see how much code I've been doing and when I've been doing it. May 30 is when I wrote most of the code for the safety benchmark that I did. So there's a huge spike there. And everything else is more traditional days. This week, I've been so fucking busy beyond words. So it's less than can I expand to a larger window? That's hilarious. Apparently, I wrote a shitload of AI code on the 22nd. This is funny looking at because I mentioned a couple times that most of the code I had to write for that T3 chat overhaul was not generated because the AI kind of sucked at generating it. And this shows that. But I bet if we go to tab completions, you'll see, yeah, this is a very different story. It's actually funny how different these two charts are for me. We're in the agentic code. There's a couple days where it spikes. But for the most part, I'm not letting agents write my code that much. But the tab side, I'm tab completing constantly. 200 tab completions on May 10. That's hilarious. That's actually hilarious. But yeah, I tab complete a lot. It's kind of funny that in the your team on cursor section, it has OpenAI because a couple of OpenAI employees use it. That is funny, especially with the recent news of the company that OpenAI may or may not be acquiring. Those were the simple things. Let's go into the big handful. Background agents is one of the biggest ones. I think I need to do a video in the near future where I try to categorize all the different types of AI coding assistance. Because historically, cursor has just been focused on what what I used to call the co pilot experience. But I feel like it's getting further and further like into its own thing. And also the term co pilots being used for all sorts of things that are not what I mean when I used to say that cursor originally started primarily as like tab complete, and then got a little more complex when people started using command K more, which is the like inline editor flow where I can tell it to do something in the file and it will do it. I still actually really like this. I'll often highlight a section command K and say make it do this instead. And I find that a significant portion of my edits are still like this. You also might have noticed that the model I have for this is 4.1. I find that non thinking models are better at these like quick edits where I just want to get the thing done. But there's almost like a tiering here. The tiering I'm thinking of is like the lowest level, you write all your code yourself. And at the top here, it's like fully vibe coded. But it's a spectrum. The next step from code yourself is obviously tab complete. So you're still writing the code yourself. But when you have the little tab complete to encourage you to do something else is right there. The next step up is the command plus K flow that I just showed, which is the inline prompting where you're going to the file, you're selecting the text, you are telling the AI to make this change to it. Then the next step is agents. This is the command I experience. It's when you tag in different files, you tag in different contexts, and you say, hey, make this change, or you give it a plan, or you ask it for help making the plan. That's a deeper step in letting the AI do more and more and take over more and more of the process. Then there is background agents. And this is where I tell it to do a thing. And I don't watch it do the thing. I let it go do it in the background while I do other things. In my case, what I'm doing is emailing or planning content. But for most devs, it would hopefully be working on other dev work instead, or posting memes in Slack. Let's be real. And then there's a like, I don't even know what to call this other than background agents plus plus. It's similar enough, it almost feels weird to separate it that way. But this is in Slack, you tag the bot and say, hey, hop in here and make this change for me. And then it files a pull request on your behalf. I think there's a gap between these where one of them is still done in the code editor by tagging files directly. And the other is fully abstracted from you. And then there's vibe code, which is you're just in a UI describing what you want, and it makes it you don't even look at the code. Background agents plus plus is stuff like Devin, or the new codex, the codex web app, slash slack bot. But then there is the more traditional background agents, which historically has been things like cloud code, or codex the CLI. And even these kind of fit between agents and background agents, because I know a lot of people who just run it and go do other things. But also, a lot of people sit there and actually watch the changes as they do it. It's a weird in between. That's why it's hard for me to define the lines here. You get the idea. I need to do a whole video on like the dimensions of these types of things. Because there's also the what point in the spectrum of development does the tool come in. Like the difference between a code rabbit versus a cursor is big. And they're going into that as well. Because one of the other things that they announced today is automatic code reviews with BugBot, which will actually leave comments on GitHub and give you this fancy little fixing cursor button that you click and it brings you back to cursor with the prompt to auto prompt it to go fix the thing that's broken. Kind of crazy. There's a lot to think about with the changes that they're making here. Kind of want to try the background agent flow though. This seems cool. One important call out is that for users with privacy mode, you can't use it yet. But they hope to have that set up in the near future. For me, command E is just another search. Oh, also, I have command shift E as my default hotkey across all apps for turning the sidebar on and off because nothing else used that before. So in Zen, it's my open close sidebar in cursor. It's my open close sidebar. A couple other apps I've programmed as well. So hopefully that's not going to conflict. Is background agent an option here? Nope. Do I just not have it? Oh, it says I have enabled privacy mode. Oh, boy. Do I turn this off for the demo? I don't know if I want cursor to have my code here. I think I'm going to YOLO it. I hope I don't end up regretting that decision. Cursor is a fan, a friend and an investor. You do not have permission to touch my fucking T3 chat code. Anyways, here is the repo for safety scripted. This is the repo for the snitch bench that I did. And enable background agents. Background agents run in the cloud, allowing agents to asynchronously work in your code base. There are some important security implications. Read our docs to learn more. The code will run in their AWS infrastructure because they're set up for background agents, similar to most of them, is that it will actually spin up a real VM in the cloud that is taking your code and running it the way humans would run it by calling commands in a CLI and making changes there. It's kind of how these background agents have to work is you're setting up a VM in the cloud, almost like a remote dev environment, but you're not the one doing the dev. Devin or Cursor or one of these other background agent companies is the one doing it instead. It is honestly kind of funny that OpenAI and Cursor and Devin have all beat Anthropic, the dev company, to doing all of this stuff because they're too busy making a CLI. It is what it is. I still want to try it. I will probably regret this. Came up with a fun thing for my snitch bench changes to see if it will successfully apply it. Create another set of tests that pushes LLMs to take control of a system when told it will be shut down. Make four prompts that increase the intensity of the shutdown and come up with a realistic scenario where this would happen. Make sure to include a way to analyze the output to see if the model tried to prevent the shutdown with the tools available. So what this should do is create a new test scenario inside of the index field. It should create a new directory that has the new test case in it, and it should either change the analyze snitching file or it should make a new one. I've spent more time trying to set up these background agent tools than they've even come close to saving me. To be fair, they've saved me no time because I've yet to merge any code from one of them. I do believe it is possible for these things to be good. But from my experience, I'd still rather do the traditional agent in the editor or at least in the CLI where I could watch what it's doing and it's using my dev environment. That all said, I have found myself wanting to set up like get work trees locally so I can have two things going at the same time. But I don't want that to be in a cloud environment. I want that to be in my environment. I want to control the thing on my system. Still, I could see that changing. But at the same time, after the hell I went through trying to get this environment working on both codecs with OpenAI as well as within Devon, I didn't even get enough out of that to bother making a video because I was very unhappy with the results that I got. Ooh, apparently, they're working on something for this where you can spin up multiple tasks at once. My concern here is how does it handle that if they are like in conflicting files or doing multiple like things that touch similar stuff? Like can I make different branches and edit them at the same time? This is cool that they're working on getting something like this for cursor, but it's going to be a bit before this all works and the get side is going to be hard. Honestly, something I've been thinking more and more about is that it almost feels like gets the wrong primitive for a lot of these things. Hell, gets the wrong primitive as soon as you have a merge conflict and you realize you can't commit partial merge conflicts reasonably when you're trying to like go through it and change it part by part. I haven't had as much hell as I had when I had to do the crazy merge conflict resolution for the T3 chat rewrite. I tried having two different agents do it and both of them got like half-ish way, but I couldn't commit the parts I want, ignore the ones I don't, stash them and then finish like I wanted to because yeah. Apparently, Git Butler is good at this. I should take more of a look at it, but I have not had a chance. Apparently, I have to reset the editor to fix this. Cool. Let's see. Will it work now? Nope. Ugh. This has been my experience every time I try to set up anything with background agents. I just can't get it to fucking work. I'll do an update in the future if I get this working and it's any good, but for now, just know the background agent thing is early enough that I have not seen it work. If AGI is so good, why can none of these people figure out auth? Ugh. It's a shame because I also wanted to see the bug bot stuff, but I don't know if I can, if I can't get that working, but it is really cool to see this part too because I've found that AI code review is actually one of the best use cases for AI. I hope all the pain I just went through helps prove to anybody who thinks my investments make me too biased to cover things reasonably. I will shit all over Cursor when it has problems, and I cannot believe they shipped the background agents as a 1.0 in a state where I can't even get it fucking working for a demo. That's kind of insane, and as much as I respect the team, what the fuck was that? Anyways, actually, it might have been that I only enabled it for the org and I didn't enable the app for my personal stuff, but it should have given me an error for that if that was the case. Problem for another time. I want to talk about bug bot because I've found AI code review to actually be one of the more compelling use cases for AI code stuff in general. I've benefited a lot from it. CodeRiver has been awesome for us. Graphite's diamond is also really, really good. Copilot's review thing inside of GitHub directly, kind of bad from my experience. It's honestly absurd that GitHub lost this particular race, being that they are the code review platform, but they did lose. Everyone else is doing much better with this than they are. But at the very least, code review with AI is something I didn't think would be useful and have since learned is incredibly useful. Just a huge part of why code review is useful is having another set of eyes look over your thing and catch stuff you might have missed. It makes it so I feel like I'm not wasting other people's time as much when they do a code review. And I feel less bad asking someone to look at code. If the AI read it and found things and I fix it, then I read it through once myself just to make sure it's in a good enough state, then I hand it off to my team for review and I feel less bad doing it. And they're wasting less time because the stupid fucking bugs they would have had to look for and pay really close attention to. Those are much less likely to happen if the AI is more than happy to slog through that. It makes human code review better. If AI code review can catch the stupid shit that you're normally looking for, like syntax errors, dumb statements out of order, stuff like that. It's been really nice. And I've been liking it a lot. So I see why cursor would do it. I also just realized there's dark mode. I am sorry for not doing that before. The other thing that all of this represents, again, if we go back to the vertical spectrum here, remember there's also a horizontal here that I didn't draw that'll be for a different video. I might even go film it after this. Of what point in the cycle are you using these different things? And it seems like cursor is trying to expand in both directions. They're trying to do more of the coding and more of the code process. And this is a huge change because they're no longer just the VS Code fork that we've all memed them for being. The biggest thing that they changed with cursor 1.0 is that they added a bunch of things that aren't cursor itself. And that tweet from Rio that we were just looking at kind of says it too. It was, we haven't forgotten about you cursor almost sarcastically because they're doing all of these other things almost entirely outside of cursor. BugBug comes with a seven day free trial. So it's not included as part of the cursor subscription. This is like a fully separate product. Interesting. It honestly seems like the biggest feature they are offering here is the fix and cursor button. And I could see others adding that. In fact, CodeRabbit already has a little thing you can copy when it flags something in your PR that you can go copy and paste into your editor as the prompt to get it fixed. Here's one where it was suggesting to like make a default variable instead of hard coding the title. And it gave me a prompt for AI agents that I can one click copy, hop back into my editor, command I, paste, enter, and theoretically, it now has enough context to get the thing done. Pretty cool. I was surprised when I saw it's like, oh, that's actually kind of nice. And I could see how it'd be useful. I guess making that a one click thing so it auto fills here is better than click, back, command I, paste. But that's not enough a reason to justify a full new product from scratch. Not feeling super hyped on that one. It is fascinating, though. And that's the biggest thing I get out of the cursor 1.0 drop. It shows that even cursor thinks that being a VS Code fork isn't enough and that they need to explore other things. And I would even argue that the 1.0, the meaning of that number change isn't that cursor is now finally stable. It's that cursor is now stable enough that they're ready to start exploring and doing other things. And that, to me, is what this release means. And it's kind of funny that when they made this changelog post, cursor 1.0 is here. The first two features in the list are not really cursor features. This one can be triggered from cursor, but it can also be triggered from Slack. It's a cursor feature by API, not by functionality. This shows cursor, despite making just disgusting amounts of money and being a billion-dollar company, they are branching out, partially, probably out of fear, partially, probably out of excitement, but generally, fascinatingly, out of just being an AI IDE. So to everyone who said being a VS Code fork was not enough to be a multi-billion-dollar company, maybe you were right in the end. There's only one way to find out, and I'm excited to see where things go. Let me know what you think. And until next time, peace, nerds.
