# Coding with GPT-5

## Метаданные

- **Канал:** OpenAI
- **YouTube:** https://www.youtube.com/watch?v=PQUcIbSEBCM
- **Дата:** 07.08.2025
- **Длительность:** 7:07
- **Просмотры:** 417,869
- **Источник:** https://ekstraktznaniy.ru/video/11285

## Описание

Greg Brockman sits down with Michael Truell, Cursor Co-Founder and CEO, to chat about GPT-5's coding capabilities.

## Транскрипт

### Segment 1 (00:00 - 05:00) []

What was the first experience you had with GPT5? — When you use Cheap D5, you feel the intelligence aspect shine through. It's fast that it can do things like entire tasks interactively like refactors to a degree that feels like it hasn't been possible before. We had run into uh a pernicious bug in our dev build that affected performance and used GB5 to scope out the initial investigation was shocked at the intelligence of the model. If you go back before the era of LMS what would have taken someone weeks of onboarding to a codebase GB5 did in couple minutes. — That is just a wild fact. Yes. — Just absolutely wild. Well, I feel like today we should show how far this model has come. — Yeah. We're not just going to do kind of a small sketch of a tiny game. It's going to be a real working software application. We have a very nice wireframe here. — Yes. — Of an application that we're going to try building with GBD5 and cursor. We're going to bring this into cursor and we're going to put this wireframe in ask you 5 to go and try and build an application from scratch. We have been using 5 as our daily driver. when one uses it as a daily driver within cursor, you know, that's for things like finding bugs, planning out of PR, that's for things like in cases where, you know, a feature is pretty scoped, actually trying to do it end to end with the model over the course of a multi-turn conversation. — So maybe we should take a look at what the model's been up to. — Yes. Let's see. — Oh my gosh. Yeah. No, but I mean uh that is actually uh pretty amazing. — Fully interactive. — This has come a long way since the napkin of uh a little joke website — before. — Yeah, — fully interactive. — Work. — Oh, that delete functionality works. — Could also make these windows resizable. These panes. — That's a good idea. — I feel I see the line and I — I'm not a front-end programmer at all. — Okay. And so thing that I find very nice about these kinds of tools is that for areas where you're not a deep domain expert, you're suddenly empowered. It just removes this whole barrier to entry. — The correction ability both from, you know, doing things like getting lints for the code and running the code, but then also just from human instructions has improved a ton. Used to be they would go down a path and then they would be stuck in that rabbit hole — and they would it would be hopeless. — Yes. — And GPD5 can correct itself. Yeah, — even for onboarding, — it's kind of amazing when you have someone new on boarding to a codebase, — just the power of asking the AI, you know, where things are, how things work, and the speed at which you can come on board. — What's been the most surprising thing to you about GP5 so far? — The intelligence for the speed. GP4 comes out. G4 is really exciting. It's a little bit slow. GDP5 packs in a ton of intelligence to something that's really fast and fast enough that you can work with it interactively. GB5 is really good at detecting bugs actually and detecting thorny bugs. When you give it a lot of detail up front, it is shockingly good at actually groing that and understanding all of that detail. It doesn't have just like some kind of like fuzzy smooth pass over all the instructions that you're giving it. It actually understands where you're tank, which is a big step forward. Code migrations are kind of a killer feature for the enterprise that code migrations are one of those things that are so expensive for companies to do that there's like if you lower the barrier to that there will be just 10x more of them that happen. It's not just about the raw model capability but about surfaces like cursor that actually deliver value to the developer community. We have been surprised at its ability on real hard refactors that require trying things correcting oneself uh working across many files over a long period of time. But I think that the magic really happens when you're going about your day as normal and then you can do your work 20% faster immediately. 30% can be delegated to an AI. — I think that just shows that we're starting to be able to solve really hard problems, right? And have AIs that go off and think not just for minutes, but think for hours. And of course, where we want to go is think for days, think for months in order to solve problems that are worthy of that level of waiting. When they're quite incapable, you really have to do a lot of work to bring the tool to the model. But if the model's able to use a browser, it's desktop, suddenly integration is just a very different kind of thing. — Yeah. Feels like the start of the whole software development life cycle changing to go from this world where you're working one-on-one with kind of an autocomplete system that can be really powerful for programmers and one-on-one with a local agent to one where you're working with multiple in parallel and um orchestrating them all and having that be a fun, productive, fast development experience. I'm really excited for a world where there's like a heads-up display and it's like a little bit I mean one way to think about it is like it's a little bit more like an RTS where yeah you've got all your kind of parallel agents going working alongside

### Segment 2 (05:00 - 07:00) [5:00]

each other and like understanding what the status is of each of those being able to jump in gain context quickly intervene. Programming is this human compilation step of you know what's in your head, you know what you want to see on the screen and then you have to go turn it into formal programming languages and to the extent we can kind of remove the like weird jump through 10,000 hoops human compilation step and just have it be about the thing that you want to show up and how you want it to work. — Promise of computers is to empower humans, right? So it's like the human wants to do a thing and now there's a helper but the whole history of computers has been the human contorting ourselves to the machine, right? going writing assembly code and we've moved further up that layer of traction but AI is like finally fulfilling the promise of what computers were meant to be the whole time. — It has always felt like — or it's felt like you know we've done AI's done a lot on the like one person few person building something from scratch front but just I feel like it's underrated by people who aren't working in professional development just how far away we are from the ceiling of speeding up development in an environment where you have hundreds of people millions of lines tens of millions of lines of code. here we are having a nice conversation about the model while the model's out there doing all of this labor uh and that you know we are still involved right we still provide oversight we're still trying to understand what it's doing — over the course of the next year uh we are excited for how programming languages are going to change as this tech gets better and maybe it will be possible for you to look less at you know things like uh the underlying JavaScript and look at things that are higher level both for code review but also maybe also you know for helping you actually edit and work with the AI Do you think it is becoming more or less fun to be a programmer? — I think it's definitely becoming more fun. It's a little bit less about digging into lots of various web pages for a few hours and it's more about what do I want to show up, make it happen, make it happen quickly and keep going. — The amount of progress in this field that is still possible. We are at the 1% the. 1% of where we will be. It's been really great working with your team. What I think really makes models like GP5 shine, it's in interactive experiences like cursor. — Thank you for having me.