Ghostty creator just proved why I'm right (again)

So, a while back, I created a YouTube video titled "AI Can't Automate Anyone." In this video, I talked about how LLMs, at least in their current state, don't have the capacity to automate any kind of skilled work, especially someone who's actually good at their job. And I just came across a tweet from the creator of Ghostty, the terminal app, who in this tweet is basically just spending like 12 paragraphs trying to prove me right. So we're gonna go through it, and I want to go over this tweet because I think it's such a good example of how LLMs and AI agents fall short, and how you still need human ingenuity to solve complex problems. You can pause the video and read the whole thing if you want, but basically, this guy's got an agent running in a loop optimizing a renderer, and the goal is to minimize frame times, right? And he's written some deterministic tests to measure the result, right? So to start off, he wrote, he hand-wrote a naive renderer, a purposely unoptimized renderer that does 150,000 allocations per frame, and it takes 88 milliseconds to render a frame, right? Now, that is pretty horrible. So he started an AI agent to bring the frame times down, right, to optimize the whole thing. It's been going for four hours. It spent around $350, and here are the results. We went from 88 milliseconds to 1 and 150,000 allocations to just 500, which is a huge, huge optimization when you think about it, right? Like, this is, I don't even know the math, but this is like 50 times better or something, right? Well, as it turns out, the handwritten renderer that this guy has literally written by hand, not even using AI, has frame times of 20 microseconds, not milliseconds, microseconds. That's a whole order of magnitude of a difference, and it does zero allocations in the update path. And then it just goes on to talk about the exact same things I was talking about in my video, where basically, if you yourself don't know what the hell you're doing with AI, you can't guide properly, and you're not going to see solutions like these, where you can get 75 times better output from the agent just because you know how everything works. But if you're just some bloke who doesn't understand systems, doesn't understand coding, you're not going to be able to see solutions like these. Especially because AI is trained to sound so confident, and you're just going to end up believing it, right? Because you have no base of knowledge to draw from. So let this be your reminder to keep learning and keep building new good software. No matter how good these AI models get, they're still going to need you at the steering wheel to guide them properly. That's all for today. Thank you for watching.