←back to thread

Andrej Karpathy: Software in the era of AI [video]

(www.youtube.com)
1479 points sandslash | 2 comments | 19 Jun 25 00:33 UTC | HN request time: 0.422s | source
Show context
lubujackson ◴[19 Jun 25 16:57 UTC] No.44320427[source]▶
Generally, people behind big revolutionary tech are the worst suited for understanding how it will do "in the wild". Forest for the trees and all that.

Some good nuggets in this talk, specifically his concept that Software 1.0, 2.0 and 3.0 will all persist and all have unique use cases. I definitely agree with that. I disagree with his belief that "anyone can vibe code" mindset - this works to a certain level of fidelity ("make an asteroids clone") but what he overlooks is his ability, honed over many years, to precisely document requirements that will translate directly to code that works in an expected way. If you can't write up a Jira epic that covers all bases of a project, you probably can't vibe code something beyond a toy project (or an obvious clone). LLM code falls apart under its own weight without a solid structure, and I don't think that will ever fundamentally change.

Where we are going next, and a lot of effort is being put behind, is figuring out exactly how to "lengthen the leash" of AI through smart framing, careful context manipulation and structured requests. We obviously can have anyone vibe code a lot further if we abstract different elements into known areas and simply allow LLMs to stitch things together. This would allow much larger projects with a much higher success rate. In other words, I expect an AI Zapier/Yahoo Pipes evolution.

Lastly, I think his concept of only having AI pushing "under 1000 line PRs" that he carefully reviews is more short-sighted. We are very, very early in learning how to control these big stupid brains. Incrementally, we will define sub-tasks that the AI can take over completely without anyone ever having to look at the code, because the output will always be within an accepted and tested range. The revolution will be at the middleware level.

replies(3): >>44320619 #>>44320811 #>>44321020 #
AlexCoventry ◴[19 Jun 25 17:14 UTC] No.44320619[source]▶
I've seen evidence of "anyone can vibe code", but at this stage the result tends to be a 5,000-line application intricately entangled with 500,000 lines of irrelevant slop. Still, the wonder is that the bear can dance at all. That's a new thing under the sun.
replies(1): >>44320878 #
1. nsagent ◴[19 Jun 25 17:46 UTC] No.44320878[source]▶
Having worked with game designers writing code for their missions/levels in a scripting language, I'd say this has been the case for quite a long while.

They start with the code from another level, then modify it until it seems to do what they want. During the alpha testing phase, we'd have a programmer read through the code and remove all the useless cruft and fix any associated bugs.

In some sense that's what vibe coding with an AI is like if you don't know how to code. You have the AI make some initial set of code that you can't evaluate for correctness, then slowly modify it until it seems to behave generally like you want. You might even learn to recognize a few things in the code over time, at which point you can directly change some variables or structures in the code directly.

replies(1): >>44321284 #
2. AlexCoventry ◴[19 Jun 25 18:35 UTC] No.44321284[source]▶
I'm not kidding about the orders of magnitude, though. It's been literally roughly 100 lines to per line required to competently implement the app. It doesn't seem economically feasible to me, at this stage. I would prefer to just rewrite. (I know it's a common bias.)