People creating impressive software without any sort of programming experience is one of the coolest parts of LLMs I’ve seen.
It’d be nice to know if a human is performing these checks, or if this is another automated process that Codex is already quite capable of. I’m surprised this isn’t made explicit.
In my experience bad code comes from a failure of communication: the concept is muddy, abstractions overlap and leak, direction changes constantly, and the user speaks on the layer they understand while the model operates in a different reality.
Hard disagree. AI is quickly making SWE experience less dependant in development. Someone - let’s say an accountant - doesn’t need to understand the machinery if the outputs are correct. 30 seconds of processing instead of 5 is irrelevant if the output is worth more.
The whole point here is that builders don’t need to understand or maintain their system. They simply need to trust and be capable of verifying the outputs.
I think there’s a deeper truth here with LLM-usage, because AI absolutely can cover fundamentals of most topics.
An inexperienced doesn’t know how to communicate fundamentals, and AI tends to discuss using the user’s vocabulary and depth.
Shallow Input -> Shallow Output
I wouldn’t be surprised to learn that this is where a lot of the cybersecurity issues arise from. Not from the model purposefully hardcoding an API key, but from the user being frustrated/unknowing, and explicitly demanding it’s hard-coded rather than deal with pesky things like environment variables or middleware.
I would also like to make one thing brutally clear:
If someone vibe-codes a project, then so can everyone else. Even better, they can point a model towards the existing project for grounding and say “copy this”, rather than build from scratch.
It’s for this exact reason I keep my vibe-coded projects to myself, and integrate them into workflows for myself and clients.