Hacker News Reader

Comment by ashleyn | Hacker News Reader

How many were introduced by misuse of AI coding/vibe coding though?

highly unlikely for many of them. SharePoint, bitlocker, Active directory, hyper-v, rdp, DHCP and MSMQ are all software/technologies that have decades of history and long pre-dated LLMs. seriously, do people not realise it was entirely possible to write insecure or bad code before LLMs?

Sure, that's true.

It is also true that Copilot is currently in use developing Bitlocker and Sharepoint. So I wouldn't be confident saying it was one or the other.

Especially if they made heavy use of offshoring, which I would bet they did.

> seriously, do people not realise it was entirely possible to write insecure or bad code before LLMs?

Some of these threads make me think every line of code written pre-LLMs was apparently perfect in all ways. Feels like romanticizing the past.

It’s like people don’t remember the whole outsourcing trend and all the awful code that came from that.

Don't forget the damned interns!

Everyone but me!

Hey, I wrote slop at Microsoft way before it was cool.

It's funny because SharePoint and AD are so god-forsakenly awful you would think they're vibe coded if you didn't know any better

For what it's worth, at my workplace AI has uncovered quite a few issues that have been there for a decade or two and survived countless rounds of careful reviews, external security analysis, pen testing and so forth for all those years.

Vibe-coded apps probably have loads, but mostly because they're using less capable models than the people who're doing the bug-hunting. Once vibe-coders are using models like Mythos too you should expect the number of bugs in vibe-coded apps to collapse quickly, because the LLM will write the bugs but will also fix them (assuming the system prompt tells it to.)

For some reasons, models are blind to their own output...

One reason seems obvious/intuitive: because their own output matches their own biases, that is, is a direct result of their model walks.

I think this could be true if you use a single Claude context, since it has its own reasoning in the context.

But a separate code review agent does much better, in my experience.

Not my experience with my homie Claudius.

The code review agent usually has feedback to be resolved before committing, which includes bugs and unhandled edge cases. Sometimes the primary context is understandably embarrassed.

Sometimes it truly be your own people.

Not my experience.

I use Claude to build me a small web app I needed for ages, I'm using its superpowers to discuss/talk about architecture /brainstorm/build plans, and always a fresh context with the built plans.

A few times the plan implementer (subagent writing a code assigned to the task) found something lacking (more often it was test issue related to the code, not the actual flaw of the implementation plan), fixed it on the fly, or found something in the review (last step of each task). Also there's always a technical review for each "slice" (set of tasks), consisting of code review and e2e tests.

Only when it passes I do the fresh code review of the changes with the Opus or Fable again. Happens rarely rarely, but it did found a few issues.

Code works every time, I have yet to find the fault myself.

Of course there are issues and I need to read the output especially in the planning phase very carefully and yes, Opus disappointed me many times trying to weasel out from something it "agreed with me" (and entered into architecture + todos)...

Of course right after agreeing to use devcontainers it proceeded to attempt installing a handful of node modules in my os, so intended up running it in the bwrap (pain in the ass in itself).

But it works, it's fascinating, and I have the app I actually needed.

Not magical but useful.

The distinction I'd draw is between AI-assisted and AI-generated. Using AI to write isolated functions you understand and review is different from prompting your way to a complete system you can't debug. The second case is where you get surprising failures at runtime that no amount of linting catches.

It is hard to tell, the code may genuinely be decent quality or not.

That is the issue with vibe coding. Increased output but reduced understanding. So if something does go wrong, one has to hope that there is still enough understanding to address it quickly.

Is it worse than what humans were doing on their own anyway?

No, just more prolific.

Well yes if they expect to find and correct more bugs every update which is pretty much what they are saying.

If AI ever manages to make something half as atrocious as Windows XP that would probably be the first proof of AGI.

3 or 4

How many were known, and put on the roadmap because war got hot?

At Microslop? Evidently, lots.

Reply on news.ycombinator.com