[roflmaostc]

Many of the issues could potentially be solved by modern LLMs?

Reading, analyzing and assembling documentation could be probably done by LLMs.

And by including old code and snippets into the training set, the LLM could be fairly proficient in writing this code probably too?

Maybe someone knows more about the use/not-use of LLMs in this context?

[doodlesdev]

I don't think you want LLMs touching projects that cost over $800.000.000, even to assemble "documentation" (since the LLM can't really document in as much it's translating what it's reading, because documentation includes much more information than what's stored in the code itself).

It's a cool idea, though, I'd like to see this done as an experiment :)

[philipwhiuk]

> It's a cool idea, though, I'd like to see this done as an experiment :)

Since the Bun AI LLM 'exeriment' conversion got merged I'm a lot less trusting of 'experiments' with LLMs. They seem to get shipped

[gabbagool]

My thought on this is that LLMs should be allowed to touch high stakes projects, but they shouldn't be left completely unsupervised.

Here me out.

Would you let AI manage your investments/retirement savings/etc. completely autonomously and unsupervised? Or would that make you a little nervous?

What if you had to undergo a X-Ray or a MRI. Would you ONLY want AI reading the images? Would you ONLY want a human radiologist reading the images? This is an interesting one because I would want both. In fact, I would find it questionable if AI wasn't given the opportunity to look at the images.

[roflmaostc]

I am fully aware of the costs and so on. But i can certainly imagine that LLMs help with the process of understanding and editing old code.

And of course you need to test and debug before you ship to production.

[Rohansi]

> i can certainly imagine that LLMs help with the process of understanding and editing old code.

Can you trust it with assembly language for a custom-built CPU with its own instruction set? Even if you digitized all of the documentation you have no idea what was lost since 1977 (or earlier), or what was never written down to begin with.

[rvnx]

Anthropic / Google / Meta / Airbus / Boeing / ASML EUV machines, etc, etc... are all developed using LLMs and they are much more expensive than this

[doodlesdev]

Arguably, Anthropic and Boeing are NOT what you should use to determine whether your coding practices are reasonable. Software problems with Boeing have literally killed hundreds of people due to faulty code and Claude Code is know to be a pretty buggy CLI (although extremely useful, sure).

Though I agree with the expensive factor. Perhaps, what I actually believe is that LLMs shouldn't touch code that's as mission critical as what NASA works on, even though it might be great to develop user-facing frontend software and CRUD backend code for huge corporations and projects.

[nickjj]

> Many of the issues could potentially be solved by modern LLMs?

Yesterday I asked an LLM what customizations niri made to the KDL language.

It said niri modified the language to add single line comments with //. However if you visit the official home page of KDL https://kdl.dev/, the very first example shows single line comments as being part of the official spec. There's also a whole page dedicated to comments in the spec that mention this.

The moral of the story is LLMs are honestly really really bad and I'm sincerely concerned at how they manipulate people into thinking what they produce is accurate or trustable. I didn't believe the AI because my spidey sense said that doesn't feel right, so I double checked the real source.

It's gotten to the point where I'm finding a huge majority of the time, it provides incorrect information on really basic things. Pure hallucinations. I'm at the stage now where even if you paid me, I wouldn't use it. There's a 0% chance I'd ever consider paying for it in its current state and it's upsetting because it is killing off the web in real-time. It's becoming harder and harder to find useful and accurate information.

[sublinear]

That's insane.

Of all the possible projects, this is the one where it's both highly feasible for humans to learn and historically critical that we have a full understanding and control of its operation.

[philipwhiuk]

You don't want your spacecraft that's 15.8 billion miles away pointing backwards because the AI's fairly proficient code misunderstand something.

[teeray]

“You’re absolutely right, I shouldn’t have jettisoned the RTG”