Asking LLMs to find flaws, then verifying those flaws are real, is more powerful than just looking for flaws without the tool.

No it isn’t. Your example is so contrived. What was happening before these documents were reviewed by LLMs? Were the lines that were incorrect causing issues? If so then a human would have found it years ago. If not then they weren’t causing problems and all the LLM has done is invent work that a human now needs to clean up. The LLM didn’t solve a problem, it created a new one!

Now you need to pay a human to make sure that the LLM isn’t actually making things worse, which wasn’t even a job before! Your company that won’t pay a human to review the document in the first place is now needing more humans to clean up a mess that wasn’t a problem in the first place.

you’re acting like the solution here is LLMs

No, you're imagining that's what I wrote. LLMs are a tool, not a solution.

Fine, you’re acting like the tool for the solution is an LLM.

A helpful tool, very helpful in some situations.

X to doubt. My job is currently filled with people using LLMs and most of what I’m doing is undoing their shit our detangling whatever LLM output they paste

The real problem that people and LLMs struggle with is: what should we be doing in the first place? LLMs don't solve that,

Correct

they do accelerate the making of a software thing when you know what you want that software thing to be.

Incorrect. They accelerate making a ton of code that is impossible to maintain, update, etc. If you know what the software is, it was never writing it that was the hard part. That’s why senior+ developers have been saying for decades that the hard part of the job isn’t coding. It’s everything else.

They also accelerating the finding of flaws during the process, enabling cleaning those flaws out faster.

Incorrect. Like I said, half my job is reading over extremely verbose LLM bullshit from colleagues to tell them every single thing the LLM said incorrectly. That didn’t happen before LLMs. People checked their work. LLMs don’t have a mind, they cannot check their work.

That's not a problem for tools to solve, that's a societal problem that goes back to - at least - the start of agriculture.

I didn’t say it was a problem for tools to solve! But it sure is a problem tools can cause! And that’s exactly what’s happening here. It’s making everything so incredibly worse that there is literally no justification for it!

Nah the EU legislation made it so that it has to be watermarked (LLM output was not watermarked until that passed, it was just image and music output) and discernible with a tool to tell. So you can check if something is AI generated. Anthropic says it should work on code.

I’ve been waiting for this. Did you use that dude’s project that was posted about last year? Wondering if it will work on the Ioniq 5

No that doesn’t work for me either.

Isn’t that what they use in Europe?

Sounds nice. We’ve got ~500 employees at our “site” (which is now distributed across 3+ states with remote workers), of those, about 8 do software development, so developers end up doing virtually all of the translation of customer “needs” to specifications / requirements. We also have a marketing department that changes their mind about twice as fast as the product development / delivery cycle.

you're acting like the solution here is LLMs, when the solution is just hiring more people. LLMs aren't 'solving' anything, they're just masking the problem and allowing the millionaires at the top to keep making tons of money.

Simple questions to an LLM like: “review this procedure, is it internally self consistent” will find issues, real issues - usually minor - but undeniable when you read the review, the problem it describes is absolutely present in the document, page 14 says “X” and page 23 says “not X” and the context is basically identical - so which are we really supposed to do: “X” or “not X”? Well, that depends. Depends on what? Additional context and detail that the already too long 50 page procedure doesn’t address.

lol and I find issues like this in LLM output of only 1 or 2 pages. LLMs can't even be consistent across paragraphs, much less hundreds of pages of context window. what an insane claim.

tell us of a software tool that isn't deterministic.

that is the argument linus is making... he is comparing a deterministic tool (a compiler that outputs the same assembly every time) to a non-deterministic tool and saying that they are the same type of tool. that require the same level of review, when they do not.

AI detection doesn’t work and leads to a lot of false positives

Anthropic have begun watermarking their outputs (which is possible thanks to EU legislation) so yes you can detect ai output.

'all he should care about is that we increased profits, the effect on everything external has no bearing on my decisions"

the earth hates people like you.