No natural law that says tech gets better either. The current set of LLMs have peaked. Not enough unpoisoned training data out there. They can do it faster with better algorithms and cheaper with better chips but they can't remove the confident errors. Unlike AI, _those_ are here to stay. - ThreadSky

homebrewandhacking.bsky.social • 14 days ago

No natural law that says tech gets better either.

The current set of LLMs have peaked. Not enough unpoisoned training data out there.

They can do it faster with better algorithms and cheaper with better chips but they can't remove the confident errors.

Unlike AI, _those_ are here to stay.

Comments

soyweiser.bsky.social•14 days ago

There might even be reasons to think it cannot get much better (or at least not easily) : https://bsky.app/profile/irisvanrooij.bsky.social/post/3lpe5uuvlhk2c (note I have not looked into this research, so it could be saying something else, sorry if I misused your post then Iris).

homebrewandhacking.bsky.social•14 days ago

I just saw that post from 9 months ago today! 😀

It's saying, very rigorously, and I approximate here, that even in the best case "sure let's say you have magic flying pigs and all the atoms in the universe" that AI can't achieve human level intelligence.

soyweiser.bsky.social•14 days ago

I don't think we can draw that conclusion from this research yet (proving it would be impossible is a bit different than what is being said here I think but also I have not looked into it so don't want to speak for the research) but possible in theory vs possible in practice.

homebrewandhacking.bsky.social•14 days ago

You should read what it says.

It's an NP hard problem and I think that impossible will do as a nice colloquial way to describe that, because very, very, very, very, very, very, very, very, very, very hard would get tedious to type out repeatedly.

homebrewandhacking.bsky.social•14 days ago

Here you go, I did a rough translation.

https://bsky.app/profile/homebrewandhacking.bsky.social/post/3lpgwaf7ies2q

If you feel I've misinterpreted what is written there that might be interesting.

You telling me what you believe it might say without having read it is not interesting.

loukas.bsky.social•14 days ago

Yes, I think a lot of people are using an inductionist logic - since there has been rapid progress in computing that changed things, this must always keep happening.

homebrewandhacking.bsky.social•14 days ago

Exactly! Moore's Law has held so long it is treated like an actual law!

But fusion for example has always been 50 years away for about 60 years now.

I guess computer fondlers don't read around outside computers. 🤷

iplayguitarbad.bsky.social•14 days ago

I think of AI as like Tesla's "autopilot". They make outlandish claims in media, but in their ToS it's "you're always responsible this doesn't do what we say it does".

Same with AI. Companies push it but also disclaim any liability because all that matters is growth (at all cost).

drrelling.dev•13 days ago

We had someone do a presentation on AI writing code at my company recently, and in between all the hype he slipped in "but obviously you're responsible for the code you commit".

That really should have gotten more attention. If you make a mistake, fair enough. If you "vibe code" a mistake...

iplayguitarbad.bsky.social•13 days ago

Yup. That kind of thing should be priced into how we evaluate our coworkers/reports. If you submit bad code (or bad anything), and your excuse is "sorry, AI wrote it" then that means you're just really bad at your job, just like if you cribbed directly from Stack Overflow.

drrelling.dev•13 days ago

Someone tried that in a previous job and sent it to me for review.

It took six weeks for my sarcasm gland to fully heal after it catastrophically overloaded and burnt out 😅

homebrewandhacking.bsky.social•13 days ago

So, AI is bad at writing code, and there's a link in this thread to a mathematical proof that, to all intents and purposes, LLMs are a dead end. 🤔

I mean, I think I see a solution to the problem...

drrelling.dev•13 days ago

It's not just bad at writing code - it's good at writing code that looks right but isn't. This is why it makes me so twitchy.

iplayguitarbad.bsky.social•13 days ago

I equate AI-generated code with someone taking code directly from Substack, but without understanding any security, interop, performance, scaling, etc. issues. It may work in the moment, but it's likely going to break at some predictable time and place. But it's easy, so...

homebrewandhacking.bsky.social•14 days ago

Well, you'll be delighted to hear that it's not just the AI companies that make big promises and rely on technical small print to not deliver.

The insurance companies are showing them how such things are _really_ done.

https://pivot-to-ai.com/2025/05/12/lloyds-offers-corporate-insurance-against-ai-chatbot-errors-now-try-to-get-a-payout/

iplayguitarbad.bsky.social•14 days ago

Oh my god, I read about this one and thought my eyes would achieve escape velocity from rolling so hard in my head.

Comments

Posting Rules

Reply