llms-will-not-replace-you/

[giscus[bot]]

llms-will-not-replace-you/

“Any sufficiently advanced technology is indistinguishable from magic” - Arthur C. Clarke
Today I’d like to talk about LLMs. But first, I’d like to talk about an impressive invention from the late 1700s.
The Mechanical Turk The Mechanical Turk

• or “The Turk” as people called it - was an autonomous chess-playing machine. Built in 1770, it went on tour across parts of the world for 84 years as a robot playing chess against human opponents.

https://www.davidhaney.io/llms-will-not-replace-you/

[paceaux]

I've predicted a degradation of these LLMs because of the concept of linguistic register; basically we all speak to different groups in different ways (called registers) and because we don't quite talk to these AIs like we would humans, they'll never learn to talk to us like humans, either. And it'll keep getting worse; they'll progressively sound weirder and weirder.

And that's of course assuming that everyone who interacts with an LLM is a good person — and we all know that is never true.

There will always be a percentage of the population that loves chaos for the sake of chaos, and they will intentionally train an AI incorrectly à la the "Tay" incident with Microsoft.

And then of course there's the whole matter that people want these generative AIs to generate text, but not that text. You know... the censorship problem. We barely know how to solve the Scunthorpe problem, but now we have to filter out ideas.

We've seen how well Musk has done with Grok and its ridiculous overcorrection for South Africa. I wouldn't be surprised if that was a basic if-then added to the code and no real training at all.

So between us accidentally training our AIs to talk weirder, the chaos that ordinary humans inject, and then us trying (and failing) at filtering malicious ideas — you are absolutely right. These LLMs are going to get worse over time, not better.

[combinatorylogic]

[cfehunter]

Do share your insights on what he got wrong, because his explanation aligns with my understanding neural nets and large language models.
I haven't implemented a neural net for a while, but the foundational tech is decades old. What has changed in your estimation?

[Paycho]

Love the "burrr you don't understand this" without so much as a single example of something he's gotten wrong (because he's not wrong).