[nomat]

Imagine like an LLM trained on everything you've ever done since birth included as part of the data set. and then one of those for each other person in the world. they could slice and dice you into a thousand different groups.

at some point technology is going to stop being something that we interact with through a screen and a keyboard. Google maps will be there, where ever you go, all the time. No more getting lost or wondering where anything is located. You will just know. or at least, be told.

Something else to consider is the fallibility of human memory. As we generate more information we're gonna have to start abstracting and delegating some of those storage and retrieval tasks to the computer, trusting it completely. Today maybe you can remember things but what about your parents?

[jbergens]

I don't remember everything I've done or why. There is very little training data to be found for us as persons. IF they hack your phone or computer they might be able to store how you interact with the internet. Which things you browse and read, what you buy, how you write on social media. They could possible then create a digital twin (a clone) of your online presence but it would be a lot of work for very unclear gains. If they hack enough people they could learn some connections like that people interested in electric cars might also be interested in battery technology and solar panels, people who are interested in vegetarian food might also like to read/write about vegan food or the environment and so on. But I still think it would be a LOT of work just to end up creating a fake website selling fake batteries or vegetarian food that you either never deliver or that ends up being a poor replica of the good stuff.

[waihtis]

Well, currently the top LLM fails at repeatedly generating a stable set of pretty simple regexes for me, so where pretty far from a full-blown LLM dystopia at this point.

[sznio]

LLMs are too dumb to be useful, but smart enough to be destructive.

You don't need to be able to generate correct regexes to create a blogspam article about regexes that will go to the top of the search results.

[Planktonne]

> LLMs are too dumb to be useful, but smart enough to be destructive.

I think this puts the blame in the wrong place; LLMs are too dumb to be useful, but it's the users who are too dumb to realise that, or too unethical to care.

[uoaei]

It's also the creators, hosters, and startup founders who are overselling its capabilities.

[Planktonne]

I was lumping them in with 'users'; I don't think the evangelists are any less credulous than their marks.

[waihtis]

search is dying anyway and this will just accelerate that trend.

[mavhc]

The question is does the rapid progress in LLMs continue over the next few years, or do we reach another local maximum. Is "currently" going to improve rapidly?

[waihtis]

as I understand we're approaching local maximum with LLMs and that is what I base the previous comment on, but fully layman and no insight into deeper layers of LLM R&D