Post

i think tech and ai positive people are going to be in for a big shock when anti tech and anti ai sentiment becomes a major part of leftwing politics going forward especially as datacenters continue to destroy communities and raise electricity bills

Same as it ever was, tbh Not long ago Wikipedia was the leftwing devil over Assault Rifle/Armalite Rifle brigades If you're not behind solar and wind energy driven around by little AI/AI adjacent tech then you aren't actually left. Musk and Sama ruined stuff, but now they're gas/coal geezers

Given enough time we'll have open source and home made small language models that (yes, accurately, yes, without hallucinating) teach crochet and organic farming. It's far too useful-yet-horribly-boring to ignore, just like blogs, patreon, wordpress, wikis, reddit, blah blah blah

No, we will not. The problem of hallucinations is inherent and unsolvable - because LLMs do not think or reason or understand or know, they make statistical predictions of the most satisfying response, and there is no point at which truth is distinguished from falsehood in that process.

Yes, you will. The problem of hallucinations is inherent and unsolvable - and, in fact, a feature of LLMs. But that does not mean you can't get deterministic answers. If you ask for a "double puff" a specifically trained crochet model is not going to go ChatGPT on you and talk about birds.

I believe that is called "reading a book" and the tech is quite mature

I appreciate this answer :) But books don't have ctrl+f or search boxes and about the time you've screwed up hours of work because you don't know what a certain stitch is... I mean at least vector search is useful. :)

Ebooks, which predate LLMs by many years and don't require them, literally do have ctrl+f and search boxes. So the case is made here for OCR, which also existed before LLMs, and half-decent search, which existed before LLMs, and which definitionally shows its sources.

They do! But not semantic search. And OCR using an LLM now is vastly superior than any of the other approaches. So you're still using one in a roundabout way.

If I ask an LLM, what I'm going to have to do is, I'm going to have to ask it to actually cite its sources. Then, I will have to wrestle with it to provide sources, which it MIGHT do if it's a very classy model, and those sources will tell me, after I look, whether or not it gave me a line of shit.

So the LLM on its own, not as a statistical imporvement to the OCR (which, because the OCR only has to run once to scan the document, is far less resource-intensive than running a bunch of sequential queries) but on its own, will approximately be half as useful to me as google 10 years ago.

Asking an LLM to cite sources generally goes sideways. But you might enjoy the Olmo models and the ability to trace exact responses back to training data.

That's the bare minimum and yet it's being pitched to me as if restoring them to the functionality of a search engine is the wave of the future. If the LLM is just going to give me the same info on organic farming I would get from looking into a resource, I should just use the resource.

That's fair and very true. But this is a very boring technology that has quite a lot of advantages for things like 🥁 offline and accessible versions of the same information.

It's clear you view LLMs as a key part of a solarpunk future, to put it one way. But in the interim between here and there, LLMs as they currently are now, and not as they are in fantasy, requires me to deep fry the last surviving devil's hole pupfish in order to get one page of a google search.

See, this simply isn't true. You can use granite4 on 2G of RAM completely offline. It won't tell you to kill yourself. You don't have to kill wildlife. One of the earliest LLM projects I did was reorienting solar panels :) It's not fantasy. It's at least a decade old.

Yes yes and you can use gzip for text classification, "AI" has been around since the 60s, but here we've hit a circle: The knowledge required to do this requires knowledge of the subject enough to feed it the right data, and anyone using it will need to confirm it with the source material anyway.

Right - so eventually content creators may publish their own LLMs (or small language models, whatever you want to classify them as) that are complete tutorials that can be used offline and in a number of different languages with accessibility options from braille to audio in < 2G RAM. :)

You seem to already understand that this conversation has veered far from what anyone using the term "LLM" in common parlance is even talking about

If 9 people are having a conversation about Hiroshima and you chime in with "what about medical imaging," and those 9 people give the tenth a :| face, would you say "well it doesn't matter because it's a conversation about nuclear stuff"

9 people aren't talking about Hiroshima here. 2 people tried to tell me AlphaFold 3 is a type of AI they made up. If 9 people are discussing Hiroshima and someone says that bomb was never dropped and it was actually the Trident of Satan while daring people to prove them wrong... that's this thread.

I fear we're losing you to the sauce of my example here. Gratefully, you already pretty well agreed with my point two posts ago, so it's neither here nor there