Post

that seems not true? it's certainly capable of 'threatening' the liberal arts (just to mirror that language). super cool tho.

I was imprecise, this is better

But the point is still strong: we can build systems on this that are no “threat” in the sense I mean, in a way that we have heretofore struggled to do with LLMs for various reasons.

it sure feels like it could hypothetically do a whole lot of wacky shit to grading, admissions screening, editorial selection, moderation, archival classification, document interpretation, semantic search, curation, peer-review triage, rubric assessment, etc etc

actuarials in trouble :)

In other words, the entire *scaffolding* of the liberal arts, and none of the arts themselves. 🙂‍↕️

i think they need to talk more shop because I have no idea what they're actually doing different from an architecture standpoint for this to "not be an LLM" in a meaningful way not that it's impossible just that i don't understand how currently

This will make a lot of blue teamers very happy. We've had so many folks getting burned out by alert fatigue/IR/etc., having the robots running in multiple different directions gathering data and firing off runbooks before you even open your laptop is gonna really feel like the future.

I don't quite understand yet. it seems like this takes in a bounded set of options and just makes an educated decision, but like... so can anyone? we all do the same thing. every single person would come up with a different (decision, reason) tuple given the same inputs, and that's a good thing.

is the utility of this model just that it does it in ~100ms? or is it supposedly able to do it more effectively in every domain than like 95% of experts in those domains, or something (and while not 100%, 95% is damn good, which makes it have generalizable knowledge)?

Yes, people all do the same thing. That’s kind of the point of AI, to make computers do things that humans do automatically. :3 The utility from a technical perspective is that it scores about as high across domains as the best LLMs, while being much cheaper and faster.

I have Alignment Concerns

More than you had before? :3

> Not an LLM > immediate off the charts results LeCun you beautiful bastard you strike again

Would be a compliment to LLMs if its all they say it is, but not a replacement

As far as AI knowers are concerned, you’re correct. As far as policy is concerned, this provides a way to “put AI into everything” that doesn’t tend to collapse into “replace human-generated content with AI-generated content”. This is a political opportunity (even) more than a technical one.

They can pry the audio generators out of its cold dead hands :p

Ethan Mollick linked a study that showed most commercial LLM use is for decision support (That was before last November though, so the coding revolution might have changed it)

or structured semantic generation is yet another threat

i’m slowly wrapping my head around the implications, and they are vast

This is really going to throw a wrench into "stochastic parrot theory". How can it be a stochastic parrot if all it does is an elaborate set of mathematical equations, producing a statistical probability, with no words involved except the ones we feed in?

Do you mean Bender and Gebru’s position on how LLMs interact with meaning, or do you mean the phrase as used dismissively by anti-AI posters? This doesn’t threaten the former, and the latter is hardly a “theory”.

Yeah, “I can’t quite see how this fits in my workflow” is sometimes followed by “oh, oh shit”

This is an economically viable off-ramp to the LLM arms race.

I'm still trying to grok what all this means but like, this is huge for feature extraction no?

I say this as a fool who barely knows what they're talking about, but I have always envisioned a very powerful shift away from "text generation is used for literally everything and we leave it at that"

I honestly just didn’t think we’d get it before society collapsed lol. It was always a possibility, but line go up fast!

I still fondly remember the early days of the neural network boom at the turn of this decade, and for a while, voice models and even image models seemed to steal the thunder of gimmicky autocomplete and even more gimmicky chatbot applications... I want this world to come roaring back lol

What if I asked the structured output of the model to be textual token strings What then

You have to enumerate the choices in advance :3

only multiple-choice? that's a bit less useful than i imagined i was thinking it did constrained generation, so you could point the harness a syntax and it would generate that but yea "write a novel as xhtml" would then work, too

the enumerated choices are: every possible token one token per turn yo dawg i heard you like etc etc

isn't this more like, at least on paper, what typesafe is itself describing it as - a "frontier-intelligence function call" deliberately restricted to narrow judgments or specifically bounded semantic decisions?

Though this was an amusing point by a robot I asked:

So basically a pure classifier on an LLM backbone? As a scientist I would find it useful for plugging into "crystallized" workflows, removing need to sanity check or parse, while also appreciating generative outputs for other purposes. My main question would be if it can assess confidence?