Really solid write-up Key issue: "...it becomes clear that we are heading towards a place where not only are image and video generations likely to be good enough to fool most people, but that those capabilities will be widely available and, thanks to open models, very hard to regulate or control." - ThreadSky

gregdoucette.bsky.social • 3 days ago

Really solid write-up

Key issue: "...it becomes clear that we are heading towards a place where not only are image and video generations likely to be good enough to fool most people, but that those capabilities will be widely available and, thanks to open models, very hard to regulate or control."

Reposted from Ethan Mollick

I wrote a history of recent AI development in 32 images of otters using wifi on airplanes, from images to video to code.

It shows two big trends: rapid improvements in AI models of all types and the growth of open weights AI models. www.oneusefulthing.org/p/the-recent...

Comments

shortandslanted.bsky.social•3 days ago

I remember Hank Green talking about this some years back.

What happens when the cost of manufacturing bullshit becomes nothing or nearly nothing?

javamanphil.bsky.social•3 days ago

But the cost isn’t nothing. It’s billions of dollars of energy, water, and acreage. If nobody is willing to pay for this, then what!?

shortandslanted.bsky.social•2 days ago

You're right, it's not nothing but to the individual consumer at the moment of use, that cost is not obvious.
By the time enough people become aware of the damage AI has done, both in environmental damage and mis/disinformation, it's likely to be too late.

javamanphil.bsky.social•3 days ago

Energy, land, and water aren’t free. And no matter the output quality, eventually there has to be someone willing to pay for this shit to make it viable.

gregdoucette.bsky.social•3 days ago

That was the point of the column's discussion in open weights & open source models -- the tech is already able to run on your own personal laptop just using a single GPU

javamanphil.bsky.social•2 days ago

I read it. It still has to be paid for, even on a standalone machine. The costs still exist, even if private equity has funded the development so far. Eventually users will be charged: monthly subscriptions, enterprise licenses, ads. And then, does it matter how good it is? Will people pay?

connorlynch.bsky.social•3 days ago

seems weird that they can make photorealistic otter-on-plane videos but can't spit out a decent legal brief that's been properly cite checked. I suspect that will change, though I certainly won't be gambling my bar license to be an early adopter.

proudanselmo.bsky.social•3 days ago

It's multi-factorial, but it basically boils down to the focus has not been on the legal field. The number of legal datasets are miniscule in comparison to almost every other type, many are out of date, and they are seldom downloaded.

cjones47.bsky.social•3 days ago

That bit's kind of puzzling. Gemini does a pretty damn good job of showing its work now (if you want).

I'm still a big skeptic relative to hype, but we have a something machine here (nowhere close to AGI).

Specialized models with curated data are a whole other beast.

gregdoucette.bsky.social•3 days ago

I suspect it will get there sooner than later, at least as far as citing real cases with real cites goes (even if the propositions they're cited for are wrong). Something like a Mixture of Experts setup coupled with using multiple agents, one to pull the case and what to confirm it's real

gregdoucette.bsky.social•3 days ago

The Wexis folks have been testing out a blend of off-the-shelf LLMs coupled with more traditional algorithms, and results have been described as "not total dogshit". That's a big change from just a year ago

gregdoucette.bsky.social•3 days ago

I will likely be a curmudgeon about it until I die and never once use it for a brief though 🙃

proudanselmo.bsky.social•3 days ago

I think people like Altman successfully sold a lot of people on the idea that AI would be able to do all things for all people and just as it was with every salesman who came before him who made it, that promise is laughably bullshit.

proudanselmo.bsky.social•3 days ago

AI will likely be able to replace humans for much of the repetitive work we do, but contrary to claims, it cannot replace human creativity. So AI will at some point probably take over drafting the Book of Authorities (or whatever the US equivalent is), but, unless you do slip and falls, it probably

connorlynch.bsky.social•3 days ago

Yeah, it does seem like it should be a solvable problem. If it actually meaningfully makes brief writing faster (and cheaper) by an order of magnitude or more, I'm not sure it will be viable to avoid it if everyone adopts it. it doesn't do that now, but it seems like it might within 10 years.

connorlynch.bsky.social•3 days ago

(And if it could, I do not think it would result in less legal work. I think there would be more, as markets that are currently priced out open up, or more disputes that now settle because the cost of litigating is high end up going forward.)

parserexception.bsky.social•3 days ago

Have they been managing proper citations? Because my understanding of how LLMs work would indicate that citations wouldn't work without massive over fitting or a new model design break through.

gregdoucette.bsky.social•3 days ago

"Proper" is the issue 😄 The past 3-4 months or so, they now routinely cite to real pages in real cases -- but the thing cited still ends up being wrong (not relevant, or opposite of what's needed, etc)

My assumption is everyone's doing MoE now and there's some behind-the-curtain verification

parserexception.bsky.social•3 days ago

It annoys me that if there had been more focus on making reference tools instead of making better Markov chains we could have tools that take in natural language questions and generate references to sets of documentation instead. So much knowledge is in formats resistant to normal searches e.g. law

satisfiedpsychic.bsky.social•3 days ago

I think we hold text to a higher standard and if we were to look at video generation from a professional perspective (lighting, angles, spacing, etc) we would be harsher on it.

satisfiedpsychic.bsky.social•3 days ago

If AI generates a video it could take a specialist to identify it. If AI hallucinated a case citation that doesn't exist that's something that you could casually catch on a basic review. It's harder to fake it in text, somewhat surprisingly.

Comments

Posting Rules

Reply