The A.I. Thread

[Russic]

One thing that Sora is doing really well is voice matching... I'm sure copyright stuff will get wayyyy tighter in the coming weeks as IP holders lose their collective minds, but for now not only can you use a lot of likenesses, but the voices are perfect.

https://sora.chatgpt.com/p/s_68dd940...6a54e73c597952

[chemgear]

If you have a VPN (UK) this came out recently. Interesting, especially for the last minute.

[DownInFlames]

How long until AI wipes out the human race? I give it five years.

[Bill Bumface]

Originally Posted by DownInFlames

How long until AI wipes out the human race? I give it five years.

We're doing just fine doing this on our own, thanks!

[White Out 403]

I reconnected with an old friend and we talked about AI. He's super onboard and told me all about his use of agents and how he's basically a programming super hero now with this tool and he's doing things that he never thought possible. Later on the call after we had some space from it I asked how his business was going. "Not good".

I think there's this real, chasm like gap that exists between how "neat" Ai's can be and how useful. So many people want to believe with all their heart we are just a couple years from AGI when it seems like we're closer to LLM's becoming worse than achieving anything approaching AGI. We're already seeing the shifting of expectations and now people reframing what AGI is.

Understanding context isn't something you can program. There's a real squishy part of our meat brains that is programmed to handle this stuff.

[indes]

Originally Posted by White Out 403

I reconnected with an old friend and we talked about AI. He's super onboard and told me all about his use of agents and how he's basically a programming super hero now with this tool and he's doing things that he never thought possible. Later on the call after we had some space from it I asked how his business was going. "Not good".



I think there's this real, chasm like gap that exists between how "neat" Ai's can be and how useful. So many people want to believe with all their heart we are just a couple years from AGI when it seems like we're closer to LLM's becoming worse than achieving anything approaching AGI. We're already seeing the shifting of expectations and now people reframing what AGI is.



Understanding context isn't something you can program. There's a real squishy part of our meat brains that is programmed to handle this stuff.

I couldn't agree more. "Neat" is exactly the word I would use to describe AI. It does a lot of cool stuff but it never seems to do exactly what I try to do with it. As a search engine (LLM) I think it's pretty damn good. As for actually doing things properly all by itself? It seems a ways off to me. It feels like all the mindblowing AI stuff I see has had a significant amount of human intervention, which is still cool, but not AGI cool.

[Whynotnow]

I think you are seeing agentic ai use cases popping up where’s there’s pretty well defined rules and parameters for how things work and risk of it being wrong pretty low. So this is in some pretty traditionally low tech jobs with low decision and authorities, so think customer service.

It’s also good at assessing multiple data streams(if you have good data) and helping provide summaries to people.

But I’ve yet to see it change an industry completely, it’s an efficiency play is what I’m seeing now. Much is being over promised by all sorts of consultants who only have one goal, to get you to hire them. The more independent firms are taking a more cautious view of things.

[Firebot]

It won't stop companies from trying their hardest to make AI happen with zero understanding of its limitations and spending millions into a black hole.

https://petapixel.com/2025/09/23/mov...s-with-runway/

https://runwayml.com/news/runway-par...with-lionsgate

https://www.thewrap.com/lionsgate-ru...odel-concerns/

Lionsgate was trying to build a fully AI generated movie using its movie library as the training data partnering with Runway (who in my opinion is now one of the worst and lagging significantly behind). The problem is their library is miniscule in terms of data compared to the best video generation models, including Veo 3 which uses Youtube as its training data.

Ironically Disney also tried it, which was a disaster in results and in figuring out how to copyright AI generated work and intellectual property (quite ironic). Over promising on what is possible.

https://futurism.com/disney-secret-ai-disaster

https://metaphysic.ai/our-story/

Don't get me wrong, I do think we will see fully AI generated movies in the near future, but companies are rushing to can their whole production teams before they have a working solution with companies over promising on what is effectively AI snake oil. Heck some of these projects likely started at the first Will Smith eating spaghetti video era.

[Russic]

Originally Posted by White Out 403

I reconnected with an old friend and we talked about AI. He's super onboard and told me all about his use of agents and how he's basically a programming super hero now with this tool and he's doing things that he never thought possible. Later on the call after we had some space from it I asked how his business was going. "Not good".

I think there's this real, chasm like gap that exists between how "neat" Ai's can be and how useful. So many people want to believe with all their heart we are just a couple years from AGI when it seems like we're closer to LLM's becoming worse than achieving anything approaching AGI. We're already seeing the shifting of expectations and now people reframing what AGI is.

Understanding context isn't something you can program. There's a real squishy part of our meat brains that is programmed to handle this stuff.

A few things going on here, I think. I always like the quote "people will believe with all their heart what they desperately require to be true." To a business person (speaking as a former one), the idea that AI will waltz up any day now and deliver us from drudgery is a potent drug.

I make no secret that of the AI super-fans in my life, a curiously high percentage of them are in that "I work smart, not hard" camp that seems to rarely get anything done.

Now, the few that do not fall into that camp are in a very interesting category that shall inherit the earth. They're doing some stuff. They're just a very rare breed.

And it's true, as of today, the agents and whatnot aren't delivering what they hype-goblins are screaming about. Thing is, agent tools are what... months old? It's a bit odd to say there's nothing here when things are that new.

My take on it is they won't be taking over tons of jobs yet, but make a few strides on a singular issue (memory comes to mind), and things get interesting/terrifying. I can't help but feel like the leap required between what we have today and scary functional is shorter than many assume.

As for the "when will it kill us all" question... if we can make it 15 years, I think we're golden. It's all utopia from there forward. Gotta not kill ourselves with it first, though.

[peter12]

Originally Posted by indes

I couldn't agree more. "Neat" is exactly the word I would use to describe AI. It does a lot of cool stuff but it never seems to do exactly what I try to do with it. As a search engine (LLM) I think it's pretty damn good. As for actually doing things properly all by itself? It seems a ways off to me. It feels like all the mindblowing AI stuff I see has had a significant amount of human intervention, which is still cool, but not AGI cool.

World’s most overrated parlour trick

[chemgear]

https://twitter.com/user/status/1983846299586683236

[peter12]

Why are we inventing technology, which to this point, only really seems to directly benefit criminals?

[PaperBagger'14]

Originally Posted by chemgear

https://twitter.com/user/status/1983846299586683236

[Fuzz]

Because the criminals are in charge now?

[Shazam]

Bought an Anthropic sub.

Works okay. Still writes sub-par code. If you're a junior dev it'll be wrong code 50% of the time. Oh sure, it'll compile and run, but it'll write you, say, synchronous code instead of async code.

But it excels at doing stuff like "take this input and output it into this json format". You need to be very specific but it gets the job done cheap.

Also Opus uses an insane amount of tokens. It's not cost effective quite yet.

[TorqueDog]

Originally Posted by chemgear

https://twitter.com/user/status/1983846299586683236

Waving your hand quickly in front of your face usually upsets the whole thing, which I happen to notice they didn't do in this demonstration.

It's one of the things people have been instructed to request when interviewing candidates remotely -- get the candidate to wave their hand in front of their face, and watch for any anomalies or artifacts that suggest there's a filter being used.

[Fuzz]

That's good for now, but I suspect it's like the AI art demented finger thing. It'll be resolved soon enough. Unfortunately technical glitches are never going to be a long term solution to this.

[TorqueDog]

Yes, but the fact they didn't demo overcoming that flaw suggests to me that it might still be present in their current version, given how widely known it is.