Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

AI is different

antirez.com

Post reply on HN
Print view View original

871–880 of 920 posts

Re: AI is different

#871

Post by danny_codes »

danielbln wrote: post #478
jhbadger wrote: post #425

Earlier quoted context omitted.

>You can already see this with YouTube: AI-generated videos are a mild amusement, not a replacement for video creators, because made by AI is becoming a negative label in a world where the presence of AI video is widely known. But that's because, at present, AI generated video isn't very good. Consider the history of CGI. In the 1990s and early 2000s, it was common to complain about how the move away from practical s…

Ironically, while the non-CGI SFX in e.g. Interstellar looked amazing, that sad fizzle of a practical explosion in Oppenheimer did not do the real thing justice and would've been better served by proper CGI VFX.

Totally agree. Nolan is a perfectionists though, so I don’t think he could let himself go for broke on the actual boom boom

Re: AI is different

#872

Post by sneak »

ben_w wrote: post #864
oinfoalgo wrote: post #848

Earlier quoted context omitted.

The bigger issue to me is that not all geography is anything close to equal. I would much rather live on a beach front property than where I live right now. I don't because the cost trade off is too high. To bring the real estate market into equilibrium with UBI you would have to turn rural Nebraska into a giant slab city like ghetto. Or every mid sized city would have a slab city ghetto an hour outside the city. It…

> I would much rather live on a beach front property than where I live right now. I don't because the cost trade off is too high. Yes, and? My reference example was two aircraft carriers and 1:1 models of some fictional spacecraft larger than some islands, as personal private residences. > To bring the real estate market into equilibrium with UBI you would have to turn rural Nebraska into a giant slab city like ghett…

Some people’s idea of wealth is to live in high density with others.

You bump up against the limits of physics, not economics.

If every place has the population density of Wyoming, real wealth will be the ability to live in real cities. That’s much like what we have now.

Re: AI is different

#873

Post by xpe »

amanaplanacanal wrote: post #858
xpe wrote: post #844

Earlier quoted context omitted.

Let’s get specific. If we’re talking about DARPA’s research agenda or the US military’s priorities, I would say they are quite capable at planning for speculative scenarios and long-term effects - for various reasons, including decision making structure and funding. If we’re talking about shifting people’s mindsets about AI risks and building a movement, the time is now. Luckily we’ve got foundations to build on. We…

It's hard. Take global warming as an example: this is a real thing that's happening. We have measurements of CO2 concentrations and global temperatures. Most people accept that this is a real thing. And still getting anybody to do anything about it is nearly impossible. Now you have a hypothetical risk of something that may happen sometime in the distant future, but may not. I don't see how you would be able to get a…

> And still getting anybody to do anything about it is nearly impossible.

Why exaggerate like this? Significant actions have been taken.

> I don't see how you would be able to get anybody to care about that.

Why exaggerate like this? Many people care.

Re: AI is different

#874

Post by int_19h »

Eggpants wrote: post #661
vlovich123 wrote: post #646

Earlier quoted context omitted.

How is that different than the models today are actually usable for non trivial things and more capable than yesterdays and it’s also true that tomorrow’s models will also probably be more capable than today’s? For example, I dismissed AI three years ago because it couldn’t do anything I needed it to. Today I use it for certain things and it’s not quite capable of other things. Tomorrow it might be capable of a lot m…

It’s not that the LLMs are better, it’s the internal tools/functions being called that do the actual work are better. They didn’t spend millions to retrain a model to statistically output the number of r’s in strawberry, but just offloaded that trivial question to a function call. So I would say the overall service provided is better than it was, thanks to functions being built based on user queries, but not the actu…

This is a good example of making statements that are clearly not based in fact. Anyone who works with those models knows full well what a massive gap there is between e.g. GPT 3.5 and Opus 4.1 that has nothing to do with the ability to use tools.

Re: AI is different

#875

Post by int_19h »

pm wrote: post #778
libraryofbabel wrote: post #755

Earlier quoted context omitted.

What is your definition of "stochastic parrot"? Mine is something along the lines of "produces probabilistic completions of language/tokens without having any meaningful internal representation of the concepts underlying the language/tokens." Early LLMs were like that. That's not what they are now. An LLM got Gold on the Mathematical Olympiad - very difficult math problems that it hadn't seen in advance. You don't do…

Ignoring its negative connotation, it's more likely to be a highly advanced "stochastic parrot". > "You don't do that without some kind of working internal model of mathematics." This is speculation at best. Models are black boxes, even to those who make them. We can't discern a "meaningful internal representation" in a model, anymore than a human brain. > "There is just no way you can get to the right answer by spou…

> This is speculation at best. Models are black boxes, even to those who make them. We can't discern a "meaningful internal representation" in a model, anymore than a human brain.

They are not pure black boxes. They are too complex to decipher, but it doesn't mean we can't look at activations and get some very high level idea of what is going on.

For world models specifically, the paper that first demonstrated that LLM has some kind of a world model corresponding to the task it is trained on came out in 2023: https://www.neelnanda.io/mechanistic-interpretability/othell.... Now you might argue that this doesn't prove anything about generic LLMs, and that is true. But I would argue that, given this result, and given what LLMs are capable of doing, assuming that they have some kind of world model (even if it's drastically simplified and even outright wrong around the edges) should be the default at this point, and people arguing that they definitely don't have anything like that should present concrete evidence ot that effect.

> We're just highly advanced "stochastic parrots" with a game loop.

If that is your assertion, then what's the point of even talking about "stochastic parrots" at all? By this definition, _everything_ is that, so it ceases to be a meaningful distinction.

Re: AI is different

#876

Post by int_19h »

dmead wrote: post #680
gdubs wrote: post #622

AI has been improving at a very rapid pace, which means that a lot of people have really outdated priors. I see this all the time online where people are dismissive about AI in a way that suggests it's been a while since they last checked-in on the capabilities of models. They wrote off the coding ability of ChatGPT on version 3.5, for instance, and have missed all the advancements that have happened since. Or they t…

Is there anything you can tell me that will help me drop the nagging feeling that gradient descent trained models will just never be good? I understand all of what you said, but I can't get over that fact that the term AI is being used for these architectures. It seems like the industry is just trying to do a cool parlor trick in convincing the masses this is somehow AI from science fiction. Maybe I'm being overly cy…

> It seems like the industry is just trying to do a cool parlor trick in convincing the masses this is somehow AI from science fiction.

If you gave a random sci-fi writer from 1960s access to Claude, I'm fairly sure they wouldn't have any doubts over whether it is AI or not. They might argue about philosophical matters like whether it has a "soul" etc (there's plenty of that in sci-fi), but that is a separate debate.

Re: AI is different

#877

Post by int_19h »

l33tbro wrote: post #783
kelseyfrog wrote: post #777

Earlier quoted context omitted.

What are you talking about? This is common knowledge. Median forecasts indicated a 50% probability of AI systems being capable of automating 90% of current human tasks in 25 years and 99% of current human tasks in 50 years[1] The scope of work replaceable by embodied AGI and the speed of AGI saturation of vastly under estimated. The bottle necks are production of a replacement workforce, not retraining human laborers…

Work is central to identity. It may seem like it is merely toil. You may even have a meaningless corporate job or be indentured. But work is the primary social mechanism that distributes status amongst communities. A world of 99 percent of jobs being done by AGI (which there remains no convincing grounds for how this tech would ever be achieved) feels ungrounded in the reality of human experience. Dignity, rank, purp…

Have you considered that perhaps tying dignity and status to work is a major flaw in our social arrangements, and AI (that would actually be good enough to replace humans) is the ultimate fix?

If AI doing everything means that we'll finally have a truly egalitarian society where everyone is equal in dignity and rank, I'd say the faster we get there, the better.

Re: AI is different

#878

Post by int_19h »

cobbzilla wrote: post #710
chuckadams wrote: post #578

Earlier quoted context omitted.

What is creative or original thought? You are not the first person to say this after all.

Not being 100% algorithmically or mathematically derived is a good start. I’m certain there’s more but to me this is a minimum bar.

If your brain is not running algorithms (which are ultimately just math regardless of the compute substrate), how do you imagine it working then, aside from religious woo like "souls"?

Re: AI is different

#879

Post by int_19h »

wybiral wrote: post #575
zarzavat wrote: post #494

Earlier quoted context omitted.

The assumption here is that LLMs will never pass the Turing test for copywriting, i.e. AI writing will always be distinguishable from human writing. Given that models that produce intelligible writing didn't exist a few years ago, that's a very bold assumption.

AI will probably pass that test. But art is about experience and communicating more subtle things that we humans experience. AI will not be out in society being a person and gaining experience to train on. So if we're not writing it somewhere for it to regurgitate... It will always feel lacking in the subtlety of a real human writer. It depends on us creating content with context in order to mimic someone that can cr…

> AI will not be out in society being a person and gaining experience to train on.

It can absolutely do that, even today - you could update the weights after every interaction. The only reason why we don't do it is because it's insanely computationally expensive.

Re: AI is different

#880

Post by int_19h »

keiferski wrote: post #474
antirez wrote: post #465

Earlier quoted context omitted.

To understand why this is too optimistic, you have to look at things where AI is already almost human-level. Translations are more and more done exclusively with AI or with a massive AI help (with the effect of destroying many jobs anyway) at this point. Now ebook reading is switching to AI. Book and music album covers are often done with AI (even if this is most of the times NOT advertised), and so forth. If AI prog…

I can buy the idea that simple specific tasks like translation will be dramatically cut down by AI. But even then – any serious legal situation (like a contract) is going to want a human in the loop to verify that the translation is actually correct. This will require actual translator skills. AI art seems to basically only be viable when it can’t be identified as AI art. Which might not matter if the intention is to…

It's kinda hilarious to see "simple task ... like translation". If you are familiar with the history of the field, or if you remember what automated translation looked like even just 15 years ago, it should be obvious that it's not simple at all.

If it were simple, we wouldn't need neural nets for it - we'd just code the algorithm directly. Or, at least, we'd be able to explain exactly how they work by looking at the weights. But now that we have our Babelfish, we still don't know how it really works in details. This is ipso facto evidence that the task is very much not simple.

Post reply on HN