/vt/ - Virtual Youtubers

>>87953000


>If you drop a Roomba in an alien's living room it would learn the floor plan just fine too.
It can't do it 'just like that', though, which is one of the practical differences. Its navigation is dependent on exhaustive permutation of cases to build a framework. You can't boostrap a dog with existing approaches without a profusion of preexisting 'dog' to finetune on.
The question though is, are all use-cases equally incommensurate? Even supposing one needs examples of prior art to make a convolutional network that simulates it, is there no such thing as more or less general cases, more or less essential things, more transcendent or more parochial teleologies? That ability to achieve one particular function would not carry knock-on implications that allow for achievement of many other functions that are downstream of it? I would say say the answer is no, of course, so...

>Either you think it can only answer questions it has seen in its training data, or you have to admit that it is learning some abstraction of that data then deriving answers by building a train of thought from that learned abstraction.
The fact that LLMs have functional verisimilitude at all is a fairly definitive devalidation of nominalist epistemologies.
Never the less, while predictions or models of what text comes after other text has potential for extremely powerful and useful search and organization, it's not very good at oracular functions.
To paraphrase another conversation on the subject, trivial combinations of stuff that is in the databases, like an astronaut riding a horse, is not utility that goes far beyond database retrieval on steroids. Automatic summary that aggregates information from diverse sources to draw valid conclusions not in those sources doesn't exist yet. When you ask bing or chatgpt about things, you can often read its answers word-for-word from the sources it is using. It would be more useful if it serves those sources up front without people pretending it isn't doing what it is doing.

>A single AI (called MuZero) was able to learn by itself to play Chess, Go, Shogi, and dozens of different Atari games. In many cases it learnt the games better than the average human, or even better than every human. Most current LLMs are not trained on 3D data so they have no experience of using limbs, but they can be trained to use tools like web search or calculators
While I would quibble that this is not particularly cross-domain (that would be something more like going from playing Go to guiding a missile into a target to designing a new car engine), people should be thinking more about this - about circumstances that lead to explosions in possibility space, and purposefully seeking to create them.
Existing LLMs are bad at oracular functions. You ask it what moves in pokemon you should use, you'll have a bad time (https://www.youtube.com/watch?v=IGpIZhT4Tw8).). You ask it who in the war in ukraine is getting attrited more, you will receive "some random facts mixed with a pile of rote official bullshit that verged on incoherence. Because its database contains mutually conflicting stories, it produced an answer that was profoundly unhelpful and uninformative. Because the data in its database did not add up, produced an answer that did not add up. It would have been appropriate to say “some say this, some say that, and some say the other”, but it reproduced many conflicting voices as a single voice. Much like putting an astronaut on a cowboy horse. It retrieves the data, and integrates diverse data into a response that has normal flow of syntax, but not normal flow of reason."
Can you imagine an oracular model made for directing business strategy, predicting world events, or organizing a war effort? Most if not all would likely not be usefully functional in such a capacity especially is used naively; but those are exactly the sort of deep cuts dependent on world-formation that people *should* be thinking about. I would be very interested in seeing what a model that *can* do those things would look like.

>and that could be extended to arbitrary software functions if people were stupid enough to give it that much control.
Which is exactly what is needed, though. How Nature or Nature's God themselves got interesting results with life: stochasticity and recursion.
Of course substrate determines what sorts of things can be done with the substrate, and animal-like phenomena are naturally a function of animal-like substrates. Never the less, when speaking of boolean substrates, and the things we can do with them, these tools are functions of their programming; therefore, an obvious path to getting interesting results with these substrates is programs that can make programs that fit to purpose; programs that can self-program to fit to purpose. Create the circumstances for explosions.
Models for all sorts of objectives, useful. Things that can 'see' objectives, that can give themselves objectives, that's explosion.

>>87602881
I did and honestly it was shocking that one small AI could be so racist

>>83199748

>>82728907
what do you think?

>>79270718

>>77326361

>>75034470
explains why he programmed neuro to be racist

>>50527608

>>50263980
As was I, she hasn't been based in months

>>47006512