OpenAI's Sora model generates its videos by moving a camera through its conception of the world indicated by the prompt. The recent video with the milk being offered to the cat on the throne - when milk wasn't mentioned in the prompt - shows that the model is reasoning as well as any LLM.
Given that the model can start from an image and generate backward to predict what would cause that image, hasn't Sora achieved AGI according to the definition of any reasonable person 20 years ago? Isn't the ability to show the user a tutorial of what to do to get to the desired image general intelligence?
1,000People are also trading
Isn't sora videos? That's not AGI. The G in AGI stands for general. And the I stands for intelligence.
@jim you can have general intelligence in video form. god communicating through a video model could show sign of arbitrary intelligence and have arbitrary capabilities
I don't know if I'm supposed to vote on this or not since I was not around in 2000 to have any opinion on AGI, but I don't see how a model that only generates short videos can possibly be called an AGI in any sense of the word. Unless other people are defining the term AGI in a way so different from how I think about it that we aren't even thinking about close to the same thing, Sora is unambiguously not an AGI.
@Sss19971997 I don't see how ChatGPT could be considered an AGI. It still just does one thing: respond to whatever text inputs the user gives it.
@PlasmaBallin "responding to whatever text inputs the user gives it" counts as agi if done well enough, surely? So you would have to say it responds in a sub-human way to whatever text inputs the user gives it?
@Bayesian When I think of an AGI, I'm usually thinking of something that would be capable of actually doing some real-world project on its own by using multiple different skills. I'm not sure if even a super-advanced chatbot could do that when it's only allowed to respond to what a user says rather than act autonomously. But maybe that is too strong of a definition.
@PlasmaBallin connect its output to physical actuators and if it's arbitrarily powerful it will be able to control a robot and have non-destructive feedback loops and be able to have embedded agency and all the fun stuff, after a small python for-loop that connects its outputs to itself or wtv. in that sense I consider question-answering to be agi-complete; in the same way, a text to video system is agi-complete, because an arbitrarily powerful version of sora can have prompts and answer those prompts by typing out answers on the video feed, and many other things. imo the issue with current systems is that they are not powerful enough, but the medium is a technicality and not fundamentally limiting. ofc it might be that the structure is practically limiting and that to get those arbitrary capabilities you could do it more easily with some other structure we haven't found yet
@PlasmaBallin I am saying if u tell some one in 2000 about ChatGPT, they would consider it as AGI. Because, in their minds, ChatGPT must mean we have solved vision/robotics/self driving/ alpha go etc, because language was considered a much harder cognitive task. They would think AI today can out compete Yann LeCun’s cat
@Bayesian Okay, yeah, I think you're right that that would count as an AGI. But since ChatGPT isn't powerful enough to do anything like that yet, that still makes it clearly not an AGI, right?
@Sss19971997 There might have been people in 2000 who defined AGI in such a way that ChatGPT counts as one. I suspect that if you time-traveled back then and showed them ChatGPT, they would be more likely to revise their definition of AGI than say, "Yeah, that's one."
Sora is just a probabilistic machine; it takes your input and creates an output based on millions of invisible variables, each with certain degrees of randomness, based on its training data.
Milk being offered to the cat on the throne doesn't demonstrate any kind of intelligence. I think it's just a really complicated random number generator good for spitting out low-quality videos and nothing else. I think me 20 years ago would have still felt the same.
@thepurplebull Yeah but isn't the argument against this stochastic parrot view that humans might fundamentally be no different in how they generate content, even if it doesn't seem this way to us? One of my bigger problem with ascribing the AGI crown is that these systems have absolutely no agency, and the me of 2000 would have assumed agency an essential element in a truly intelligent system (you can argue with 2000-me on if agency is a valid component to require of an AGI, but you're not going to change 2000-me's mind!).