Meta's chief AI researcher says OpenAI's "world simulator" Sora is a dead end | Hacker News Reader

Meta's chief AI researcher says OpenAI's "world simulator" Sora is a dead end

the-decoder.com

3 pointssh_tomer12 comments

Wouldn’t it be easier to do what Sora might want to do by training models inside a physics engine? For example, training a model exclusively to generate videos from Unreal Engine 5. Based on predefined physics rules in the engine, it would be impossible to generate something that breaks the rules.

To model physics and not just a playground pretend version of it, you need to model the complete understanding of physics within the entire universe. We'd need answers to questions we're UNKNOWN-TIME away from understanding. This means modelling everything from the visual response of a human eye to the known-truths of the universe and beyond.

We can't even trust our own nose to smell bullshit but we're supposed to have a model for it.

Please think critically folks!

A model is by definition "a playground pretend version". We can't know the real physics, it is a philosophical question if they even exist, that doesn't mean you can't make an approximate version, ie simulation

We agree. But remember, approximations here are nothing but lies masked as convenience.

Sora is not even a “world simulator”. It just creates one plausible scene based on a textual description. I hates this kind of click-baity headlines.

When one focuses on the competition, one is not the leader.

His team did release a separate approach which he claims is more effective, so it’s not like he’s just dissing OpenAI without a rationale. In time we’ll get to see which (if any) approach was closer to correct.

Does OpenAI/SamA diss other companies’ work? I can think of examples.

When you’re the leader and secure in your position and you have the emotional maturity, you don’t feel the need to diss otgers explicitly. It is a sign of weakness.

*ALL* current public proof of concepts for AI are a dead end. They are strictly past information ingest machines. Until 'Generative' solution thinking finally stops chasing the tail of history and believing that what it has ingested from then to now is just as simple as adding more 9's... and the solution will manifest -- is ignoring that without the ability to also quantify stimuli in every definition - of the 'present' - it's nothing but a well dressed toy. We can't even define 'present' and yet we think a farm of GPU's will magically solve singularity. If anything, this is infinitely short sighted.

Any read-only AI model or entity is a dead end. If the system cannot modify itself to adopt improvements (self-provided, externally provided), it can't go beyond its initial training, "per-se".

Unless, some kind of ad-hoc architecture permits the AI to grow up, which could be a project or two already, behind the information firewall of the research teams. Not going here to "caution, AGI explosion possible", but more a grounded "some could guess our trick and stole the market before us".

For the singularity thing, too many people worried about, and most probably we will hit singularity a lot of time before we actually acknowledge that we have done. Not AGI stuff again, remember that extremely advanced (AI) technology is more than enough to radically change society, not AGI required.

You haven't once ingested anything but past information. Unless you're clairvoyant?

Oh, I'm sorry, did I argue my omnipotence? Or the fallacy of marketing? You shifted the goalposts pretty far.

Our existence in the 'now' means we're reactive to the now and the universe of events influence us as much if not more than our concept of free will. I don't think to make my skin react to cold air. If these models can't ingest, quantify or even understand/intuit current stimuli, they're a dead end solution and the problems are bigger, harder and still being ignored during the cash grab era.

I haven't seen or heard a single word of progress on that front. Any sources?

Reply on news.ycombinator.com