As I often mention when this subject pops up: while the current statistics-based generative models might see some application, I believe that they’ll be eventually replaced by better models that are actually aware of what they’re generating, instead of simply reproducing patterns. With the current models being seen as “that cute 20s toy”.
In text generation (currently dominated by LLMs), for example, this means that the main “bulk” of the model would do three things:
convert input tokens into sememes (units of meaning)
perform logic operations with the sememes
convert sememes back into tokens for the output
Because, as it stands, LLMs are only chaining tokens. They might do this in an incredibly complex way, but that’s it. That’s obvious when you look at what LLM-fuelled bots output as “hallucination” - they aren’t the result of some internal error, they’re simply an undesired product of a model that sometimes outputs desirable stuff too.
Sub “tokens” and “sememes” with “pixels” and “objects” and this probably holds true for image generating models, too. Probably.
Now, am I some sort of genius for noticing this? Probably not; I’m just some nobody with a chimp avatar, rambling in the Fediverse. Odds are that people behind those tech giants already noticed the same ages ago, and at least some of them reached the same conclusion - that better gen models need more awareness. If they are not doing this already, it means that this shit would be painfully expensive to implement, so the “better models” that I mentioned at the start will probably not appear too soon.
Most cracks will stay there; Google will hide them with an obnoxious band-aid, OpenAI will leave them in plain daylight, but the magic trick will still not be perfect, at least in the foreseeable future.
And some might say “use MOAR processing power!”, or “input MOAR training data!”, in the hopes that the current approach will “magically” fix itself. For those, imagine yourself trying to drain the Atlantic with a bucket: does it really matter if you use more buckets, or larger buckets? Brute-forcing problems only go so far.
I agree 100%, and I think Zuckerberg’s attempt at a massive 340,000 of Nvidia’s H100 GPUs AI based on LLM with the aim to create a generel AI sounds stupid. Unless there’s a lot more to their attempt, it’s doomed to fail.
I suppose the idea is something about achieving critical mass, but it’s pretty obvious, that that is far from the only factor missing to achieve general AI.
I still think it’s impressive what they can do with LLM. And it seems to be a pretty huge step forward. But It’s taken about 40 years from we had decent “pattern recognition” to get here, the next step could be another 40 years?
I think that Zuckerberg’s attempt is a mix of publicity stunt and “I want [you] to believe!”. Trying to reach AGI through a large enough LLM sounds silly, on the same level as “ants build, right? If we gather enough ants, they’ll build a skyscraper! Chrust me.”
In fact I wonder if the opposite direction wouldn’t be a bit more feasible - start with some extremely primitive AGI, then “teach” it Language (as a skill) and a language (like Mandarin or English or whatever).
I’m not sure on how many years it’ll take for an AGI to pop up. 100 years perhaps, but I’m just guessing.
I don’t know much about LLMs but latent diffusion models already have “meaning” encoded into the model. The whole concept of the u-net is that as it reduces the spacial resolution of the image, it increases the semantic resolution by adding extra dimensions of information. It came from medical image analysis where the idea of labelling something as a tumor would be really useful.
This is why you get body dysmorphic results on earlier (and even current) models. It’s identified something as a human limb, but isn’t quite sure on where the hand is, so it adds one on to what we know is a leg.
There was an interesting paper published just recently titled Generative Models: What do they know? Do they know things? Let’s find out! (a lot of fun names and titles in the AI field these days :) ) That does a lot of work in actually analyzing what an AI image generator “knows” about what they’re depicting. They seem to have an awareness of three dimensional space, of light and shadow and reflectivity, lots of things you wouldn’t necessarily expect from something trained just on 2-D images tagged with a few short descriptive sentences. This article from a few months ago also delved into this, it showed that when you ask a generative AI to create a picture of a physical object the first thing the AI does is come up with the three-dimensional shape of the scene before it starts figuring out what it looks like. Quite interesting stuff.
That’s perhaps why image generators are comparatively better than text generators. But there’s still something off, by your example it seems that the model cannot reliably use clues like position to understand “this is a «leg»”. And I don’t know much about image generators but I think that they’re still statistics- and probability-based.
As I often mention when this subject pops up: while the current statistics-based generative models might see some application, I believe that they’ll be eventually replaced by better models that are actually aware of what they’re generating, instead of simply reproducing patterns. With the current models being seen as “that cute 20s toy”.
In text generation (currently dominated by LLMs), for example, this means that the main “bulk” of the model would do three things:
Because, as it stands, LLMs are only chaining tokens. They might do this in an incredibly complex way, but that’s it. That’s obvious when you look at what LLM-fuelled bots output as “hallucination” - they aren’t the result of some internal error, they’re simply an undesired product of a model that sometimes outputs desirable stuff too.
Sub “tokens” and “sememes” with “pixels” and “objects” and this probably holds true for image generating models, too. Probably.
Now, am I some sort of genius for noticing this? Probably not; I’m just some nobody with a chimp avatar, rambling in the Fediverse. Odds are that people behind those tech giants already noticed the same ages ago, and at least some of them reached the same conclusion - that better gen models need more awareness. If they are not doing this already, it means that this shit would be painfully expensive to implement, so the “better models” that I mentioned at the start will probably not appear too soon.
Most cracks will stay there; Google will hide them with an obnoxious band-aid, OpenAI will leave them in plain daylight, but the magic trick will still not be perfect, at least in the foreseeable future.
And some might say “use MOAR processing power!”, or “input MOAR training data!”, in the hopes that the current approach will “magically” fix itself. For those, imagine yourself trying to drain the Atlantic with a bucket: does it really matter if you use more buckets, or larger buckets? Brute-forcing problems only go so far.
Just my two cents.
I agree 100%, and I think Zuckerberg’s attempt at a massive 340,000 of Nvidia’s H100 GPUs AI based on LLM with the aim to create a generel AI sounds stupid. Unless there’s a lot more to their attempt, it’s doomed to fail.
I suppose the idea is something about achieving critical mass, but it’s pretty obvious, that that is far from the only factor missing to achieve general AI.
I still think it’s impressive what they can do with LLM. And it seems to be a pretty huge step forward. But It’s taken about 40 years from we had decent “pattern recognition” to get here, the next step could be another 40 years?
I think that Zuckerberg’s attempt is a mix of publicity stunt and “I want [you] to believe!”. Trying to reach AGI through a large enough LLM sounds silly, on the same level as “ants build, right? If we gather enough ants, they’ll build a skyscraper! Chrust me.”
In fact I wonder if the opposite direction wouldn’t be a bit more feasible - start with some extremely primitive AGI, then “teach” it Language (as a skill) and a language (like Mandarin or English or whatever).
I’m not sure on how many years it’ll take for an AGI to pop up. 100 years perhaps, but I’m just guessing.
I don’t know much about LLMs but latent diffusion models already have “meaning” encoded into the model. The whole concept of the u-net is that as it reduces the spacial resolution of the image, it increases the semantic resolution by adding extra dimensions of information. It came from medical image analysis where the idea of labelling something as a tumor would be really useful.
This is why you get body dysmorphic results on earlier (and even current) models. It’s identified something as a human limb, but isn’t quite sure on where the hand is, so it adds one on to what we know is a leg.
There was an interesting paper published just recently titled Generative Models: What do they know? Do they know things? Let’s find out! (a lot of fun names and titles in the AI field these days :) ) That does a lot of work in actually analyzing what an AI image generator “knows” about what they’re depicting. They seem to have an awareness of three dimensional space, of light and shadow and reflectivity, lots of things you wouldn’t necessarily expect from something trained just on 2-D images tagged with a few short descriptive sentences. This article from a few months ago also delved into this, it showed that when you ask a generative AI to create a picture of a physical object the first thing the AI does is come up with the three-dimensional shape of the scene before it starts figuring out what it looks like. Quite interesting stuff.
That’s perhaps why image generators are comparatively better than text generators. But there’s still something off, by your example it seems that the model cannot reliably use clues like position to understand “this is a «leg»”. And I don’t know much about image generators but I think that they’re still statistics- and probability-based.