Lots of misconception about GenAI ...
Latent Space as a New Medium
31–40 of 41 posts
Re: Latent Space as a New Medium
#32That's a mischaracterization. Latent space is simply a (multidimensionally) sorted collection, it's only a piece of the pie. A massive amount of structure is held in the unembedding layer. Generative AI models are a very specific ordering, LLMs a very specific subset of that, and they're hardly the only users of the concept. I get what the author is going for, and they're on the right track. There is something intere…
>Latent space is simply a (multidimensionally) sorted collection, it's only a piece of the pie. I agree about the small piece of the pie, but I can't really see it as a sorted collection. The essence of many dimensions is that you can head off in another direction that doesn't impact the relationship of other dimensions. It seems common to consider a latent space as a encoding of meaning. It is certainly a mapping of…
The Alethiometer! 30ish concepts in distinct order can convey most things, but it's coarse. We have a trillion symbol Alethometer!
Re: Latent Space as a New Medium
#33There's a lot of very strong claims made at the start of this article. (emphasis mine) > A Large Language Model (LLM) is like a small zip file that contains all human knowledge. > In a strange but real way the resulting tiny file contains all the information that is on the internet and in our libraries. > Likewise, the LLM could recognize the face of almost any person, and it could generate any possible human face, A…
Can you provide reasons other than "he's wrong"? There is no other technology that compresses so my knowledge into such a small space. A 24 GB diffusion model can generate a good approximation of any human relevant image from a description. Ask it to generate an image of the Tivoli fountain and it will do an impressive job. You can't use it as a map, but it gives an excellent representation.
Finally I don't think there are "dangerous ways" to think of any technology. This is just another tool.
Re: Latent Space as a New Medium
#34There's a lot of very strong claims made at the start of this article. (emphasis mine) > A Large Language Model (LLM) is like a small zip file that contains all human knowledge. > In a strange but real way the resulting tiny file contains all the information that is on the internet and in our libraries. > Likewise, the LLM could recognize the face of almost any person, and it could generate any possible human face, A…
Can you provide reasons other than "he's wrong"? There is no other technology that compresses so my knowledge into such a small space. A 24 GB diffusion model can generate a good approximation of any human relevant image from a description. Ask it to generate an image of the Tivoli fountain and it will do an impressive job. You can't use it as a map, but it gives an excellent representation. Finally I don't think the…
If all that the author was saying was "LLM weights are highly compressed encodings of knowledge", I wouldn't have any complaints. He is making much bolder claims than that.
And if he was describing LLMs as "just another tool", I wouldn't be complaining either.
Re: Latent Space as a New Medium
#35>In other words, correctness, truth, cohesiveness, completeness, comprehension, etc are all essentially patterns that are mapped in this space. The caveat is that truth in latent space is just a reflection of the consensus from the corpus, and you find truth by comparing the answer in latent space to what is in reality. But I just hate this idea that truth and facts are no longer real, they're just "directions". The…
> But I just hate this idea that truth and facts are no longer real, they're just "directions". There's another way to look at this which is just as pernicious, and you've already mentioned it. That's the idea that "truth" is the consensus of the training data. Even if you include several layers of meta-training-data (i.e training data that comments on the accuracy/truthfulness of other training data), this idea that…
Indeed how do we break free from the tyranny of group think?
Re: Latent Space as a New Medium
#36[dead]
Re: Latent Space as a New Medium
#37There's a lot of very strong claims made at the start of this article. (emphasis mine) > A Large Language Model (LLM) is like a small zip file that contains all human knowledge. > In a strange but real way the resulting tiny file contains all the information that is on the internet and in our libraries. > Likewise, the LLM could recognize the face of almost any person, and it could generate any possible human face, A…
Tbh a trivial generative image model could generate any human face, because it could generate any image, just randomly generating pixels is enough.
Re: Latent Space as a New Medium
#38Earlier quoted context omitted.
That's a really interesting claim, can you give an example of a question that could reasonably be on a K-12 exam that a frontier model can't answer?
I learned to drive in a drivers ed course at school in tenth grade. There's also home ec, wood shop, metal shop, auto shop, sculpture, choir, theater arts, marching band, phys ed.
Surely the ability to drive or work wood is separate to the knowledge required to gain the ability? An LLM can explain how to drive, how a car works, wood working processes… Some knowledge is difficult to communicate through text, and consumption of text might be a poor way to gain the ability to use it, but the knowledge is there?
Re: Latent Space as a New Medium
#39There's a lot of very strong claims made at the start of this article. (emphasis mine) > A Large Language Model (LLM) is like a small zip file that contains all human knowledge. > In a strange but real way the resulting tiny file contains all the information that is on the internet and in our libraries. > Likewise, the LLM could recognize the face of almost any person, and it could generate any possible human face, A…
Can you provide reasons other than "he's wrong"? There is no other technology that compresses so my knowledge into such a small space. A 24 GB diffusion model can generate a good approximation of any human relevant image from a description. Ask it to generate an image of the Tivoli fountain and it will do an impressive job. You can't use it as a map, but it gives an excellent representation. Finally I don't think the…
The onus is not on the commenter. The author made badly qualified statements. Thinking of anything as an oracle is dangerous, an “oracle” is not just another “tool” it’s a religion
Re: Latent Space as a New Medium
#40I like how Kevin Kelly brings together the ideas about what could be next. That’s always been his thing, hasn’t it? He is probably on to something here and it’s ok to be vague, I think. I don’t mind at all whether it’s all 100% technically correct.
Here are the bits I enjoyed:
LLMs (or GenAI, in general) as (the most?) efficient compression machines.
Creative or new ideas are the gaps or white space of the latent space. They already exist but become harder and harder for humans to find. It’s the perfect job for computers, therefore.
The concept of “true” is defined in the latent space (or compression) itself.