4 ms·

> AI is astonishingly good at laundering and remixing without exposing the training set They just about manage to make a good multimodal transformer that can g

by golol 2y ago

> AI is astonishingly good at laundering and remixing without exposing the training set They just about manage to make a good multimodal transformer that can generate audio and you expect that right away they can also interpolate in latent space? How does that actually work? It's not so simple. What benefit do they have from training on Scarlett Johansson's data, because they sure as hell have a big risk. They clearly hired a voice actress and they clearly told her to sound like Scarlett Johansson in "Her", and the end result perfectly fits with that. The voice doesn't "uncannily sound like SJ", no it just vaguely resembles her voice and mostly just mimicks the mannerisms from the movie. For me this is a perfect example of Occam's razor. One explanation is simple and realistic. The other explanation requires significantly more advanced AI control than OpenAI has claim/demonstrated, and it requires Altman to be so obsessed with the SJ idea that he goes out of his way to secretly train on her voice, risking legal exposure, while still hiring a voice actress.

klabb3 2y ago
To be clear I’m not saying it did happen for sure. Only that it’s realistic and not exactly far fetched given everything that’s publicly known. > they can also interpolate in latent space? How does that actually work? I don’t know. I assumed it is the case because of the vast and excellent interpolation AI appears to do in other domains.
- golol 2y ago
  > I don’t know. I assumed it is the case because of the vast and excellent interpolation AI appears to do in other domains. Sure it's plausible just like many other forms of baseless speculation are, I still don't think to state it as a matter of fact like it is happening in this and other articles and in comments. > I don’t know. I assumed it is the case because of the vast and excellent interpolation AI appears to do in other domains. Yes but do you know of any examples with sequential data, and in particular transformers? I don't and I think it is a somewhat different thing. With pictures you just train an encoder into some low dimensional latent space and there you can interpolate, but working with sequences you don't want to give the whole sequence as an input to the model and generate the whole sequence as an output. The natural way to do interpolation with an audio transformer model, which it can in principle indeed do, would be to somehow audio-prompt engineer it to speak in a voice that interpolates SJ and your voice actor. But that would be a pretty fragile method.

by golol 2y ago

klabb3 2y ago

To be clear I’m not saying it did happen for sure. Only that it’s realistic and not exactly far fetched given everything that’s publicly known. > they can also interpolate in latent space? How does that actually work? I don’t know. I assumed it is the case because of the vast and excellent interpolation AI appears to do in other domains.

golol 2y ago
> I don’t know. I assumed it is the case because of the vast and excellent interpolation AI appears to do in other domains. Sure it's plausible just like many other forms of baseless speculation are, I still don't think to state it as a matter of fact like it is happening in this and other articles and in comments. > I don’t know. I assumed it is the case because of the vast and excellent interpolation AI appears to do in other domains. Yes but do you know of any examples with sequential data, and in particular transformers? I don't and I think it is a somewhat different thing. With pictures you just train an encoder into some low dimensional latent space and there you can interpolate, but working with sequences you don't want to give the whole sequence as an input to the model and generate the whole sequence as an output. The natural way to do interpolation with an audio transformer model, which it can in principle indeed do, would be to somehow audio-prompt engineer it to speak in a voice that interpolates SJ and your voice actor. But that would be a pretty fragile method.