>I remember when OpenAI practically cloned Scarlett Johanssen's voice...
There is absolutely zero evidence for this. I find it infuriating that this keeps being stated as a fact. So they go and hire a voice actor and clearly use her voice to train, but then they also scrape Scarlett Johansson from youtube and splice it into the training data to make the voice a bit more like hers? Really does that sound realistic?
Motive: Altman had some weird boyish thing for her and they asked her first, she said no.
Means: Lots of available data to use from her movies. They probably trained a model first without releasing it just because it’s ridiculously easy. Especially for OpenAI.
Opportunity: AI is astonishingly good at laundering and remixing without exposing the training set, for previously-unseen levels of plausible deniability.
> 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.
There is absolutely zero evidence for this. I find it infuriating that this keeps being stated as a fact. So they go and hire a voice actor and clearly use her voice to train, but then they also scrape Scarlett Johansson from youtube and splice it into the training data to make the voice a bit more like hers? Really does that sound realistic?