Stability AI

“So we train the whole internet and like stable diffusion is a 100,000 gigabytes in a two gigabyte file.”

“We had datacomp, which is the next generation lion that we funded the compute for, whereby at a quarter of the parameters of OpenAI's clip, it outperforms for the beta data quality.”

“TPUs are the most scalable architecture, like we have zero failure rate with our TPU language model training, whereas with GPUs it's like there's an ECC error”

“We had hundreds of millions of images opt out of stable diffusion because we were the only company in the world To offer opt out of data sets.”

“People are offering us tens of millions per model”

“I call them organic free-range models, the ones we're building with no web scrape data.”

“Well, this is what we're doing right now. We're working with our multinational partners, lots more to be announced soon, and multiple governments for a framework for what good data looks like to feed these models, to stimulate innovation and localization.”

“If you want to have distribution, what do we do? We went to Amazon and said Bedrock, because then it gives us a 100,000 SageMaker SMEs, and we just have to give them the models that they can then take to the private data, and we get a share of all of that.”

“I think that there's only going to be five or six foundation model companies in the world in three years, five years. ... I think it's going to be us, NVIDIA Google, Microsoft OpenAI, and Meta and Apple probably are the ones that train these models.”

“From next week, we're going to build our language models in the open and share what works and what doesn't work.”

“You need these chips lined up with interconnect, and we've got 7000 A-one hundreds now.”

“If you look at models closer to Runway ML or Stability AI or like PhotoRoom in Paris, you see these AI native startups that are actually building, training their own models. And how they can, in my opinion, build much better things than the ones that just use …”

“The reality is no models that are out today will be used in a year.”

“I think that there's only going to be five or six foundation model companies in the world in three years, five years. I think it's going to be us, Nvidia, Google, Microsoft OpenAI, and Meta and Apple probably are the ones that train these models.”