Not so stable after all.
This is the best summary I could come up with:
The massive GPU clusters needed to train Stability AI’s popular text-to-image generation model Stable Diffusion are apparently also at least partially responsible for former CEO Emad Mostaque’s downfall – because he couldn’t find a way to pay for them.
According to an extensive exposé citing company documents and dozens of persons familiar with the matter, it’s indicated that the British model builder’s extreme infrastructure costs drained its coffers, leaving the biz with just $4 million in reserve by last October.
It’s not clear whether those bills were ultimately paid, but it’s reported that the company – once valued at a billion dollars – weighed delaying tax payments to the UK government rather than skimping on its American payroll and risking legal penalties.
By late 2023, venture capital outfit Lightspeed had reportedly raised alarm bells expressing its shock after discovering the poor state of Stability’s cash flow, which apparently had not previously been disclosed.
The situation came to a head late last month when Mostaque revealed on social media he had resigned stating that “the concentration of power in AI is bad for us all,” and that in order to address it he had “decided to step down to fix this at Stability and elsewhere.”
Even if the biz manages to turn its financials around, it’s still facing down multiple copyright infringement cases brought by Getty and other artists, who allege their works were used without permission to train its signature model.
The original article contains 868 words, the summary contains 241 words. Saved 72%. I’m a bot and I’m open source!
If stability ai fold, what’s the alternative for stable diffusion?
As far as I know there aren’t any other open source alternatives at the moment that are comparable to commercial models. The main roadblock for open source models is the cost of initial training. As we see with Stability AI, relying on companies to do this isn’t really a sustainable approach. I’d really like to see more work going into stuff like Petals to allow training and running models using a distributed network.
Network latency will make distributed training a very time-consuming task.
Sure, but I don’t think that’s a show stopper since you don’t need to do comprehensive training often. Also worth noting that stuff like LoRAs allow extending functionality of models without retraining from scratch. So, most training might be relatively small within a specific context.
You don’t need to do it often, but initial training requires huge ressources and someone has to do it, if you want to create new models from scratch. And for this you need your compute packed as close as possible.
Not sure what your point is here. The whole point of stuff like Petals is to facilitate a way to do this by harnessing a lot of computers around the world. It would be slower than doing it in a data center, but it’s not a show stopper if this is something that only needs to be done occasionally.
Sorry, I thought that we might be underestimating the factor of “slower”, but I couldn’t quickly find numbers to prove my point. I might be wrong after all. I wish you a good night. 😊
Time to use a competing service to pump out a bunch od content marketing articles to bring in more VC $$
