Stability AI Raises $101M at $1B valuation | Hacker News Reader

Stability AI Raises $101M at $1B valuation

techcrunch.com

101 pointsobiefernandez55 comments

I've spent the last half hour trying to figure out what this company is, and I still have no idea. My first stop was their website and I could not find a single sentence providing concrete details regarding their value proposition. It's just a bunch of vacuous statements which paint themselves as some sort of iconoclastic visionaries. Please do check it out for yourself and tell me it's not just a bunch of nonsense. For further evidence that this company hates concrete ideas and coherent sentences, head over to the founder's twitter and read his bio.

Anyways, if you scroll past the hand waving on their site, the first thing they tout is their developer communities. What exactly is a developer community? Do these communities belong to Stability? Does Stability own the research/models produced by these communities? I don't understand the value of this section at all. It seems to be just a list of ill-defined and loosely connected organizations with no clear relation to Stability.

Lastly, I am well aware of what stable diffusion is, but I still am not sure what the relationship is between Stability AI and the researchers that developed the stable diffusion model. As far as I can tell, Stability AI simply provided the compute for model training. None of the researchers list any affiliation with Stability in their publications (or elsewhere that I looked). And if this is the case, is there any reason to expect Stability to stay on the forefront of AI research? Or are they just going to open source models in the wake of actual research done by other organizations? How is lending compute to other researchers a viable business model, let alone one worth a 1B valuation? Is their stable diffusion API/Studio that much better than all the others?

Are these investors reputable? My first impression would be that they are drastically overestimating the competitive edge that Stability has.

> I've spent the last half hour trying to figure out what this company is, and I still have no idea. My first stop was their website and I could not find a single sentence providing concrete details regarding their value proposition.

As much as it sucks, they don't owe the general public an explanation of their business plan. They apparently have access to investors and can explain it to them directly.

> but I still am not sure what the relationship is between Stability AI and the researchers that developed the stable diffusion model.

Another post trending on HN today about data laundering says that Stable AI funds the university research behind Stable Diffusion. The data laundering comes from the university using otherwise copyrighted data for "research purposes" and Stable AI can in turn use the trained model commercially, sidestepping the fact that the underlying data is not available for commercial purposes.

Another way to read that situation is simply that it is easier to outsource your research than build your own team.

>As much as it sucks, they don't owe the general public an explanation of their business plan. They apparently have access to investors and can explain it to them directly.

Like Theranos and Magic Leap!

Their first deliverable was wildly successful, shot a hole in the business model of open ai and was praised by generally everyone.

So not quite.

I am curious what the pitch is though. As investors would want return, I assume

Reminds of of that old as rocks joke about the most powerful computer available being graduate students.

Graduate student descent.

> As far as I can tell, Stability AI simply provided the compute for model training.

As models grow complex, Compute is what's holding back AI from mass production and consumption [0]. And so whoever figures out a way to give away Compute for inference and training is likely going to be able to provide the resulting service at much competitive rates to keep the flywheel going?

Sort of similar to Google and Chrome/Search/YouTube/Android, where even if untold billions are spent to build and run the software (Compute), to the end-user (AI researchers) it costs virtually nothing, while Ads continue to bring in revenue (APIs / services).

I see StabilityAI as Yang to OpenAI's Ying. And I believe, that's where the upstart's evaluation stems from.

[0] The new business of AI (future.com), https://archive.is/g4bEe

There is an interesting interview with "the man behind Stable Diffusion" at youtube.

https://www.youtube.com/watch?v=YQ2QtKcK2dA

They wrap up studies into API calls or applications, which being sold to developers or end users.

These second-tier AI company don't need the be the best. "Good enough" with competitive pricing will do.

Thanks. So they will offer similar APIs as OpenAI and Hugging Face?

> I've spent the last half hour trying to figure out what this company is, and I still have no idea

That's not unusual.

As long as the investors understand it, that's fine.

Even though the investors have no clue what it does, it's still fine as long as the investors are willing to write the cheque.

>And if this is the case, is there any reason to expect Stability to stay on the forefront of AI research? Or are they just going to open source models in the wake of actual research done by other organizations?

Even if this was the case, turns out this is more valuable than just holding the model close to your chest and not letting people use it, build off it.

Coatue is literally the definition of an unsophisticated investor. Their AI investments and initiatives have been disasters thus far.

They're the ones that released Stable Diffusion (the ML model that can generate images from text prompts).

See also "Stability event happening now. News so far: Animation API next week, DreamStudio Pro in progress (automatic gen of video from music, latent space exploration) will fund 100 PHDs this year. Their cluster is 4000 A100s on AWS and wants to grow 5x-10x next year. will reduce by half price of Dreamstudio"

https://old.reddit.com/r/StableDiffusion/comments/y6v0v9/sta...

How does this change their organizational model now they have VC money?

> “Nobody has any voting rights except our employees — no billionaires, big funds, governments or anyone else with control of the company or the communities we support. We’re completely independent,” Mostaque told TechCrunch in a previous interview. “We plan to use our compute to accelerate open source, foundational AI.”

> We’re completely independent

After taking $100 million of other people's money, were they born yesterday?

New way of founding allows founders to retain 100% control. Facebook and Google are examples in their early years.

Owning voting shares is quite different from being independent. Investors had plenty of influence on those companies regardless of the founders setting up dual class shares.

If you think a startup with investors operates the same way as one without, and that it can be independent the same way, I'm not sure what to say.

Stable Diffusion is simply fantastic, we live in great times to have such a thing as open-source to play with. And however justified the skepticism expressed in other HN comments might be, I truly hope they figure out a business model that’s compatible with continuing to make these releases. Otherwise, it’s forever going to be DeepMind’s “we actually have much better models, no you can’t see them, they go to a different school”, or alternatively, OpenAI’s insane prices and restrictions. For this field to become a game changer, the public needs access.

Are OpenAI’s prices really “insane”? I have been using their GPT-3 based APIs for probably close to a year, and they seem really inexpensive compared to running a server myself with the required memory and compute, not to mention convenience. I am just starting to use some Hugging Face APIs, but so far I don’t have enough experience to talk about them.

Makes the stalling on releasing 1.5 publicly more understandable. Hopefully they won't become the next "OpenAI" but the material incentives are not aligned with public releases.

Emad said on Twitter that this funding round completed 8 weeks ago and it was the last few weeks he has been stalling.

For over a month after the funding round completed he was eager to release as much as he could, including 1.5

My read on the issue is that either of (or all) the following is causing some concern:

1. The severe pushback from artists regarding the fact that the model is trained on living artists artworks.

2. Almost every implementation removed the nudity filters.

https://mobile.twitter.com/EMostaque/status/1582055542960230...

This is quite stunning. OpenAI valuations are quoted as $3B. Stability AI just popped up6 months ago and is able to raise rounds in very tough economy at $1B valuation. The real differentiator between two is that Stability AI has took upon itself to o democratize AI ignoring all social wokness and thereby multiplying their efforts many manifold through community labor. OpenAI meanwhile is still deeply caught up in all kind of social/ethical/philosophical delima as well as Apple-isc secrecy as devotional offerings to Steve Jobs.

With this level of funding and community following, Stability AI is en route to dethrone OpenAI in almost no time.

Microsoft invested $1B in OpenAI in 2019, the valuation is probably somewhere in the neighborhood of $5-20B.

Wasn't that only in Azure credits or somesuch?

In the articlr it says Stability AI was founded 2 years ago, so they are not popping out from nothing.

both numbers are total BS

I'd love to see the SEC start requiring VCs to underwrite these idiotic "valuation" numbers...you say its worth $1 bln, you need to be prepared and required to potentially buy back pre-IPO shares from invested parties at that valuation

now watch those values drop, rapidly

did we not learn anything from Uber being "valued" at $150bln?

The job of the SEC is not tell if a stock price is too low or too high. The SEC’s can only protect investors by ensuring companies do not lie. At the moment the SEC starts to decide something is too expensive then the stock market effectively becomes socialism.

It should be everyone’s own resposibility to decide if they invest in something and also be prepared to take losses on that investing. There cannot be investing without the risk taking aspect. Whoever invested in Uber at too high valuation should bear the resposibility of losses of their investment.

On what business model exactly?

According to the article, they plan to train private models. They claim to have done R&D that makes them more efficient at training models.

Does that sound like a $1 billion business to you? Even at 6 figure profit per model, they are going to need to build up a huge sales arm to do enough business for that valuation.

I think they are expecting to become some sort of artist's tool, but since the model is open source and artist's tend to have good GPUs, Adobe can crush them by putting their own model (or a similar proprietary model from Adobe) into photoshop.

It does not imo but also seems like a cheap way to back a possibly asymmetric group. The valuation is not the hurdle for investors, as they will have seniority (likely with a multiplier) in any liquidity event.

Does it seem like this group is moving quickly and making waves in a potentially huge market? Yes. Is this still a lot of money and very risky? Also yes.

It does not at all. I agree with you.

DreamStudio got one million users already, generating 17 million images. Granting that most people don't pay in (I got $2 credit for 200 generations by logging in), a $10/1000 images is still, ah, good enough I'd say. Especially if they stayed fresh by agglutinating all the advances in the open ecosystem.

You aren’t factoring in the cost of running the model on GPU instances. They might make a bit, not sure if it’s enough to satisfy their investors.

The investors are buying into a toolset for what will potentially become open source AGI that is already blowing the competition out of the water in terms of capabilities, speed of development, and access/reach. Add together large language models with text and code interpretation/generation + diffusion + vision + speech and sound + embodiment = what might you get?

It's unlikely investors care more than an inkling about the cashflow positivity of a (really fun to use, admittedly awesome) stock image synthesis engine other than its demonstrated impact as a stepping stone and buzz generator. I suspect their pitch was a little bit broader in scope than making Getty redundant.

VC exit is over a 10+ year horizon -- people writing cheques are looking to own a piece of the future, as is the case with all catchy narratives, but in this case the future is already here and publicly disclosed progress in the field appears to be accelerating rapidly in no small part due to Stability AI and the research communities they support.

Leading to AGI? I have been using deep learning professionally for about seven years (and neural networks since the 1980s with a commercial product and I wrote the model for SAIC’s bomb detector they did for the FAA).

Personally, as awesomely useful as deep learning is, I don’t see the path to AGI. I think AGI will be a hybrid of DL, GOFAI, and things not yet invented.

Couldn't agree more - generalised task capable systems will likely be composable architectures that draw upon a grounding in variety of methods, not just DL by any means. Synthetic sensoria capable of integrating, interpreting and generating general task capabilities are likely to lead us somewhere interesting. DL / LLMs are akin to Wintermutes before waking; with whatever passes for consciousness instantiating and evaporating with each prompt input (which some might argue is a good thing for foom-resistant biospheres) and no continuum vector state pseudoawareness. However, strongly feel we'll have instruct-GPT models on the edge as an interface modality, but they're not the holy grail or be all end all.

Sure but the comment is still an appealing pitch to investors, and probably can be read broadly to include "things not yet invented".

Out of curiosity - what is GOFAI?

Good Old Fashioned (symbolic) Artificial Intelligence. And thanks, absolutely my clumsily worded investor pitch intent was to cover techniques not yet invented, thanks for picking up on that.

Access to patents?

This is the one that took over the community stable diffusion subreddit moderation from a kid right?

And gave it back.

Oh yeah? This space is moving fast

Video of the investment announcement event/keynote: https://youtu.be/1Uy_8YPWrXo

They're leading the way with open source models, I wish them the best!

That didn't take long to sell out.

How did you expect them to pay for the $50M in infrastructure and R&D costs spent on giving you and millions of developers worldwide access to their models for free?

Is this comment supposed to be satire?

Stable diffusion cost just $600,000 to model.

> Stability AI has a cluster of more than 4,000 Nvidia A100 GPUs running in AWS, which it uses to train AI systems, including Stable Diffusion. It’s quite costly to maintain — Business Insider reports that Stability AI’s operations and cloud expenditures exceeded $50 million.

Directly from the article.

VCs just got fleeced.....

Not if they can flip it to a bigger fool.

No idea what they do.

Taken from their website, they seem to have a range of products like:

- eleutherai: A decentralised collective focusing on AI alignment

- Laion: A community-driven non-profit organisation devoted to making large-scale AI openly accessible.

- DreamStudio: Open-sourced image generation model that cultivates autonomous freedom to produce incredible imagery.

- ... few more things with a bunch of nice words I can't make sense of like their coming soon DeepFloyd that is about "Discovering new facets of Multimodality."

The only thing I could get that point to something concrete is the link to their github which was created end of August of this year from which they seem to claim: "Our vibrant communities consist of experts, leaders and partners across the globe. They are developing cutting-edge open AI models for Image, Language, Audio, Video, 3D and Biology"

Reply on news.ycombinator.com