The company behind Stable Diffusion appears to be at risk of going under
futurism.com
futurism.com
As far as I know they pretty much just paid for servers to train a big shiny model on that was based on research they had no hand in. Throwing money at researchers after they came up with something good, just to let them build a big shiny version of it, does not retroactively make their accomplishments yours.
Basically they hold no rights to anything relevant, no patents, no secret sauce, nothing. Them going under after exhausting their money will hardly have any effect.
If you look at the stable-diffusion repo, it says that SD is based off a colab with StabilityAI and Runway.
Most of the value from these models is the training + dataset. The architecture is open source, and we've had flavours of it for several years. SD has some improvements on how it handles diffusions, but most architecture out there use about the same, but with wildly different results.
The datasets used are provided by LAION.
Here is a honest summary of Stability.Ai's involvement[2]:
> In their project, the LMU scientists had the support of the start-up Stability.Ai, on whose servers the AI model was trained. “This additional computing power and the extra training examples turned our AI model into one of the most powerful image synthesis algorithms,” says the computer scientist with a smile.
[1]: https://ommer-lab.com/research/latent-diffusion-models/
[2]: https://www.lmu.de/en/newsroom/news-overview/news/revolution...
Which is their business model.
Provide compute to people who can’t afford compute so the only people doing AI research aren’t doing so behind closed doors.
Now, it seems, giving away your product isn’t all that profitable and they need to “pivot” to find a way to keep the business running.
Judging from the interview posted elsewhere in the discussion they make no claim to be inventing anything but just wanted to democratize AI research.
There is a ton of momentum behind the general public’s belief and perception that image generation is free for small uses and cheap for large uses. Most of the other players can afford to keep those losing prices for a long time, too. I think it’s going to be an uphill battle to charge for image generation at amounts that turn a profit. I wouldn’t rule out a more creative way of monetizing it, but the obvious routes look unlikely.
I've noticed that they've recently rolled out some new features (e.g., "/describe") that were at first limited to the higher tiers, but after a short time of evaluating the increased GPU load, were enabled for the lower tiers as well.
To me, that seems to indicate "can't install GPUs fast enough" rather than "running out of money", but of course it's just an indication.
I believe their stated reason for getting rid of the free trial accounts (temporarily, they said) was that people were using bots to register hundreds of sockpuppet accounts to get the freebies. Perhaps they'll come back when and if they can deploy some effective anti-bot measures.
What would be the incentive for a person, not a company, to pay Stability AI company instead of downloading and doing a bit of setup to have their own uncensored model?
If your product is tested and guaranteed to certain standards, like MySQL and RedHat, then that is something a company may pay for. But a user doesn't really have that high standard so they can be satisfied with just the off brand, derived stuff floating around.
We were a bootstrap so were immediately profitable.
Things like "ok, we can't use one drive - need a different tool," "can't use Sharepoint, need to use a different tool," and so on.
The market is there. RHEL and similar are well established.
In order to make this a "we should do this" either government IT funding needs to be significantly increased (that is difficult in the current political climate in most places) or the support offerings and staff needed for the average user (using Windows, Sharepoint, Word, Excel, Teams, and Project) needs to be competitive with the pricing that Microsoft offers.
That should be "simple" - make a company that offers the same level of support as Microsoft does for a packaged suite of software that includes easy installations, appropriately locked down desktops, call center, and so on.
And if that can't be found at the same price that Microsoft offers - then we return to the "increase government funding."
Saying "we should outlaw government spending on closed source" misses a lot of the tools out there that are needed to keep things running. Is there a FOSS (with support contracts for the stack) alternative to Cerner or Epic? SAP? ArcGIS? And that's not even getting deep at all into the niche SaaS tools that some pieces use for specific problems.
The market is there and state and local governments would likely jump at the opportunity to switch if there's a company that can offer the same functional stack with the same support for the same price or cheaper built on top of FOSS. Otherwise... persuade those state and local governments to staff up to the necessary levels to be able to hire people able to customize and support the FOSS to fit their needs and be prepared to financially support that decision.
Of those only ArcGIS isn't a basic CRUD, and yes, there are plenty of open source replacements for it. It's in fact way behind the open source tools.
There aren't for the others, because they are about paying a company so you can use its other customers data, or paying it to get overpriced consultants. None of them are about software.
Sure it’d be great for government acquisitions to subsidize open source, but at what cost?
This is for 200 users and it needs 24/7 support. The budget is $30k/y.
Ahh, you could spin up Redmine on AWS for much less...
Maybe, but I'm going to need to hire a Redmine consultant to help configure it for our needs, add maintaining and updates for that instance to the sysadmin and add it to the helpdesk support - and if we really want 24/7 support, our helpdesk is only 7-5 business hours (the support is for the 'it crashed' which means that the sysadmin gets it on pager duty... but if there's an issue that's application support, so wake up an application team person who isn't on pager duty...). And there's a bug in Redmine, so either get a developer to learn Ruby (this is a Java shop) or beg the core team to fix it...
You know, Jira is looking more attractive as these likely costs start adding up - and it's a fixed price... not going to have an AWSh.. surprise if a bitcoin miner exploits an issue on the base image that didn't get updated.
Local and state governments do look at open source first. The "no cost to start" is very attractive until they get to a "this is a hard problem, you're on your own" and the costs go way past what a commercial application costs.
For government, a predictable budget so that you can ask for next year is valuable. Unknown budgetary costs are were overruns happen. A budget line item of $26,500 per year is much better than a budget line of "$0 to $50,000."
Metabase is another company I knew where they do this, but I don't know how successful they are (their hosting costs start at "way too high considering how easy the thing is to spin up", which was great for me at $PREV_JOB I suppose but)
I'd suspect the moment you touch anything with external auditing, it looks better to say "we've got a fully paid up support contract from the vendor" than "we're running v1.23.456-ubuntu-patch-357-with-chives-and-salsa that we downloaded last week."
None of which are profitable
I do not know the details, but based on the fact that some of the orgs involved in the project literally exists to democratise ai, I believe that many of the stake holders in that project were adamant about open sourcing it.
In this case they were actually the first to the punch(along with openAI)!
In fact, I know of a huge company that does all their research in the EU, where they can get gov't funding, but then the process to turn them into products is in the US (less regulation, larger market).
So lots of highly talented people just do open source in their free time or work at universities.
> As mentioned earlier,[MIR](Sec. 21) when only consulting surveys from the Anglosphere, it is not always clear[DLC] that Deep Learning was first conceived outside of it. It started in 1965 in the Ukraine (back then the USSR) with the first nets of arbitrary depth that really learned.[DEEP1-2][R8] Soon afterwards, multilayer perceptrons learned internal representations through stochastic gradient descent in Japan.[GD1-2a] A few years later, modern backpropagation was published in Finland (1970).[BP1] The basic deep convolutional NN architecture (now widely used) was invented in the 1970s in Japan[CNN1] where NNs with convolutions were later (1987) also combined with "weight sharing" and backpropagation.[CNN1a]
The Stable Diffusion models are proprietary freeware, not open source.
I wonder if, then, Stability AI wanted to sell license exceptions—namely, the ability to use the software for amoral and (mildly) immoral uses.
Nobody benefits from their failure.
If Runway and Stability can cut costs they will become cherished institutions.
without a stable source of revenue, cutting costs mean nothing. What are they selling, and why isn't there buyers?
A business cannot survive without revenue, and revenue only comes from people who want to buy something from you that they cannot otherwise get else where.
I applaud StabilityAI for releasing SD for free. They could've done what openAI did, and monetized it (which other diffusion model services are doing). By releasing it for free, stabilityAI contributed to the common good. Unfortunately, if there's nothing else that can be sold as a product, they cannot be sustainable.
An alternative, which i'm not too big a fan of, is collectivization of AI models, and make it a global commons for which taxpayers will fund.
...
> cut costs
Unless they can cut costs to zero, something has to change.
Like, example. I use SD in Blender sometimes as part of the compositor. I have maybe a 10% acceptance rate for SD output: sometimes the water isn't right, or the clouds look goofy, or something keeps getting rendered as an anime pillow for some godforsaken reason. If SD captured my prompt history and some of the final model tweaks between runs, they could ostensibly get really solid HITL test data. Then they could be the curator of that "super model" which they could upsell, maybe along with very high rez stuff, or a higher priority on jobs. Again, not an expert, so who knows. And also, having the model local, that gives you back some of the same benefits, but without the scale.
Not too surprised about funding issues from the casual answer.
I’m not saying it was bad to self fund a project, but having to choosing between your life and fun (and potentially very profitable) projects is not easy.
[1] https://github.com/BlinkDL/RWKV-LM
[2] https://huggingface.co/laion/CLIP-ViT-L-14-laion2B-s32B-b82K
[3] https://github.com/lucidrains/gigagan-pytorch#appreciation
This time around though, the means of production are available to anybody with a credit card, and AWS/GCP/Azure throw credits at startups that apply, just to have them locked into their cloud. Break free of the chains of work, with generative pretrained transformers!
is it “company was planned vehicle for Napster except freeing images not music”
Or is it “an instance of companies wasting money because ZIRP and some companies are bad like some stuff I disagree with that happened at Uber”
In fact, I think one could argue that many companies raising that much money early on are indeed trying to become a first mover and don't care about making money (yet).
This won’t be possible without accelerator compute for training. Open source developers can’t afford this compute.
Apple benefits because they can deploy on iOS devices. Amazon benefits because they can tarball the models and sell it as a managed service.
I don’t think either wants to be in the business of figuring out which researchers is going to use the compute for deep fakes or for the next big model.
So something like Stability should exist, Amazon and Apple should figure out how to make that happen.
There's a heap of bad banks and terrible debt floating around, the golden goose has been cooked. Turns out risks aren't just things you can ignore for 'growth.'
Focusing on hype and growth over profitability and burning hundreds of millions of VC cash on data scientists and AWS for training, fine-tuning AI models. This is even before mentioning the mounting lawsuits they are already facing.
The inefficient training of deep learning models is unsustainable for these pre-profit companies.