I refuse (yes, refuse) to make a distinction between people looking at images and learning from them and software looking at images and learning from them, and it is my fervant hope that the legal system agrees with me. I don’t want to see amazing innovation crippled by luddites.
And Stable Diffusion is free and open source. You can download it to your computer right now absolutely free and use limitlessly absolutely free. You can also modify, extend and enhance it and share those enhancements absolutely free. Massive amounts of innovation are being done for free by third parties, such as Controlnet. It is a Linux analogue, not a (for instance) Facebook one. And the cat is way, way, way out of the bag
But software isn’t just looking at images and learning from them.
I don’t see how it can. I certainly hope that it doesn’t.
The “amazing innovation” are the result of the real world hard work and creative output of the artists that populated the dataset. They deserve all the credit. They deserve the money.
If its being stored, it isn’t “just being looked at.”
That data is now subject to a whole host of laws and regulations. The GDPR in the EU, for starters. If I look at a photo that contains personally identifying information, I’m not breaching the GDPR. But if you store that same photo on a computer as part of a dataset for machine learning, then you’ve gone beyond merely looking at it, and you need to be in compliance with the applicable laws.
With your deep and fundamental mistake being the fact that the data is not stored. The Stable Diffusion training set was around 25 TB. The set of learned concepts is a downloadable file of less than 2.5 GB that never looks at the training set again. That learned data file is 10,000 times smaller than the training data set. (Almost) no images are stored, it is the result of looking. At. The. Images. And. Learning.
Then you probably don’t get a pass. The fact that the images are hosted somewhere else is irrelevant, especially if that image contains personally identifiable information.
If just the engine, we may be getting close real early to the same sort of commodification that happened to processors and OSes.
IOW What matters / will matter for a useful AI is how well it was trained on what. Not whichever underlying soon-to-be-commodity engine it’s running under and which underlying soon-to-be-commodity mega-hardware is running the engine.
And once the training data is really the secret sauce, then the fussing upthread about ownership of the raw material of the training data becomes very central.
Right now I feel a bit like a landowner who sold his land in Titusville PA just before the discovery that underground oil was useful and has economic value. And sold to somebody who did know those things. See
Many of us have set up our lives around e.g. social media posts, imgur photo sharing, email through gmail or outlook.com, document storage on cloud servers, and all along we’ve been handing very valuable IP out to the techlords for free. IP that they knew, or at least strongly suspected, would be very valuable to them. And for which they have paid and intend to pay exactly zero. While certainly charging for the fruits of our raw material.
The old joke that “if you’re not paying for it you’re not the customer; you’re the product.” has developed a sinister new sibling: “If you’re not being paid for it; you’re going to be paying for the consequences of not being paid for it.” That’s called lose-lose in my book.
Fully trained. The “engines” are totally trivial and have no secret sauce whatsoever. A reasonably experienced engineer can put one together in days, and there are a number of tutorials out there for building your own.
The training remains expensive, as does creating the training set, doing the fine-tuning, and so on. The free models out there aren’t necessarily going to give you ChatGPT out of the box–the OpenAI GPT models themselves required fine-tuning before they were really usable. But they still have a great deal of value built into them.
I watched the Sam Altman / Lex Fridman interview yesterday, and one of Sam’s points was that the usability was a key factor in ChatGPT’s success. It wasn’t that big a leap over the existing state of the art, but the interface, plus all the fine-tuning they did to make it pleasing to interact with, ended up being the most useful part. Sort of an iPhone-like situation. No one piece was really that big a leap, but the combination of small things and a ton of polish made it compelling.
That said, a lot of this will end up being commoditized in the long run. It’ll be cool if we can all get our own fine-tuned AI agents. Maybe fine-tuned on our own document archives. I’ve saved every email I’ve ever sent…
It’s not just social media and email (and texts and IMs); all of your fitness data via personal fitness trackers, financial data you shared with online mortgage and investment sites, essentially everything about your ‘digital persona’ than can be collected online. And even if you are ‘assured’ that it is protected or anonymous, you may rest assured that it isn’t; with sufficient cumulation of data will allow algorithms to extract your personal data from ‘anonymized’ sources with a shocking degree of efficacy.
You have fallen into the common belief-trap that this is all in service of you as a consumer of a beneficial service. But it isn’t; any ‘service’ you are being offered is in terms of what will make you a better “client”, or as it was known in previous generations, a “mark”. In terms of any commercially-motivated ‘artificial intelligence’ system, you are a source of profit to be exercised to the maximum extent possible, and if not directly by payment then by utilizing your personal information to manipulate you and others for increased profitability.
That’s likely the common case. But given the prevalence of these open models, and the decreasing costs of hardware, I think there is likely room for technical users to run their own models. In a similar way to how I run my own domain and website instead of just putting my shit up on social media. Already, to the limited extent I’ve used the image generation models, I’ve mostly stuck to a locally-run Stable Diffusion as opposed to any of the online ones.
You remind me of how the www was when a few brave pioneers put up the first websites. “We’re out here being the voice of topic X in locale Y. The EFF is so proud of us. Information wants to be free my fellow Netizens!”
Cute hobbyists. Who were simply bulldozed out of existence by the first of the commercial operators. Much less the monopolist behemoths that soon followed.
Yes, you and I might maintain our own AI and use it as perhaps an alter-ego mind map, having been fed all our life’s work atop ordinary generic training data. We might even loose the thing to make posts on the Dope on our behalf while we sleep.
But that’ll be a fart in a hurricane compared to what’s going on in the larger society powered exclusively by systems designed solely for money grubbing. The vast majority of humanity won’t know of anything else.
It is pretty scary. It’s a continued source of astonishment that the larger internet continues to exist, given that it’s 99% paid for by advertising, and yet it’s completely trivial to block all advertising with various tools. And yet Google and Facebook and the rest are still monsters.
Somehow, us hobbyists have still survived. Probably by flying beneath the radar. We just aren’t a real threat.