Discover more from jonstokes.com
The Story So Far: AI Makes For Strange Bedfellows
We're in an era of weird alliances and strange bedfellows, and there are even more realignments coming.
The previous “The Story So Far,” where I tried out narrative-style updates about goings-on in the world of generative AI, was a success. The results were good enough to keep at it, so here’s the next installment. Notice that instead of putting the date in the title this time, I’ve opted for text so that it’ll be easier for me to remember which installment I put what ideas in.
jonstokes.com is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.
Here’s a plot twist I didn’t see coming but should have: rumor has it that Google is worried enough about the supposed dangers of AI that it’s quietly pushing regulators to rein it in.
This is just one tweet from Balaji, but a few things lend it credence:
Balaji is extremely well-connected and hears all kinds of things from all kinds of people.
It’s well-reported at this point that Google has pressed some kind of internal PANIC button about AI — though whether it’s a business model panic button in response to ChatGPT or an AI x-risk panic button in response to what it knows about how close we are to AGI is not always clear. Could be both.
An open-source ChatGPT, on the pattern of Stable Diffusion, would no doubt be a real, existential threat to Google’s search business. And such a model is definitely coming, almost surely this year.
Big, centralized entities are always gonna use regulatory capture to squash potential threats. Tale as old as time.
🔀 The reason this is a bit of a plot twist for me is that I started this Substack with an examination into the controversy around Timnit Gebru’s famous paper, On the Dangers of Stochastic Parrots: Can Large Language Models Be Too Big? [PDF]. This paper that argues that large language models (LLMs) are dangerous didn’t sit well with Google higher-ups, and it ultimately played a major role in the authors’ departure from Google under a storm of controversy.
But I’d be willing to bet money that some of the anti-LLM arguments in this very paper are showing up in Google’s dark warnings to lawmakers the need to get AI under control.
(🦜 I actually think the “Stochastic Parrots” paper was just bad, at least initial leaked draft of it was, and that’s why Google didn’t want it published. This is literally what they told the authors — it wasn’t up to the company’s quality bar. I heard on background from multiple other ML researchers that the parts on LLM energy use and environmental impact were weak in the first draft. I also think that even in the draft that was published, the supposed harms of LLMs are overhyped. If I were Google, I wouldn’t have wanted it to go out under my name, either. But in the popular lore around this document, Google tried to block it from getting published because they didn’t want their people warning the public about how dangerous LLMs are.)
I talked to one of the Googlers who’s lobbying congress on this issue, and he told me: “It's about trying to put your foot down before this whole AI thing runs amok, and get people to start a debate about, to get Congress to start setting relative parameters about where technology is going.”
Just kidding, that’s the famous Lars Ulrich quote about getting congress to crack down on the whole entire internet over the narrow problem of music piracy, but with the word “AI” swapped in for the original’s “internet.”
🤝 At any rate, I think Google, Microsoft (plus OpenAI, which Microsoft owns almost half of), Facebook, and other large players are going to make common cause in the war against decentralized, open-sourced AI with the following players:
AI ethicists like Gebru, Mitchell, Bender, and others who hate any powerful cultural force that they can’t capture and impose their own values and language norms on.
Big Content shops like Disney, who are threatened by the proliferation of open-source AI tools for essentially the same reason the aforementioned AI ethicists are — i.e., they’ll lose control of a large patch of culture that they quite literally own and profit from.
Reporters and journalists who are threatened by the prospect of robots taking what’s left of their jobs. (Robots started taking journalism jobs way back with the original launch of Google News, which was editorially curated by algos. Also, the advertising robots inside Google and Facebook have already taken most of the journalism jobs by killing the industry’s business model. So with generative AI, we’re really only talking about a coup de grâce here — just finishing off what’s left.)
Independent artists who are threatened by the deflationary implications of generative AI for the cultural products they make. There’s also a real, justified alarm in these quarters over a loss of artistic control over one’s signature style. There were a set of norms and incentives around style copying that have been blown apart by generative AI, so these folks will be looking to replace those fuzzy norms and incentives with black-and-white laws and litigation.
The anti-“disinfo” industry, which is mostly made up of disinformation spreaders who are paid to gin up scares about the spread of any narratives that they can’t control.
The DEI industry, which will oppose decentralized AI for the same reasons the AI ethicists, Big Content, and Big Disfino do (i.e., narrative control).
So there are many forces of varying levels of financial and cultural power that will come together to stop the spread of open-source AI models. This is going to be a big, ugly fight that will make for some exceedingly strange alliances. Indeed, most of these groups have historically been enemies, but they’re coming together to oppose the common enemy of decentralized AI.
✊ Much of the anti-AI push is already being framed in the language of “creators’ rights,” which is ironic because what these folks are fighting against is the right of all creators everywhere to freely access some of the most productive tools humanity has yet invented. It’s as if a group of Stone Age hunter-gatherers came together to ban agriculture in the name of protecting food producers’ rights.
For Big Tech, of course, the reason they want to squash decentralized AI is clear: it’s a threat to their ability to sell ads at scale. If users are skipping your big, centralized search engine and instead going to a thousand different niche, AI-powered sites to answer questions, generate ideas, find useful images, and create new things, then you can’t have a business selling ads against an enormous pool of search queries.
Big Tech needs big-time scale to make its business models work, so if you take away that scale then you kill Big Tech.
( It’s also within the realm of possibility that Big Tech knows something about the potential scariness of AI that the public doesn’t — this was the subject of yesterday’s newsletter. But “follow the money” > “hunt for Bigfoot,” so I think the business model threat is more compelling as an explanation for what’s going on than the AGI threat. I do entertain that both could be true, though.)
Democracy is the real threat
🔭 I said above that I should’ve seen all this coming — I should’ve predicted that Google would eventually get into the “LLMs are dangerous” game and press regulators to take a close look at these models.
I say this because one thing has been clear to me since this newsletter’s first post: the AI ethics crew that makes a living warning about how toxic and bad LLMs are quite conspicuously absent from the anti-trust and anti-Big Tech discussions that are happening on both the left and the right.
As I’ve pointed out on Twitter (and I think in this newsletter, though I can’t find a good example to link) they like big, centralized pools of power because these offer single points of control to groups that can position themselves the right way.
These types also don’t like democracy because:
It’s not a reliable guarantor of rights and protections for vulnerable minorities, and
Virtuous, educated elites always know better than the benighted mob how society should be run.
🤨 What’s funny and sort of awkward (for me, at least) is that the founding fathers were right there with them on both of the above points. I think it’s true that #1 is undoubtedly correct, and #2… eh, I admit I go back and forth on that one.
I have a thread, linked below, on the woke disdain for democracy with some examples in it. There are two incidents in the thread where they first held a vote (both times it was on renaming something), and when the vote doesn’t go their way they ignored it and did what they wanted to anyway.
💪 Think about what’s required to pull off these stunts (i.e., hold a vote, then ignore the results if you don’t like them): centralized power where your side has admin privileges.
Now think about a single large ML model locked away behind an API at a BigCo vs. an open-source ML model that anyone can copy and tweak and use as they see fit. The former is centralized power over what billions of users can do with AI, and the latter is decentralized power where nobody has admin rights and can make the rules of AI for everyone else.
Decentralized AI in action
🏃 If you want to know what decentralized AI looks like in action — what all the groups I listed in the first section above are banding together to fight — check out Civitai.com. There are hundreds of customized checkpoint files, which are copies of the Stable Diffusion model weights that have been altered by users via further training runs.
There’s an amazing variety of models on display there — just everything you can imagine. There are models that specialize in generating anime ladies with oversized boobs; models for photorealistic ladies with oversized boobs; models for Pixar-style ladies with oversized boobs; models for redheads with oversized boobs; models for Japanese ladies with oversized boobs; models for sci-fi cyborg ladies with oversized boobs… So many different models and so much variety, it’s a wonder of technology.
🦌 I also see this same amazing variety in the Facebook AI art groups I’m in. There’s post after post featuring women of different art styles and ethnicities with huge breasts. It makes me think about how all those GPUs that went on sale after crypto crashed were finally to a purpose far more important to society than mining bitcoin: training Stable Diffusion to render impossibly large boobs.
😳 Guys, I feel like I was going someplace more optimistic with this section but I got sidetracked by the boob situation, which really is as bad as I’m telling you it is. This will have all kinds of weird implications for still more strange bedfellows in the anti-AI camp. People (like me) who think porn is bad and there’s way too much of it on the internet are going to have to figure out what to do about people (like me) who want to see everyone do whatever they like with open-source image generation models.
I don’t have any answers for this horny dilemma, but if I do come up with one I’ll probably post it over at RETURN.