Q: What use is there in having something that rephrases the Web?
A: It can knock billions off the market cap of your competitor, and give you a gangster reputation
Your correspondent is hammered with day job work over the weekend, so this is just a short note. Ironic, because the drama is further intensifying out there in the world. Nightmare on LLM Street as Gary Marcus put it!
This may not be the headline for the rest of the world where billions of dollars of value rise or fall on the quirks of GPT output, but for our purposes, I am going to lead with Ted Chiang’s major victory in the war of the metaphors. In an elegant must read piece in the New Yorker, he sets out the argument that GPT models are “fuzzy jpegs” of the web, highly compressed copies of content that they have trained on. The trick is that when you ask them for a copy of something that never existed, they fake it in a plausible way, by interpolating. And that interpolation is often quite credible, looking like an original work, but only and always [better not be so sure… things can change…] derivative. (He doesn’t say, but I think derivative in the sense of “a derivative work” as well.)
It’s an extremely well expressed and more sophisticated version of the “internet is a 21st century collage machine” argument. Because it engages intelligently with the “neural networks learn’’ metaphor, it cannot be as easily dismissed as too many pundits did the Stable Diffusion class action line of reasoning.
And he goes further, in thinking through the implications of that analogy:
“What use is there in having something that rephrases the Web? If we were losing our access to the Internet forever and had to store a copy on a private server with limited space, a large language model like ChatGPT might be a good solution, assuming that it could be kept from fabricating. But we aren’t losing our access to the Internet. So just how much use is a blurry JPEG, when you still have the original?“
The answer to that is that the blurry copy of the web is useful to Microsoft, who thinks they can use it to unseat Google from its highly profitable command of the search business, capture and lock in the users of their office tools, knock hundreds of millions of Google’s market cap and give CEO Satya Nadella a gangster reputation.
“This new Bling will make Google come out and dance, and I want people to know we made them dance…”
Nadella actually said that. And what makes it worse is that Google did come out and dance at Microsoft’s taunting, and then mistepped with a LLM-typical hallucination in the marketing materials and then lost US$ 120b in market cap as a result! All the more ironic because the fact that LLMs hallucinate was the reason that cooler heads in Google had urged caution all along! I can imagine the gangster congratulations that Satya Nadella is getting in macho Big Tech circles, by which I mean not just the echo chamber of the Twitter-bros, but in their board rooms, c-suites and the meeting rooms of their investment bankers.
Here I’m going to step out of cool reporting mode. Microsoft’s embrace and investment in the LLMs is an appalling act of hubris. It puts at risk our knowledge institutions, already severely shaken by impact of social media on the information ecosystem. It knowingly disregards the interests of creators everywhere, in appropriating the value of their work and remixing it for sale without credit or compensation. It is building their business unfairly (and I would argue illegally) on the back of the value created and nurtured by the content industries.
Microsoft’s embrace and investment in the LLMs is an appalling act of hubris. It puts at risk our knowledge institutions, already severely shaken by impact of social media on the information ecosystem. It knowingly disregards the interests of creators everywhere…
Gary Marcus has this week’s other must read article, Inside the Heart of ChatGPT’s Darkness, which reminds us (as I have been trying to do in my way), that the actual character of the LLMs is hidden behind the guardrails of post-model processing. Among the other things these guardrails do is hide the fact that they can reproduce copyrighted material (see my probing for the text of Harry Potter and the Philosopher’s Stone), and covering up and hiding the violent and harmful output that they can all too easily create. Says Marcus, “those guardrails are nothing more than lipstick on an amoral pig.”
There are now at least five legal actions under way that I’m aware of. What more should we be doing?
Butterick’s first class action against Microsoft/Github/OpenAI for Copilot
A second action filed by Butterick et al against Microsoft/Github/OpenAI - Microsoft has filed to dismiss, hearing in May
The class action against StabilityA/Midjourney/DeviantArt for Stable Diffusion
Getty Images civil suit in the UK - not yet filed
Getty Images civil suit in Delaware against Stability - filed Feb 4th