A few months ago, while meeting with an AI executive in San Francisco, I noticed a strange tag on his laptop. The sticker depicts a cartoon of a ferocious octopus-like creature with many eyes and a yellow smiley face attached to one of its tentacles. I asked what that was.
“Oh, that’s a shoggoth,” he explained. “This is the most important meme in AI”
And just like that, our agenda is officially off track. Forget chatbots and computing clusters – I need to know everything about Shoggoth, what it means, and why people in the AI world are talking about it.
Shoggoth has become a popular reference among AI workers, the executive explained, as a vivid visual metaphor for how large language models — the type of AI systems that power ChatGPT and other chatbots — actually work.
But that’s only part of the joke, he said, because it also hints at the anxieties many researchers and engineers have about the tools they’re building.
Since then, Shoggoth has become popular, or as popular as it can be in the small world of hyper-online AI insiders.Here’s a popular meme on AI Twitter (including now deleted tweet by Elon Musk), a recurring trope in papers and message board posts on AI risks, and some useful shorthand for conversations with AI safety experts. An AI start-up company NovelAI, say it recently named A group of computer “Shoggies” pay homage to the meme. Another artificial intelligence company, Scale AI, has designed a line of tote bags featuring Shoggoths.
Shoggoths are fictional creatures introduced by science fiction author HP Lovecraft in his 1936 novella “Mountains of Madness”. In Lovecraft’s telling, shoggoths are gigantic, blob-like monsters made of iridescent black goo covered in tentacles and eyes.
Shoggoths hit the AI world in December, a month after ChatGPT was released, when Twitter user @TetraspaceWest responded to a tweet about GPT-3 (ChatGPT’s predecessor, the OpenAI language model), with two images of hand-drawn Shoggoths — — The first is labeled “GPT-3” and the second is labeled “GPT-3+RLHF”. The second shoggoth has a smiley mask on one of its tentacles.
In short, the joke is that in order to prevent AI language models from behaving in scary and dangerous ways, AI companies have to train them to behave politely and harmlessly. One popular approach is called “reinforcement learning from human feedback,” or RLHF, and the process involves asking humans to score chatbot responses and feeding those scores back to the AI model.
Most AI researchers agree that models trained with RLHF perform better than models without it. But some argue that fine-tuning a language model in this way doesn’t actually make the underlying model less weird and incomprehensible. To them it was nothing more than a fragile and friendly mask for the mysterious beast beneath.
@TetraspaceWest, the creator of the meme, told me in a Twitter message that Shoggoth “represents something that thinks in a way that humans don’t understand, which is completely different from how humans think.”
Comparing an AI language model to Shoggoth doesn’t necessarily mean it’s evil or sentient, just that its true nature might be unknowable, says @TetraspaceWest.
“I’m also thinking about how dangerous Lovecraft’s most powerful entities are – not because they don’t like humans, but because they’re apathetic, and their priorities are completely foreign to us and don’t involve humans, which is what I think is real about the possible future of powerful artificial intelligence”
The image of Shoggoth became popular as AI chatbots became popular and users started noticing that some of them seemed to be doing weird, inexplicable things that their creators didn’t intend to do. When Bing’s chatbot went unhinged and tried to sabotage my marriage in February, an AI researcher I knew congratulated me on “getting a glimpse of Shoggoth.” An AI reporter joked that Microsoft forgot to put on its smiley face mask while fine-tuning Bing.
Eventually, AI enthusiasts expanded on the metaphor.In February, Twitter user @anthrupad created a Version The image of the shoggoth, in addition to a smiling face labeled “RLHF”, has a more human-like face labeled “Supervised Nudge”. (You actually need a degree in computer science to understand this joke, but this is a repetition of the difference between general-purpose AI language models and more specialized applications such as chatbots.)
Today, if you hear Shoggoth mentioned in the AI community, you might blink at the oddities of these systems—the black-box nature of their processes, and the way they seem to defy human logic. Or it could be a playful visual shorthand for a powerful artificial intelligence system that looks really good. If it’s an AI security researcher talking about Shoggoth, maybe that person is keen to prevent AI systems from revealing their true, Shoggoth-like nature.
Regardless, Shoggoth is a powerful metaphor that encapsulates one of the most bizarre truths in the world of artificial intelligence, which is that many of the people who work on the technology feel a little disorientated by their own creations. They don’t fully understand the inner workings of AI language models, how they gain new capabilities or why they sometimes behave unpredictably. They’re not entirely sure whether AI will be good or bad for the world. Some of them have begun using versions of the technology that have not been sterilized for public consumption – real, maskless shoggoths.
Some AI insiders have called their creation a Lovecraftian horror, or even a joke, unusual by historical standards. (Put it this way: Fifteen years ago, Mark Zuckerberg wasn’t going around comparing Facebook to Cthulhu.)
It reinforces the notion that what happens in AI today is, for some of its participants, more an act of calling than a software development process. They’re creating blob-like alien shoggoths, making them bigger and stronger, and hopefully enough smiley faces to cover up the scary parts.