A few months ago, while meeting with an AI executive in San Francisco, I saw a strange sticker on his laptop. The sticker depicted a cartoon of a menacing octopus-like creature with many eyes and a yellow smiley face attached to one of its tentacles. I asked what it was.
“Oh, that’s the Shoggoth,” he explained. “It is the most important meme of AI”
And with that, our schedule was officially derailed. Forget chatbots and computer clusters – I needed to know everything about Shoggoth, what it stood for, and why people in the AI world were talking about it.
The executive explained that Shoggoth had become a popular reference among AI workers, as a vivid visual metaphor for how a great language model (the kind of AI system that powers ChatGPT and other chatbots) really works.
But it was only partly a joke, he said, because it also hinted at the anxieties many researchers and engineers have about the tools they’re building.
Since then, the Shoggoth has gone viral, or as viral as it gets in the small world of hyper-online AI experts. It’s a popular meme on AI Twitter (which includes a tweet now deleted by Elon Musk), a recurring metaphor in essays and message board posts about AI risk, and a handy bit of shorthand in conversations with AI safety experts. An AI startup, NovelAI, said he recently named a group of computers “Shoggy” in homage to the meme. Another artificial intelligence company, Scale AI, designed a line of handbags featuring the Shoggoth.
Shoggoths are fictional creatures, introduced by science fiction author HP Lovecraft in his 1936 novel “At the Mountains of Madness.” In Lovecraft’s narration, Shoggoths were huge, blob-like monsters made of an iridescent black goo, covered in tentacles and eyes.
The Shoggoths landed in the AI world in December, a month after the release of ChatGPT, when Twitter user @TetraspaceWest responded to a tweet about GPT-3 (an OpenAI language model that was the predecessor to ChatGPT) with a image of two hand-drawn Shoggoths: the first labeled “GPT-3” and the second labeled “GPT-3 + RLHF”. The second Shoggoth had, perched on one of its tentacles, a mask of a smiling face.
Simply put, the joke was that to prevent AI language models from behaving in scary and dangerous ways, AI companies had to train them to act in polite and harmless ways. One popular way to do this is called “reinforcement learning from human feedback,” or RLHF, a process that involves asking humans to rate the chatbot’s 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 adjusting a language model in this way doesn’t actually make the underlying model any less strange and inscrutable. In his opinion, it’s just a flimsy, friendly mask that obscures the mysterious beast underneath.
@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 and is totally different from the way humans think.”
Comparing an AI language model to a Shoggoth, @TetraspaceWest said, did not necessarily imply that it was evil or sentient, just that its true nature might be unknowable.
“I was also thinking about how Lovecraft’s most powerful entities are dangerous, not because they don’t like humans, but because they’re indifferent and their priorities are totally foreign to us and don’t involve humans, which is what I think it will be. be true about possible powerful AI in the future”
The image of Shoggoth caught on as AI chatbots became popular and users began to notice that some of them seemed to be doing strange and inexplicable things that their creators didn’t intend. In February, when the Bing chatbot went berserk and tried to break up my marriage, an AI researcher I know congratulated me on “catching a glimpse of the Shoggoth.” A fellow AI journalist joked that when it came to tweaking Bing, Microsoft had forgotten to put on its smiley face mask.
Finally, AI enthusiasts expanded on the metaphor. In February, Twitter user @anthrupad created a version of a Shoggoth that had, in addition to a smiling face labeled “RLHF”, a more human face labeled “supervised fine tuning”. (You pretty much need a computer science degree to get the joke, but it’s a reference to the difference between general AI language models and more specialized applications like chatbots.)
Today, if you hear mentions of Shoggoth in the AI community, it may be a nod to the strangeness of these systems: the black-box nature of their processes, the way they seem to defy human logic. Or maybe it’s a joking visual shorthand for powerful AI systems that look suspiciously nice. If it’s an AI security researcher talking about Shoggoth, maybe that person is passionate about preventing AI systems from showing their true Shoggoth-like nature.
In any case, Shoggoth is a powerful metaphor that sums up one of the strangest facts about the world of AI, which is that many of the people who work on this technology are somewhat baffled 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 are not entirely sure if AI will be net good or bad for the world. And some of them have come to play with the versions of this technology that have not yet been sanitized for public consumption: the true unmasked Shoggoths.
That some AI experts refer to his creations as Lovecraftian horrors, even as a joke, is unusual by historical standards. (Put it this way: Fifteen years ago, Mark Zuckerberg didn’t compare Facebook to Cthulhu.)
And it reinforces the notion that what is happening in AI today feels, to some of its participants, more like a convening act than a software development process. They are creating the alien Shoggoths, making them bigger and more powerful, and hoping there will be enough smiling faces to cover up the scary parts.