A few months ago, while meeting with an AI executive in San Francisco, I noticed a strange sticker on his laptop. This sticker featured a cartoon of a terrifying octopus-like creature with a yellow smiley face attached to his many eyes and one of his tentacles. I asked what it was.
“Oh, those are Shoggoths,” he explained. “This is the most important meme in AI.”
This officially derails our agenda. Forget chatbots and compute clusters. I needed to know all about Shoggoth, what it meant, and why people in the AI world were talking about it.
The executive said shoggoths have been used in the field of artificial intelligence as a vivid visual metaphor for how large-scale language models (the type of AI systems that power ChatGPT and other chatbots) actually work. I explained that it has become a popular reference among working people.
But he said it was partly a joke and also hinted at the trepidation many researchers and engineers have about the tool they were building.
Since then, Shoggoth has gone viral, or as viral as it can be in the tiny world of hyperonline AI insiders. This is a popular meme on AI Twitter ( now deleted tweet by Elon Musk), a recurring trope in AI risk essays and message board posts, and a useful little abbreviation in conversations with AI safety experts. NovelAI, one of his AI startups, you said you named it recently Shoggy, a cluster of computers that pays homage to memes. Another of his AI companies, Scale AI, has designed a series of tote bags featuring shoggoths.
Shoggoths are fictional creatures introduced by science fiction author H.P. Lovecraft in his 1936 novel At the Mountains of Madness. According to Lovecraft, the shoggoth was a gigantic blob-like monster of iridescent black mucus, covered with his tentacles and eyes.
Shoggoth landed in the AI world in December, one month after the release of ChatGPT. At that time, Twitter user @TetraspaceWest responded to a tweet about GPT-3 (his OpenAI language model, the predecessor of ChatGPT) with an image of two hand-drawn shoggoths (the first shoggoth). The first is labeled “GPT-3” and the second is labeled “GPT-3 + RLHF”. The second Shoggoth was perched on one of the tentacles and had a smiley face mask.
In a nutshell, the joke was that in order to prevent AI language models from behaving in horribly dangerous ways, AI companies need to train their language models to behave politely and harmlessly. . One popular way to do this is called “reinforcement learning from human feedback” (RLHF), a process that asks a human to score a chatbot’s response and feeds that score back into her AI model. .
Most AI researchers agree that models trained with RLHF perform better than models without RLHF. However, some argue that fine-tuning the language model in this way does not actually make the underlying model less weird or cryptic. In their view, it’s just a flimsy, friendly mask that obscures the mysterious beast beneath.
The meme’s creator, @TetraspaceWest, told me in a Twitter message that shoggoths “represent something that thinks in a way that humans cannot comprehend, and that’s completely different from how humans think.” .
@TetraspaceWest said that comparing an AI language model to a shoggoth does not necessarily mean that it is evil or perceptive, it just might be unaware of its nature. said.
“I was also thinking about how some of Lovecraft’s most powerful beings are dangerous. Not because they don’t like humans, but because they are indifferent and their priorities are us.” Because it’s completely alien to and doesn’t involve humans, and that’s what I think “be true about the potential for powerful AI in the future.”
As AI chatbots grew in popularity and users began noticing that some chatbots seemed to do strange and inexplicable things that their creators hadn’t intended, the image of shoggoths became widespread. rice field. In February, when Bing’s chatbot stopped working and was about to destroy my marriage, an AI researcher friend of mine congratulated me for “getting a glimpse of the shoggoth.” A fellow AI journalist of his joked about the Bing tweak that Microsoft forgot to put on the smiley face mask.
Eventually, AI enthusiasts expanded on this trope. Created in February by Twitter user @anthrupad version Shoggoth photo with a smiley face labeled “RLHF” plus a more human face labeled “supervised tweaks”. (You actually need a computer science degree to understand the jokes, but this is just a brief explanation of the difference between his AI language model in general and more specialized applications like chatbots.)
When you hear mention of shoggoths in the AI community today, you might notice the strangeness of these systems: the black-box nature of their processes, their apparent defying human logic. Or maybe it’s a joking visual shorthand for a powerful AI system that looks questionably great. If an AI safety researcher is talking about shoggoths, she may be passionate about preventing her AI systems from exhibiting their true shoggoth-like qualities.
Either way, Shoggoth is a powerful metaphor that encapsulates one of the strangest facts about the world of AI. The thing is that many of the people working on this technology are somewhat confused by their creations. They don’t fully understand the inner workings of AI language models, how they acquire new features, and why they sometimes behave unpredictably. They aren’t entirely sure whether AI will be pure good or pure evil for the world. And some of them even got to play around with a version of this technology that hasn’t yet been sanitized for general consumption: real shoggoths that aren’t masked.
It is unusual by historical standards for some AI insiders to even jokingly call their work Lovecraftian horror. (Speaking of which, 15 years ago Mark Zuckerberg wasn’t comparing his Facebook to Cthulhu.)
And this reinforces the idea that for some AI participants, what is happening in AI today feels more like an act of subpoena than a software development process. They want to create a puffy alien shoggoth, make it bigger, more powerful, and have enough smiles to cover up the scary parts.
