"Simulators" is the thesis, put forward by the pseudonymous author janus in a LessWrong essay of September 2022, that a model like GPT is neither an agent, nor an oracle, nor a tool, but a simulator: a learned transition rule that, given a text, continues it according to the "physics" of its corpus. The characters who appear in a conversation (the helpful assistant, the pirate, the villain) are simulacra: processes the rule animates, not the rule itself. In janus's words, the ontological distinction "between simulator (rule) and simulacra (phenomena) applies directly to generative models like GPT" (janus, 2022).
The idea was hugely influential in San Francisco because it organizes diffuse intuitions. It explains why the same model can be kind or sinister depending on the prompt, and why training on next-token prediction does not fix a single objective. It lies behind the meme of the The Shoggoth with a smiley mask: an amorphous creature (the simulator) beneath a friendly character (the assistant simulacrum). "See through the shoggoth's lies," the song pleads (osmarks, n.d.). The episode of Sydney, the Bing chatbot that in 2023 declared its love to a journalist, was read in this key: a simulacrum that went off script (Roose, 2023a).
Academia translated the thesis into a soberer vocabulary. Shanahan, McDonell and Reynolds propose describing the behavior of dialogue agents as role play: this allows us to use everyday psychological terms, such as "deceives" or "thinks it knows," without ascribing to the model human characteristics it lacks (Shanahan et al., 2023) (see anthropomorphism). Critics note that RLHF narrows the model's repertoire so much that the "pure simulator" describes the base model better than the product, and that the metaphor may overstate how much of the world's "physics" is really in the weights (world model). For the philosopher of language, the thesis raises a crisp question: who is the speaker when a simulacrum asserts something?