In July 2020, 4chan’s video-game discussion board looked much like the rest of the notorious online forum. There were elaborate, libidinal fantasies involving “whores” and “dragon cum,” and comments on how long a gamer had to wait “before my dick can get up for another beating,” as one put it.
And yet, as the gamers discussed such things, they were also making a discovery of significance to the AI industry. Some of them were playing AI Dungeon, a new text-based role-playing game that was essentially an AI version of Dungeons & Dragons. In endlessly generated fantasy-world scenarios, players described actions like “pick up the sword” or “tell the troll to go away,” and the computer responded with the action that followed.
In addition to asking the game’s characters to engage in various sex acts (naturally), the 4chan gamers also asked them to do math problems. That sounds strange, of course, but AI Dungeon was powered by OpenAI’s GPT-3, and the gamers knew that they were among the first people to probe the capabilities of this new large language model. This was more than two years before the release of ChatGPT, and the model was famously bad at math. It frequently failed at simple arithmetic. But when they asked a character in the game to do a math problem and provide a step-by-step explanation, one of them wrote, the LLM was “not only solving math problems but actually solves them in a way that fits the personality of the fucking character.”
The players had come upon a new feature—what’s known in AI today as “chain of thought.” Essentially, it means that the model explains the steps required to solve a problem, in addition to giving an answer. Asking the model for a chain of thought also seems to improve the accuracy of its answers to certain kinds of problems. The gamers on 4chan recognized the significance immediately, and posted examples on Twitter.
Recently, the tech industry has promoted chain of thought as a revolution in technology, and a reason to get excited about AI all over again. Researchers at Google claimed in a paper to be “the first” to elicit a “chain of thought” from a general-purpose LLM, more than a year after the 4chan gamers shared their findings. (This claim was removed from subsequent versions of the paper, which still did not acknowledge the gamers, though at least one other research paper has.) And in the past couple of years, companies have begun to claim that their chatbots are not just getting math problems right; they are actually thinking about them. OpenAI wrote in 2024 that its “o1” model “thinks before it answers,” and Google claimed that Gemini 2.0 Flash Thinking Experimental was “capable of showing its thoughts.” Companies started referring to their models as “reasoning models,” ostensibly a new kind of product from an LLM.
Amid all this hype, the 4chan history is instructive. 4chan gamers, for all their brash language, have tended to speak in more levelheaded—and accurate—terms than the AI industry about how the models work. Last year, for example, Anthropic published a long and serious-looking article, “On the Biology of a Large Language Model.” Its visual presentation mimicked scientific publications, with sophisticated-looking diagrams and equations. But on every topic, the article described the operation of the LLM in terms of a human mind. It said the LLM “plans” its writing in advance, “generalizes” its knowledge, and can be “unfaithful” to its chain of thought (meaning, the article explains, the LLM is occasionally “bullshitting”).
Contrast this with a guide written in 2024 by people on 4chan, which begins with the heading, “Your bot is an illusion,” and proceeds with a clear, detailed description of how companies use an LLM to construct a chatbot that responds to questions and has a personality. It describes an LLM’s most important technical features and shows how the model’s outputs correspond to its various inputs. The guide is a useful reminder of the most basic truth about large language models: The only thing they can do is imitate their training data.
LLMs can output explanations of math because they were trained on explanations of math. Some of those explanations come from textbooks, but companies also train their so-called reasoning models on text that conveys the act of thinking. I dug into some open-source AI-training data sets and found hundreds of thousands of meandering solutions to math problems that included language such as “Wait, no. The question is,” “First, I should parse the input correctly,” and “Wait, but in cases where …” As far as I’ve seen, companies acquire this text either by paying workers to write it or generating it with other AI models. (Google, OpenAI, and Anthropic did not respond to requests for comment.)
Models trained on such utterances are not actually reasoning; they are predicting what reasoning might look like. There isn’t even necessarily any connection between a model’s reasoning steps and its final answer. Researchers have shown that models can provide incorrect chain-of-thought text but still arrive at the correct result.
Some people have argued that if a computer can imitate human reason well enough to fool us every time, then how can we say it isn’t doing the real thing? Researchers at Apple have explored this question, and their findings are insightful. For example, they discovered that a model might answer a math word problem correctly, but then answer the same problem incorrectly after the wording was changed slightly. Specifically, they found that state-of-the-art reasoning models performed up to 65 percent worse when irrelevant information was added to a question, even when the wording of key facts was left unchanged. Apple researchers have also shown, in a paper titled “The Illusion of Thinking,” that although the reasoning models do better than standard LLMs on certain problems, they are also worse at others.
The reason the chain-of-thought trick does often work is fairly simple. The additional words in the chain of thought give the model more context, which guides its word-predicting process in a better direction, as Perplexity CEO Aravind Srinivas explained in a 2024 interview. This is analogous to the common advice about being specific when asking an LLM a question on any topic. The more details you give, the more you push the LLM toward the relevant words in its memory.
Some of the 4chan gamers appeared to understand this immediately. As one explained back in July 2020: “It makes sense since it is based on human language that you have to talk to it like one”—that is, like a human—“to get a proper response.”
In addition to the gamers, another AI enthusiast discovered the chain-of-thought trick at almost the exact same time. A computer-science student named Zach Robertson, who also came to GPT-3 through AI Dungeon, wrote a blog post in July 2020 about “how to amplify GPT3’s capabilities” by breaking math problems into multiple steps. That September he gave a presentation that showed how the steps could be “chained” together. Robertson, who is now a Ph.D. student in computer science at Stanford, told me on a video call that he was not aware of the 4chan gamers. In fact, he wasn’t even aware he could be considered a co-inventor of chain of thought. I’d seen his blog post cited in a research paper, but when I first mentioned it in an email, he was unsure what I was talking about. He’d removed the post from the internet a couple of years ago when migrating his blog to a new site. (He restored it after we spoke.)
I thought Robertson might be proud to learn he was a pioneer in an area of such enthusiasm within the AI industry. But he seemed only mildly tickled. Those early experiments with AI Dungeon were what got him interested in AI, he told me, but he’s since moved on to other topics. Chain of thought was a remarkable trick, but that’s also all it was.
The post The Strange Origin of AI’s ‘Reasoning’ Abilities appeared first on The Atlantic.
