Chain of Thought: How 4chan Gamers Discovered AI's Most Hyped 'Reasoning' Breakthrough
Key Takeaways
- ▸4chan gamers discovered chain of thought in 2020 while playing AI Dungeon with GPT-3, a year before Google claimed to be first
- ▸Chain of thought—asking models to explain step-by-step reasoning—improves accuracy on certain problem types and is now a cornerstone of AI company marketing
- ▸AI companies use human-like language ("thinking," "reasoning," "planning") to describe LLM outputs, while online communities describe the same technology in more grounded, technical terms
Summary
A new investigative article reveals that 4chan gamers discovered "chain of thought"—now celebrated as a revolutionary AI reasoning breakthrough—in July 2020 while playing AI Dungeon, a text-based RPG powered by OpenAI's GPT-3. The players noticed that when they asked game characters to solve math problems with step-by-step explanations, the model performed better and responded in character, posting their findings on Twitter more than a year before Google researchers claimed to be "the first" to elicit chain of thought from a general-purpose LLM.
The discovery has since become central to industry hype around "reasoning models," with OpenAI's o1 model and Google's Gemini 2.0 Flash Thinking marketed as systems that "think" before answering. However, the article highlights a stark contrast between how 4chan communities describe AI capabilities—in accurate, technical terms—and how AI companies market their products, often anthropomorphizing LLMs by claiming they "plan," "generalize," and "think." The piece suggests that despite the industry's sophisticated framing, early internet communities understood the technology's actual mechanics more honestly.
- The rebranding of chain of thought as "reasoning" reflects industry efforts to create hype around incremental improvements rather than fundamental breakthroughs
Editorial Opinion
The chain of thought discovery exemplifies how grassroots experimentation often precedes formal academic credit in AI research. More concerning is the widening gap between technical reality and corporate messaging—while 4chan users accurately characterized GPT-3's limitations, the industry increasingly sells deterministic pattern-matching as "thinking" and "reasoning." This semantic inflation risks misleading both investors and the public about the nature of current AI capabilities and may obscure what actual progress in machine reasoning would look like.
