Minotaur: a reasoning architecture with no language model and no GPU
What happened
On October 22, Hacker News' front page covered Minotaur, an AI architecture that uses no language model and no GPU. Built on ELC, it derives answers by measurement rather than prediction. The report says it handles deduction, induction, abduction, diagnosis and belief revision at about 0.25 seconds per problem on a single ordinary CPU core. Blind tests were designed by people who had not seen its code, including a police interrogation scenario where questions can alter a witness's memory, and tasks where the system learns, diagnoses and predicts from data alone. On these, Minotaur's performance was reported as comparable to language models that reason in text.
Written by AI from the coverage · updated 1 hour ago
Coverage
Follow the reports to see the story from different sides.
- Hacker News front pageAchieving AI-style reasoning without an LLM
Minotaur 是一个不依赖语言模型和 GPU 的 AI 架构，基于 ELC 通过测量而非预测来推导答案，在单个普通 CPU 核心上每题约 0.25 秒完成演绎、归纳、溯因、诊断、信念修正等推理。在未见其代码的人设计的盲测中，它在提问会改变证人记忆的警察审讯场景，以及仅凭数据学习、诊断和预测系统等任务上，表现与当前用文字推理的语言模型相当。
Heat over time
Not enough continuous observations to draw a trend yet.