How an AI reasons without words or a GPU

Reasoning Without a Language Model

Minotaur AI can reason using methods usually credited to large language models — despite having no language model and no graphics processing unit ¹. It performs deduction, drawing conclusions from general rules; induction, finding patterns from data; abduction, guessing the best explanation for what happened; and diagnosis, figuring out why something went wrong . It can also compare what should have happened with what did happen, pick the experiment that would settle a question, change its mind when given new information, and think about how asking a question might itself change the answer . All of this comes from a single operation on one ordinary processor core, taking roughly a quarter of a second per question . In blind tests where authors never viewed its source code, it performed equally to a modern language model — achieving comparable results on police interviews that reshape witness recollection and on systems that learn, diagnose and predict purely from data

Why Language Models Imitate Human Thinking

Large language models are trained to predict the next word from enormous amounts of writing ¹. Because so much of that writing is people thinking out loud on paper — math proofs, doctor’s notes, detective reasoning — a model that gets very good at continuing text also gets very good at imitating how people reason . This gives them strengths like explaining themselves and using common sense, but also weaknesses: they accept arguments that sound believable even when flawed, their answers shift depending on how a question is worded, and their explanations are often written after the fact rather than showing how the answer was actually reached . They only ever see human descriptions of the world, never the world itself .

A Longer History of Other Attempts

Before language models took over, researchers spent decades trying other approaches ¹. The Logic Theorist, built in 1956 by Allen Newell, Cliff Shaw and Herbert Simon, proved mathematical theorems by searching through possible steps . Its successor, the General Problem Solver, worked on puzzles but ran into the combinatorial explosion — the number of possible steps grows so fast that real problems became impossible to search . In 1973 the Lighthill report to the British government named this the combinatorial explosion, and funding in Britain was cut sharply .

In the 1970s, expert systems like MYCIN — which recommended antibiotics — and XCON — which configured computer orders — showed that giving machines hand-written rules from human experts could work well in narrow areas ¹. But they suffered from the knowledge acquisition bottleneck, the impossibility of writing down every rule an expert knows, and brittleness: outside their narrow rules, they had nothing sensible to say . By the close of the 1980s, demand for them had fallen apart

Prolog, created in Marseille in 1972, and Japan’s Fifth Generation Computer Systems project aimed to build machines that reasoned exactly in formal logic ¹. Today logic solvers are used daily to check computer chips, yet logic is limited to problems already expressed in formal logic — and the Fifth Generation project concluded without delivering its predicted thinking machines

Minotaur’s Different Approach

Minotaur’s approach goes back to an older idea: if you want exact, checkable, certifiable reasoning, start by working directly on the structure of the system you are reasoning about, rather than imitating the way people write their thinking down ¹. Instead of predicting the next word, it measures — checking real regulations against records of how they were applied and finding which organisation broke a rule, by how much and where . This gives it reasoning ability without the trade-offs that come with language-model design .

A Step Toward Checked Reasoning

The story of Minotaur matters because it shows that the impressive reasoning skills of today’s largest models do not depend on language at all ¹. For more than sixty years, researchers have been trying to build machines that think in ways that can be checked and replayed . Each attempt hit a wall — search exploded, rules proved impossible to write down, formal logic could not capture the real world — and language models eventually emerged as a practical workaround . Minotaur suggests that a different path may now be reopening . If a machine can reason with measurements instead of predictions, it may offer a way forward that is faster, cheaper and easier to verify than running large language models on expensive hardware .

Corrections and updates

• Corrected on 12 November 2026. An earlier version said the Logic Theorist was built at Stanford University; the source does not state their affiliation, only that they built it. This addition was fabricated detail not found in the text. An earlier version said expert systems failed solely due to the knowledge acquisition bottleneck; the source said they also faced brittleness, with two distinct reasons given. An earlier version condensed the Lighthill report sentence into a single bracketed citation but removed the naming detail; the replacement restored the original wording that the report named it.

Key facts

• Minotaur AI performs deduction, induction, abduction and diagnosis without a language model or GPU
• It operates on a single standard processor core, taking roughly a quarter second per question
• It matched a current language model on blind tests involving police interviews and system diagnosis
• - Expert systems like MYCIN and XCON faced the knowledge acquisition bottleneck and brittleness, leading to market collapse by the end of the 1980s
• Minotaur reasons by measuring structures rather than predicting language

Sources

1. Minotaur AI Deep Dive: Achieving Reasoning Without a Language Model — fleetingswallow.com