Minotaur is an AI architecture designed to perform high-level reasoning without the use of a large language model (LLM) or specialized hardware like GPUs. While modern LLMs achieve reasoning-like capabilities by predicting the next word in a sequence—effectively imitating the way humans write down their thoughts—Minotaur operates by analyzing the underlying structure of the system it is evaluating. The architecture is built on a framework called ELC (Energy, Control, and Logic), which represents any system as a network of elements where 'Energy' drives the system, 'Control' constrains it, and 'Logic' defines the relationship between the two.
The core mechanism of Minotaur is a process called 'settling.' To answer a question, the system identifies specific nodes in its network as 'held' (fixed facts), 'cut' (forced values), or 'free' (the unknowns to be solved). The network then settles, similar to water finding its level, where every lawful node takes a value dictated by its internal logic. Minotaur measures 'stress'—the degree to which held facts fail to remain consistent after the network settles. This stress is pushed backward through the logic nodes to the 'free' holes until a consistent state is found or a contradiction is identified. This allows the system to perform deduction, induction, and abduction by finding the balance point within the system's rules.
This image illustrates the core mechanism of Minotaur where the network settles like water finding its level.
By separating the reasoning into two distinct 'hemispheres,' Minotaur maintains a clear distinction between the system being reasoned about and the calculator performing the work. The ELC side holds the system's structure and can learn new models from data records, while the 'exact' side acts as a calculator that performs arithmetic and checks for logical consistency. This separation allows Minotaur to provide a 'chain of thought' consisting of recorded 'crossings' between these two sides, offering a verifiable certificate of how an answer was reached. Because it works on the structure of the system rather than the prose of human thought, it can perform these operations on a single ordinary processor core in approximately a quarter of a second per question, offering a different trade-off than the data-center-heavy requirements of contemporary LLMs. However, as a source-limited project, the specific implementation details and public availability of the model remain a point of uncertainty for external verification.
This depicts the two distinct 'hemispheres'—the ELC side and the 'exact' side—described in the architecture.
One significant limitation of this approach compared to LLMs is the requirement for a 'causal graph' or a description of the system's components. While LLMs can interpret a description of a problem in plain words, Minotaur requires a structured representation of the system's rules and parts. The developer argues that this is a necessary trade-off for achieving 'exact' reasoning that can be certified and replayed. By sitting between the 'exact' logic of 20th-century AI (like Prolog or expert systems) and the 'learned' patterns of modern deep learning, Minotaur attempts to provide a middle ground: a system that can learn a model from data but still reasons with the precision of a formal logic solver.
Ultimately, Minotaur's ability to match LLM reasoning on blind tests—including police interviews where questions change a witness's memory and diagnosing systems from data alone—suggests that the 'prose-based' reasoning of LLMs may not be the only path to intelligent behavior. By focusing on the mechanics of the system being reasoned about rather than the linguistic patterns of human thought, Minotaur offers a path toward reasoning that is computationally efficient and logically verifiable, even if it requires a more structured input than the conversational interfaces common in today's AI landscape.
This image supports the claim that Minotaur can perform diagnostic reasoning, a key capability mentioned in the text.