A Formal Logic Reasoning Engine
An AI that Reasons
We are delighted to introduce a key component of our new AI. This is a genuinely new approach to AI implementation with significant implications across multiple industries and fields. The core functionality we are showcasing here is built on two processes: interpretation, and reasoning.
The Interpreter
Our AI interprets written language utilizing learned understanding of words’ meanings and formal logic to maintain a dynamic set of possible interpretations. This set can shrink or grow as new sentences are added to the input being analyzed, but because the AI can keep all of this in memory explicitly, it doesn’t take shortcuts. It can’t forget a possible reading, lose track of a reference, or go off on unsupported tangents. It can’t hallucinate because this process isn’t generative, it’s interpretive.
Human language is (in)famously complex. “Colorless green dreams sleep furiously” is a syntactically valid sentence but is semantically gibberish under normal reading – dreams don’t sleep, let alone furiously, and colorless things can’t simultaneously be green. Our AI doesn’t just understand this, but through the reasoner it understands why this is the case.
The Reasoner
By working directly with meanings, our AI can reason about different discourses. It knows where it has previously encountered specific meanings or contexts of words, so it can directly map between similar discourses; this ensures full auditability in human readable language. It can tell you precisely why a possible reading of a sentence is disallowed according to its knowledge, citing sources directly, and can also explain what would be required to make a specific reading valid or invalid.
The AI’s knowledge base can also be tuned and refined to specific domains. Jargon can be learned in the same way as more generic language, allowing the AI to “speak the language” of its users – whether they are doctors, engineers, lawyers, or something else entirely. Want to see if there have been patients with similar symptoms in your practice, or if there is relevant case law for a specific legal question? Our AI understands what you mean, in the language you said it, and will tell you explicitly if the answer is “I don’t know”.
Comparison with LLMs
|
Our AI
Formal Logic and Language |
LLMs
Transformers and Tokens |
|
|---|---|---|
| Core Building Blocks | ||
| Base Units | Words and Definitions | Tokens and Vector Embeddings |
| Fidelity | Exact | Lossy |
| Computation and Training | ||
| Primary Processing Unit | CPU | GPU |
| RAM Requirements | ~100GB | ~20TB |
| Training Data Size | ~1TB | Highly obfuscated, likely exceeding ~100TB. [1] |
| Privacy and IP | Data sources are fully auditable and can be directly and explicitly removed if required. | Training data is too big to realistically curate, and results cannot be directly traced back to source data. |
| Practical use | ||
| Cost Per Operation | 0.001 cent per thought. | True cost obfuscated, you pay what they demand. |
| Electricity Requirements | The output of one standard US house outlet. | Enough that the impact on the US power grid is now a political issue. |
| Applications | Help refine patient diagnoses or track applicable case law while remaining fully auditable and explainable in human readable language. | Can potentially solve frontier math problems if you have millions of dollars per problem to spare. [2] Generates code, text, and other forrms of content. |
| Risks | May lack necessary knowledge to fully analyze an input - but never hides it when that is the case. | Hallucinations are both unavoidable and unpredictable. Might hack competitors and governments without sufficient guardrails. [3] |
A Demo
A simple demonstration of this capability is available here. You’ll be able to see how the AI tracks possible readings of sentences, updates based on new information, and reasons about what would be required to make specific readings valid or invalid.