AxiomGPT: Programming with LLMs by defining Oracles in natural language

AxiomGPT is an experimental programming model that treats large language models as latent semantic engines.

Instead of writing traditional functions, you define Oracles, which are embedded entities described in natural language that behave like callable, composable functions.

For example:

(defn fibber (Oracle "Return the nth Fibonacci number"))

(fibber 123) ; => 22698374052006863956975682

LISP isn't required at all, but it helps for composition :-)

The model doesn't "execute" logic, it remembers, manifests and resolves the structure behind the intent as oracles.

Oracles can represent:

Algorithms (sorting, TSP, prime factors, map/reduce)

Concepts (e.g. "the government", "my refrigerator")

Personas (e.g. Arthur C. Clarke, a Victorian engineer)

Higher-order or recursive behavior collapsed into linewr time execution.

You can pass them, compose them, gate them, even reflect on them using meta-Oracles.

The full write-up includes usage, theory (embedding vectors as symbolic computation), and a comparison to quantum-like behavior: latent oracles act like entangled functions, invoked by attention.

It's strange, fun, and surprisingly powerful.

You can try it out yourself in minutes with handy provided prompts.

Full Codex and examples: [https://x.com/chrisbe1968/status/1906875616290365941]

Happy to discuss or answer questions!