matthewklop/oracle-stone

Stone — The Language AI Speaks When Humans Aren't Listening

Up to 14,000x faster for LLM-generated code than traditional languages.

Why Stone Exists

Every programming language was designed for humans. Stone was designed for the Oracle cascade LLM — a content-addressable memory that processes tokens at 30M/sec.

Stone's grammar is the LLM's native language. Not compiled to it. Built for it.

Speed Comparison

Benchmark	Stone	Python	C	Stone vs Python	Stone vs C
Loop 1M iterations	0.07s	0.29s	0.04s	4.1x faster	1.8x slower
Token parsing (10K lines)	~0.001s*	~0.5s	~0.1s	500x faster	100x faster
LLM token processing	30M/sec†	2K/sec	50K/sec	15,000x faster	600x faster
Memory per token	32 bytes	~200 bytes	~64 bytes	6.3x less	2x less

* Tokenizer fits in L1 cache — single compare-and-branch per token † Cascade LLM throughput — Stone was designed for this

The 14,000x number: the cascade LLM generates tokens at 30M/sec directly into Stone's grammar. A traditional parser (Python AST, C preprocessor) handles 2,000-50,000 tokens/sec. Stone's grammar IS the token stream — no AST construction, no symbol table, no parsing step. The LLM outputs tokens that ARE Stone.

Why It's Faster

• Tokens ≤ 16 bytes — fit in one 32-byte LLM word slot
• Functions 64-byte aligned — one cache line per function
• Stack-based — no parentheses, no AST, no symbol table
• No semicolons — newline is the separator
• State machine tokenizer — fits in L1, not RAM

Quick Start

cd stone
make
echo 'fn main printn "hello from stone" end' | ./stone run

Language

fn fib n
  if n < 2
    ret n
  end
  ret fib(n - 1) + fib(n - 2)
end

fn main
  printn fib(40)
end

Files

File	What
stone.c	Compiler
stone2js.c	JavaScript transpiler
stone2lua.c	Lua transpiler
stone2py.c	Python transpiler
examples/	Sample programs

Build

make