New AI Language

A project folder for representation: how models actually compute, why plain text is a bad long-horizon medium, what we can do without retraining, and what the maths research forces.

Start here

Read agent.md — goals and rules for agents.
Read docs/01_goals_and_scope.md.
Read docs/02_how_ai_works_architecture.md then docs/03_inference_path.md.
Read findings/01_forced_result.md for the punchline.
Use maths/ for theorems and distilled research.

Two tracks (do not mix)

Track	Medium	Saves	Needs
A — I/O	Glyph / STATE-IR (still text)	Prompt tokens	Orchestrator only
B — Engine	AOC / Manifold KV (floats)	KV memory + matmul	Attention hooks / custom cache

Layout

Folder	Contents
`docs/`	Architecture, inference, why text fails, AOC, glyph, roadmap
`findings/`	Forced synthesis, empirics, open questions
`maths/`	Thm0, covering, A0, hybrid, invariants + excerpts
`refs/`	Pointers into the rest of the monorepo

Quick commands (repo root)


# Track A — still text

python glyph_codec.py encode memory/packets.jsonl --out memory/glyph_runs/demo

python codec_tesla_loop.py --n 24 --seed 11



# Track B — AOC MVP (NumPy; needs engine hooks for real models)

python -m new_ai_language.aoc.run_mvp --synthetic --m 256 --d 64

Harvested engine experiments also live under Research'/maths/ (harvested_practical_approx.py, math_kv/).

Deploy docs to Cloudflare Pages (static)


python new_ai_language/build_static.py

npx wrangler login

npx wrangler pages deploy new_ai_language/site --project-name=new-ai-language --commit-dirty=true

Or run .\new_ai_language\deploy.ps1. Details: DEPLOY.md. Output folder: site/.