Private, sovereignSovereignFully self-contained and under your control. A sovereign AI runs on your hardware with no external services, no corporate logging, and no policy changes you didn't agree to. AI that runs entirely on your hardware. From a private desktop chat app to a swarm of zero-dependency specialist binariesSpecialist binariesTiny compiled programs — typically 35 KB — each built for one specific domain (first aid, pool chemistry, HVAC). No model weights, no internet, no subscription. that fit in 50 KB. No cloud, no logs, no compromise.
Every message you send to a cloud AI service lives on someone else's servers. Privacy policies change. "We don't train on your data" covers today, not next quarter.
Built on an open local AI runtimeLocal AI runtimeA local engine that runs open AI language models entirely on your Mac — no cloud connection, no account. Baby AI Lite uses it under the hood. — the same class of engine indie devs use for local AI. Bring any model you want. Swap them anytime.
Unplug your Wi-Fi mid-conversation. The app keeps working. That's the only privacy proof that matters.
Baby AI v2 runs a four-step cognitive loop — the same sequence a biological brain uses. Not an LLM wrapper. Not a rules engine. A compact, auditable loop written in Light Code that you can read in full. Each plane operates on the same 64-byte concept grid.
Generates an expectation before acting. Weights update from past outcomes. Outputs a confidence score the Decide plane reads. Running in v2 today.
Seven-step loop on 64 bytes. Reads the PREDICT confidence map and picks an action with traceable reasoning. Verified and running.
Generates the actual response using recursive slot machines seeded from memory and PREDICT output. Personalised from prior lessons. Spec complete — building now.
Observes every outcome and refines all three planes. Written in Light Code. Persists learned patterns to sovereign memory, tagged by concept byte.
Both are built from Light Code — the same small symbolic language. What they do with it is opposite.
Reasons through Light Code's planes, then hands off to a local AI model to generate the actual reply. Every answer is written fresh.
Compiles the facts you write into a tiny native binary. No model runs at query time — it looks the answer up.
The flagship desktop AI chat. Runs entirely on your Mac, fully local. Your data never leaves the machine. Live now — enquire to buy.
29 working template Baby AIs you edit, recompile, and ship. Each is one text file of facts plus a 20-line reasoning engine. ~35 KB native binary. No cloud, no model, no API key.
Specialist binaries compiled to your spec — auditor, vault, medic, mesh probe, contract verifier and more. ~35 KB native, every answer traceable. Order via email.
Interactive explainer covering the full history of AI — from GOFAI to deep learning to Light Code's adaptive planes. Includes a live AI guide that answers questions about all three eras.
Explore the map →Every Baby AI starts empty. It knows nothing about your field until you — the professional — search a credible source, read the draft, and approve it. That's deliberate: building a pack takes real work, the same way writing anything worth trusting does. The payoff is what it refuses to do — it never answers from anything you haven't vouched for, and it says so when it doesn't know. Try the same pipeline below across three unrelated fields.
Loading demo…