Research lab

Toward Level 2–3 synthetic intelligence.

"Machine learning" fits a sub-part of the AI space, but "artificial intelligence" has been stretched to cover simple state machines and hard-coded keyword matchers. This is the ongoing research behind the Ariadne AI Platform — and the honest account of what is solved and what is not.

A working definition

Categories for the intelligence of software.

Our aim is a system that learns dynamically from observation, holds long-term intents, and fulfils them in the lowest-effort, least-impacting way — while watching its own actions and their effects.

Level 0 — Conventional

Normal software and systems. Deterministic behaviour, no learning.

Level 1 — Machine learning

Attempts multiple defined or random solutions and stores successful approaches to satisfy defined goals. Useful for translation, recognition and the like.

Level 2 — Strategic

Generates, simulates and implements strategies to improve conditions or resolve problems — taking previously undefined actions toward long-term goals while storing every outcome.

Level 3 — Self-directed

Able to adjust its own goals. This is the frontier we are working at today.

Architecture

The design so far.

The two hardest components remain the output encoder and the intent engine. The intent engine can simulate a desired change in the model, but deciding the concrete action to realise it — knowing to actually draw the nought or cross — is unsolved.

The observation decoder faces the mirror problem: it can receive state about a thing, but must be told it is looking at, say, a 2-D board. For now we use a lexical analyser in that seat so progress isn't blocked. It's a problem many academics are tackling; we assume a general solution will arrive, and we'll revisit it then.

Synthetic intelligence architecture diagram

R&D progress

Where each subsystem stands.

Input

Tools like TensorFlow model observed data — e.g. an object at a location.
Done: chat-based input.

Generalisation

Deep recursive, time-variant pattern recognition with Purkinje-style suppression reinforcement.
Done: tokenisation providing a forwards-only strategy.

Reasoning

Logical extraction of correlating, distinct and opposing patterns.
To do: the intent engine.

Memory

Key–value pairs and lambda expression trees.
Vector memory database.

Intuition & generation

Done: transformer generators.
Automatic Bayes classifier; model simulator.

Goal-seeking & output

Lowest-cost, time-restricted pathing across a gradient map of action costs.
Temperature-based text generation with loop prevention.

Capabilities

From lab to product

This research ships as Ariadne.

The architecture on this page is what powers the Ariadne AI Platform. See how it comes together in production.

Explore Ariadne → See our GitHub