Farming Intelligence

We believe Artificial General Intelligence is a side effect of survival. Specify an environment and see what pops out; no objectives, no agent engineering.

Our Approach

In artificial intelligence today we carefully engineer an agent's body or neural network's architecture and train it to succeed on an objective or represent the data's structure. We call this "design and train". We propose flipping this paradigm to "specify and discover" where we specify the environment and the materials from which agents can be grown; we only require that they survive in a very general sense so that they must figure out objectives for themselves. We then discover agents with novel and useful traits automatically, some of those take on body / brain pairs that we may never have imagined. In this way we unbound the agent from an objective and engineered brain / body pair.

Environment

A landscape with fruit on stone pillars and fruit encircled in flames.

In this environment agents must gather enough fruit to survive without catching on fire. Their bodies must grow and their brains must learn how best to control their (different) bodies.

Growth and Optimization

Creatures growing bodies to get the fruit.

Agents begin to emerge but most are not successful: many catch on fire, can't grip the fruit, or are too heavy to move.

Discovery

Creatures optimized to get the fruit.

Successful agents develop useful brain / body pairs that enable them to exploit their unique characteristics. For example, springs are tricky to balance on but help agents to jump high to get the food on the pillars.

Early Research

Explore our foundational work on the conditional growth processes that laid the groundwork for intelligence farming: View on GitHub.

This video demonstrates the Stochastic Growth Process (not the more open-ended, task-less environments we describe above). It uses reinforcement learning to grow a body to reach the glowstone block. It provided early validation that we could encode an agent's morphology (body) into a differentiable encoding and efficiently optimize.

Contact Us

Interested in collaboration or learning more about our research?

Email

info@ferret-labs.com

X (Twitter)

@ferretlabs_re