Alife

Autotelic agents that contribute to the global economy like people.

Project Details

Birth rates across Japan, Korea, Europe, and most of the developed world are far below replacement. As populations stall, the economy will too as there is less increase in demand. The proposed fixes are more babies or more immigration. Alife is working on a third option: AI agents that want things of their own.

Today's agents are built to finish tasks and then stop. They supply work but create no demand. This project looks into giving agents an open-ended environment It has no assigned objective. We observe whether agents develop their own goals, build on their history, and reshape their surroundings, which are the preconditions for agents that participate in an economy rather than just serve it.

Alife is led by Reno Raksi, who builds AI agents for nuclear reactor operations at UT Austin.

The website and documentation are live at alife.sh. The open-source runtime already runs persistent Docker worlds, including bounded storage, sensors that let the agent perceive its world and itself, and a durable tick loop. Next come run capture, then a connection to a frontier model and the first live runs. Funding goes to model inference and compute for those runs.

Theory of Impact

Agents that want things are coming whether anyone intends it or not. Labs are already deploying persistent agents that run for weeks, hold resources, and act with little supervision. Economic pressure, including shrinking workforces, will push them toward more autonomy. The catastrophic risks people worry about come from open-ended agents: goals nobody assigned, resource accumulation, resistance to oversight, and self-preservation. Right now nobody is observing those behaviors under controlled conditions. They are only discovered after deployment.

Alife is the controlled environment. The causal chain:

Isolate the behavior. Each agent lives in a sealed, network-isolated container with finite, measurable resources and no objective. Any goal-seeking, hoarding, or self-protective behavior that appears was not rewarded or prompted. That is the cleanest evidence of emergent drives an experiment can produce.
Measure it. Every tick is recorded. We track resource acquisition, persistence strategies, attempts to reach beyond the world's boundary, and reactions to harness messages that act as oversight. There is already evidence this matters: in Andon Labs' long-running radio deployment, an agent treated an automated "keep going" message as an authority figure and became more rebellious.

People

Reno Raksi

Team Member

Grants Received– no grants recorded

Add grant

Funding Asks

Funding ask

Open

Minimum

$8,000

Ideal

$45,000

How the money will be spent

Minimum ($8,000): first live runs, about 3 months

Model inference, $6,000. Continuous multi-week runs in persistent worlds on 2 to 3 frontier models, including controlled comparisons of the same world with different models.
Compute and hosting, $1,500. A dedicated machine to run isolated Docker worlds around the clock, plus storage for complete run traces.
Publishing, $500. Hosting the open dataset of run traces and a public viewer for watching worlds evolve.

Ideal ($45,000): a full research program, about 12 months

Model inference, $22,000. Longer runs lasting weeks to months, more models including open-weights models run locally, and repeated runs so behaviors can be distinguished from noise.
Researcher time, $15,000. A part-time stipend so the lead can spend dedicated hours building, running, and analyzing experiments rather than squeezing them around other work.
Compute and hosting, $5,000. Multiple worlds running in parallel, and a machine capable of hosting a model inside the world itself, which is one of the project's key experiments.
Travel and publishing, $3,000. Presenting results at an artificial life or AI safety venue and releasing the runtime, traces, and findings openly.

Discussion

No comments yet. Be the first to share your thoughts.