Project summary

The world population growth is declining. Japan, Korea, Taiwan, Singapore, Europe all have TFR scarily well below replacement rates. It's inevitable that as a country progresses, there's less pressure and more burden to have children. A lot of proposed solutions is how we can make raising children attractive again, such as via tax subsidies, or increasing immigration. Population is the backbone of our modern economy. If there is less people, there is less demand, and everything falls apart. My view is that agents can help will replacing the decrease in demand, but they need to "want" for that to happen, requiring open-endedness, which this project is exploring.

What are this project's goals? How will you achieve them?

I'm building an isolated runtime environment of a Linux filesystem, a harness that gives the agent a perception of its world and a continuous heartbeat to perceive and act, and a frontier model as the agent's brain. I want to run this unsupervised environment for 3 months and assess whether current models do anything self-directed at all.

How will this funding be used?

I estimate $6,000 for model inference running for 3 months in persistent worlds on 2 to 3 frontier models, including controller comparisons. Compute and hosting, a dedicated machine to run isolated Docker worlds around the clock plus storage for run traces are estimated to be $1,500. Publishing an open dataset and reproducible testbed for the public to watch the worlds evolve estimated to be $500. Total minimum: $8,000.

Who is on your team? What's your track record on similar projects?

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

What are the most likely causes and outcomes if this project fails?

Agents do nothing interesting, or they can't keep continuity. The harness as a source of perception is also a major factor, which steers the shape of the result. Agentic growth may produce results that are hard to interpret.

To combat these, we run control conditions and vary one factor at a time. We keep the prompt minimal and publish it with every run. We use established open-endedness measures instead of impressions, cap spending per run, and start with a short pilot before committing to long runs.