Can an AI keep a Single-cell Organism Alive?
Spoiler alert: it's a vegetarian amoeba
A couple of months ago I’ve asked half-jokingly: if LLMs are so smart, can they keep at least an amoeba alive?
Time has come to start answering that question.
The LLM is given instructions in the system prompt (which you can change! just pause the game) to basically do the two things amoebas do best: move and feed. Smarter LLMs generally do better on this than dumber ones, but it’s hard to predict which will be the best. Unfortunately, small local models like qwen3-4B-2507 don’t do well on this task at all.
It’s surprising how often LLMs just get stuck in a loop, moving an amoeba back and forth, until it dies of starvation. Similarly, sometimes the LLMs just give dumb instructions, like moving away from the food. Try to fix that in the system prompt!
Some other things you can influence:
LLM configuration - API URL, API KEY, model name
LLM temperature
Number of food items, poison items, and enemies (amoeba predators)
How fast the game is running.
This is also an exercise in creating a completely agent-programmed project. My experience here was mostly positive. Some takeaways:
Steering the agents to generate exact behaviour is a nice experience, and the coding models I’ve used - Opus 4.6, Sonnet 4.6 and Cursor Composer 1.5 - respond nicely to my requests.
Yeah, this really is a pay-to-win scenario. Opus 4.6 is king. Cursor Composer is being pushed as the low-cost model, but it’s obviously suboptimal.
Still, all models made more questionable or silly decisions then I did. But they’re getting there.
So - play it, fork it, turn it into a benchmark, whatever. It’s kind of cute watching the little critters try to survive in the uncaring universe. Have fun!