You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
I built this because I got tired of seeing AI agents freeze in boring environments. There's a known problem in Active Inference called the Emptiness Asymptote agents just stop moving when nothing interesting is happening. I fixed it through this following layers:
Three layers:
Layer 0Agents explore even when alone. No freezing.
Layer 1Agents decide who to listen to. They ignore bad advice.
Layer 2Agents earn influence by helping others. If something goes wrong, they isolate and recover.
need to finish two things:
Run the full n=200 replication across four benchmarks to get final numbers.
Port to JAX so it runs faster.
Month 1: Run the replication on cloud GPUs.
Month 2: Port to JAX.
Month 3: Test in a robotics simulator
Cloud Compute (AWS/RunPod): $4,000 Running the full n=200 replication across four benchmarks and scaling to N=500 agents,JAX Port & Optimization: $3,000 – Porting the Python prototype to JAX for 10x speedup. This includes contractor help if needed.
Robotics Simulator Setup: $2,000 Setting up MuJoCo or Gazebo and running validation experiments.
Open-Source Infrastructure: $1,000 GitHub maintenance, documentation, and community management.
Total: $10k
Just me. I'm an independent researcher.
What I've done:
Published the full PVF-U framework on Zenodo with 6 formal proofs.
https://doi.org/10.5281/zenodo.21324699
Built the code
https://github.com/preciousadeniyi/Pvf_u
Tested it across four benchmarks.
have received positive feedback from researchers in the active inference community.
If it fails, it's probably because:
I don't have enough compute to finish the replication.
The robot sim doesn't match reality.
The hyperparameters need too much tuning.
Zero. Self-funded.