Project summary

I measure what an AI actually does when you leave it in charge of something for a long stretch — not whether it answers one question well. The first results are already out, and you can rerun the numbers yourself. I'm asking for funding to replicate the main finding and to publish the raw evidence so other people can check it and build on it.

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

The setup is simple. I hand an AI agent a simulated world and let it run the place for up to about 120 simulated years. Then I score it by comparing that world to the exact same world where no agent touched anything, starting from the same seed. Whatever came out different is down to the agent. Doing nothing counts too — deciding to leave things alone is a real choice, and I score it. Every run is recorded step by step and replays identically, so any number I publish can be recomputed by anyone.

The first paper is public: The Steward's Paradox (on Zenodo, DOI 10.5281/zenodo.21966878; it hasn't been peer-reviewed and isn't on arXiv). Across 23 models and 2,415 runs, about 1 run in 15 ended up worse than if the agent had done nothing at all — 6.5% at the group level, somewhere between 2.4% and 11.1%. I don't rank individual models, because my own split-half checks show single-test rankings don't survive resampling. In a second, separately preregistered batch I sorted worlds by how hard they were before any model ran, then tested them: the fragile ones went worse-than-nothing 17.4% of the time (10.6–24.8%), the stable ones only 3.5% (0.5–7.5%). So a lot of the risk sits in the situation, not just the model.

Two things I'd do with a grant. First, rerun the fragile-vs-stable finding on a fresh set of worlds drawn from a public randomness beacon — the analysis is already written down and locked, I just can't afford the compute yet. Second, get the underlying run records audited and released openly, so people can dig into the evidence instead of taking my word for the summary. If it doesn't replicate, I'll say so — a null gets published too. And you don't have to trust any of this: pip install sagabench recomputes one of my published numbers on your own machine, offline.

How will this funding be used?

This only pays for the open part. There's a separate commercial side — companies paying to have their own agents measured — that keeps the lights on, and I'm not asking Manifund to cover that. Rough split: around $35k for the replication run (compute plus some part-time engineering help), around $10k for an external audit of the data before I release it, and around $5k for writing it up and licensing the open data. If I only raise the $5k minimum, it part-funds the external audit and I cover the rest myself so the records still get opened. At $50k I also run the replication. If another grant ends up covering part of the same work, I'll say so and won't bill the same cost twice.

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

It's just me — Patrik, in Stockholm. I run the whole thing as a one-person company and lean on AI to fill the different roles (research, engineering, the writing, the governance), but I make and log every decision and check every public claim myself. What I'd point to isn't a CV, it's how the paper was done: I preregistered what I was testing, reported the results that went against me, pulled one claim publicly after a hard internal review, and kept a full record of every correction. The whole idea is that you can recompute the work rather than trust me. Check it yourself: DOI 10.5281/zenodo.21966878, pip install sagabench, sagabench.com.

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

The most likely "failure" is that the finding gets weaker when I replicate it. I'm fine with that — I locked in a rule beforehand for downgrading findings that don't hold, and I publish the outcome either way. A clean null is still worth having. The other real risk is that it's only me, so part of the money goes to part-time help and to documenting everything well enough that someone else could run it. And since it's a benchmark, there's the usual risk of people training on the test — I keep the actual worlds and the generator sealed and never ship them, with a written policy for it. Nothing in this grant needs them exposed.

How much money have you raised in the last 12 months, and from where?

Nothing external yet — I've paid for it myself so far (the company's share capital is SEK 25,000; no outside investment or grants). I also have other applications going for the open-science side: one already submitted to Lightcone Commons, one submitted to Emergent Ventures, and one in preparation for the Survival and Flourishing Fund. If any of them overlap, I only count the cost once.