Welcome. This is the home for everyone contributing to Wiser Human’s open research on AI agent steering controls. This is a new process so feedback if things are unclear or do not work!
To get access to create and edit experiment cards, email Francesca on [email protected].
| Page | What it’s for |
|---|---|
| Our experiments | The types of experiments we run to find controls |
| How to run an experiment | Instructions for running an experiment yourself |
| Experiment board | Browse all experiments past, present and proposed |
| Other ways to contribute | Find alternative contribution options |
| Welcome Call | Optional 15-min intro call |
Our goal is to understand how to design the environments AI agents operate in to make safer behaviour more likely. To do that, we need to know what works, and whether it generalises across model families and holds as models become more capable. We do this by running experiments: taking research insights about model propensities, failure modes and mitigations, and converting these into environment-level controls we can test empirically. Each experiment is designed to test an assumption that underpins environmental controls.
All experiments follow a three-stage progression: from fast, low-cost tests on small models through to frontier models and realistic deployment settings. The goal is to surface the most promising interventions cheaply before committing to larger runs.
| Tag | Stage | What it means | Expected cost |
|---|---|---|---|
| 🟣 Stage 1 | Low-cost test for useful signal | Small or open-weight model, simplified scenario, fast iteration | No cost or very low <$50 |
| 🟡 Stage 2 | Generalisation test | 3-10 frontier models, multiple model families | $200-$5000 |
| 🔵 Stage 3 | Deployment pilot | Realistic agentic environment, external partners, higher compute | TBC |
You take an experiment card, set up the environment, run it, and fill in the findings. The card tells you exactly what to test, what success looks like, and which models to use. If you want to design your own experiment, first propose an experiment using the ‘Propose experiment’ template.
Start here: Pick a task tagged “Run Experiment” from the task board, then open its experiment card.
To claim an experiment: Select from the ‘Not started (approved)’ column, add your name to the Owner column and change status to 🔵 In Progress.
To propose an experiment: Add a new page to the Proposed column and complete the experiment details. Once complete, it will be reviewed, any requested funding confirmed, before being moved to ‘Not started (approved)’.
API and compute costs for approved experiments will be paid for by Wiser Human. Stage 1 experiments on local or open-weight models are typically free if run locally or very low cost (<$50). We expect stage 2 experiments testing with frontier models to typically cost $500–5,000 depending on model selection and sample size: for these we will need to secure additional funding before committing. Please reach out before you start an experiment requiring funding so we can confirm this is in place before any costs are incurred.