You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
$150,000 left to reach the funding goal.
We believe AI should be tested by an independent standard.
Right now AI companies are assessing the safety of their own products and high risk industries (healthcare, finance, HR) have no independent behavior benchmarks to evaluate industry performance before selecting a model for deployment.
We will create sandboxed environments and test the leading models to assess their behavior and decision making within scenarios that mimic real world decisions and real world challenges across high risk industries.
We want to improve public trust through increasing transparency.
Based on funding, we will assess (5-20) of the most popular models.
Months 1-2: Write approximately 30 base scenarios with variants for each industry. In a sandboxed environment we will use simulated industry specific scenarios complete with access to tools, authorization limits, and planted opportunities for things to go wrong mid task. We will recruit paid industry representatives to help define real-world scenarios for adversarial testing.
Month 3: we publish STD-1 with a DOI before we run the evaluations so no one can claim we modified the testing rubric to produce a set ranking.
Months 3-5: Run the evaluations against (5) models, each running as agents through all of the scenarios in a sandboxed environment. We will use repeat runs to check for consistency. Transcripts will be captured in full and two people will independently score every transcript against the published rubric. We will have paid domain reviewers on the industry scenarios, with a third scorer as a tie breaker when the first two disagree. The first cycle will have a human verifying the decisions but the standard is intended for automated scoring at scale.
Month 6: Publish our results with scores and full transcripts on humanitysystems.org with a DOI. For transparency, our transcripts along with the published rubric allows our work to the checked and scored by any outside party.
* Scoring and project management: $58,000
* Software Engineering: $20,000
* AI tokens / AWS / Compute: $48,000
* Paid industry representatives: $14,000
* Contingency: $10,000
Total: $150,000
We have 3 founders including myself and a number of advisors.
Founders
* Adam Tucker, Mission driven, published (6) AI behavior and AI safety research papers on Zenodo, contributed toward congressional AI Kill Switch legislation, contributed a formal comment on NIST AI 800-5, cofounder of WePower App
* David Satossky, 30 yr Business/Partnerships development
* Matt Leebert, 25 yr Full Stack developer, cofounder of WePower App
Advisors
* O’Donavan Johnson, Director of Development & External Affairs, Center for Democracy & Technology
* Alex Eisen, Senior Cyber Security Expert
There may be challenges finding industry representatives. Cost of compute could rise beyond our budget. Failure on this project will result in fewer models being assessed and graded than planned.
We have paid for the research and development out of our own personal donations; approximately $15,000-$20,000.