Build dangerous capability evals with OpenAI governance team - seeking engineering/research contractors

[Rosie Campbell]

Hi everyone!

The OpenAI Governance team is looking for engineers and researchers to build Dangerous Capability Evaluations (DC Evals): an eval suite for assessing dangerous capabilities that AI models could have, which increase the chance of catastrophe (e.g. capabilities like persuasion, deception, contributing to weapons proliferation, self-replication or -exfiltration, etc.). Our ideal is to contract with people who have availability through the end of the year and a meaningful amount of hours to work during this time period.

We are looking to hire contractors who can work on the full range of tasks required:

• Brainstorm eval ideas to assess dangerous capabilities

• Thoughtfully think through and construct the experimental design of individual evals

• Code up evals, acquire or create datasets, run results, and iterate rapidly

• Clearly communicate materials in written reports (eval results, how others can run the eval, surprising findings, etc.)

• Document evals for broader usage, including making available to other AI labs and, in some cases, fully open-source

See document for more information:
https://docs.google.com/document/d/1wUV_wzuciF45vLWcrXX4vlD0AFpl-6sznzsQHm8lb5w/edit

Please let me know if you have any questions!

Thanks,

Rosie