Releasing Work In Progress

During my time at QURI, I wrote a lot of drafts and internal docs. As part of winding down my involvement, it made sense to get them public rather than let them sit in private folders.

You can find the QURI drafts here.

Related, I've spent recent time working with Claude to build out a few research wikis on AI safety and evaluation. These are rough and exploratory, working notes rather than settled positions, but some readers might find them interesting.

Delegation Risk treats the risk of delegating a task, whether to a person or an AI agent, as a quantity you can estimate. It then suggests strategies for decomposing tasks in order to minimize this risk. This is the longest of these wikis.

Robust Reasoning Processes discusses the study of reasoning procedures (peer review, audits, prediction markets, LLM judge pipelines, debate protocols), focusing on the cost to extract a validated conclusion and the cost for an adversary to corrupt the output. The goal is to build towards a science of reasoning processes that could withstand scheming, such as from an AI agent.

Evaluation Engineering treats the production of estimates and evaluations as a systems-engineering problem. The motivating picture is a world where we use AIs or humans to systematically evaluate large sets of important parameters about the world.