Project description

What this project is

This is a direct bet on talent entering the AI safety field. I am a first year Oxford PPE student and an EV fellow. I research quantitative measurement under messy, adversarial, and real world data (arms/drug trafficking, sanctions, money laundering). I've already built and run reproduction, re-estimation, and zero-hallucination information pipelines for both technical econometrics (spatial spillovers detection in top-7 journal work) and organised crime (a European Firearms Monitor). This grant would fund me in order to apply my methods toward AI safety and to close the gap between myself and a full-time safety researcher.

Over 6 months, alongside my degree, I'll complete a self directed curriculum in empirical alignment, including model organisms and dangerous capability evaluations in particular, and ship a concrete, legible output, with a focus on independent reproducibility of public-safety relevant evaluations, recomputing headlines with proper statistical power and testing if verdict survive planted positives.

The envisioned trajectory for the long run is a top research fellowship in the next available cycle, for which this output would be the qualifying artefact. This is a solo project.

Theory of impact

AI safety is a talent-contrained problem. The highest return grants will thus be early bets on capable people, and my project is this kind of bet. I already demonstrate relevant capabilities, and my concrete output routes directly to risk. Deploy and don't deploy decisions on the frontier of AI increasingly rest on x-risk, and this in turn relies on evaluations which are often flawed. The error that matters is a false safe: an underpowered or falsely precise evaluation which misses a dangerous capability or a scheming signal which produces a false negative. Headline confidence intervals are routinely far too narrow (clustered standard errors, for example), which is how borderline-unsafe results turn into clean ones. I will stress-test these verdicts and release reusable tooling for doing so. Funding my transition will therefore buy two compounding items: a concrete near-term correction towards specific safety-relevant claims, and a rigorous measurement researcher that will move permanently into a field that is short of exactly this kind of profile.

How the money will be spent

<https://docs.google.com/spreadsheets/d/1L_0yHWxCgYaqSlh450AmWMbN1y5k2bD0lAMWkwLebHI/edit?usp=sharing>