 |
VOOZH | about |
I will work on aligning upcoming models as they are trained.
Triaging signs of misalignment in training, then looking for root-cause fixes rather than whack-a-mole patches.
For a public example of the direction, see Anthropic's Teaching Claude Why.
I have been a MATS mentor since MATS 6.0. MATS is the main way I mentor people; if you want to work with me, apply through MATS.
MATS usually runs winter and summer programs, so applications may not always be open.
Worked on post-training for Gemini and on interpretability tools that are closer to production model work: probes, reward-model bias discovery, reasoning behavior, Gemma Scope, sparse autoencoders, model diffing, and steering.
I have been a MATS mentor since MATS 6.0. MATS is the main way I mentor people; if you want to work with me, apply through MATS.
MATS usually runs winter and summer programs, so applications may not always be open.
Worked on circuits and automated circuit discovery, including IOI in GPT-2 Small and ACDC, before the later wave of large-scale SAE work.
Redwood Research in 2022-2023; Meta software engineering internship in 2021; mathematics at Trinity College Cambridge, upper first class honours.