# hi!! I'm Kamilė Lukošiūtė^[ ![[L1003459.jpg | 200]]]. I'm an AI safety/security/evaluations researcher. I care about figuring out how to meaningfully measure meaningful things about LLMs. That means figuring out what we want to measure and how to do it well. In the past, I've worked on alignment (at Anthropic) as well as uses of LLMs in corporate security applications (while at Cisco Security). I'll be starting as a Research Scholar at [GovAI](https://www.governance.ai/) in August. I like drinking bright, acidic espressos in the mornings, flaky salt on top of my desserts, and eating lunch at dog parks. Contact me at `lukosiutekamile [at] gmail [dot] com` . I like getting emails! ### blog posts about evals - [[Building evaluations for cybersecurity assistance]]: A short post work-in-progress post. - [[You need to be spending more money on evals]]: On over-generalisation in LLM evaluations - [[Maybe don't give language models standardized tests]]: On misleading LLM evaluations ### blog posts about other topics - [[American Character]]: On the best things about America - [[Neutron star mergers and fast surrogate modeling]]: On my astrophysics research - [[What does BCEWithLogits actually do?]]: On technical details - [[When can a tensor be view()ed?]]: On more technical details ### past work - I was a resident researcher at Anthropic, where I worked on scalable oversight and evaluations. One of my publicly available projects was on [model-written evals](https://arxiv.org/abs/2212.09251). - I was a PhD in machine learning for theoretical physics but dropped out to help to make AI's safe. - During my bachelor's and master's in physics, I published a paper on [training cVAE surrogate models for kilonova parameter inference](https://arxiv.org/abs/2204.00285). ### contact! [[CV.pdf |CV/Resume]] [twitter](https://twitter.com/kamilelukosiute)