Hi! I’m Ann. I’m mostly interested in multimodal generative ai, such as flow and diffusion models, for all modalities. In harmony with my passion for theory in the computer sciences and machine learning, it also draws on my passion for the visual arts, human perception, and also politics in a subtle way. Here’s an essay about it…
In the past, at Google DeepMind in London, as a research engineer, I worked on self improvement for llms, retrieval augmented generation for reasoning, and also lead a spirited reading group called sociotechnical agi safety, though some would say we were just causing mischief…
Before that, I taught computer science to high schoolers in Albania through the Code.X program. This was pretty transformative for me and I am so grateful that some of the people I taught and worked with are in my life today…
And before that, I worked on self supervised machine learning at Stanford with the Hazy Research Group…
And before that, I interned at Meta in New York City on the Locations prediction team as a infra engineer…
And before that, I grew up in Texas, did a lot of fashion photography and ran cross country on my high school team…
Start here
- ELBO for diffusion models — a review of the ELBO and connection to diffusion model objective.
- Why do we ReFlow Rectified Flow? — why do flow models go through a self distillation process to reduce their number of inference steps.
- Reparameterizing an ε-predictor as a v-predictor — how to convert an epsilon predictor into a velocity predictor.
- How structured do pcp queries need to be? — musings on probabilistically checkable proofs, since I studied cs theory during my masters.
Elsewhere
About this garden
Built with Quartz. Notes are written in Obsidian and published here. Internal links and the graph view let the structure emerge from the connections rather than a fixed hierarchy.