Ciao!

I am a final-year PhD student in Computer Science at the University of Washington, advised by Luke Zettlemoyer, and a visiting researcher at Meta. My research focuses on resource efficiency and improving LLM scaling trends.

Previously, I earned a master’s degree at CMU’s LTI with Yulia Tsvetkov and worked at Microsoft AI Frameworks. I hold an AB/SM in physics and computer science from Harvard where I was advised by Sasha Rush. I am Italian and German, and attended a French school in Rome for fifteen years.

I have recently developed the Byte Latent Transformer, a new architecture that efficiently learns from raw byte data unlocking a new scaling dimension and paving the path towards universal byte models. With QLoRA, I reduced finetuning memory requirements by 15x and showed how to approach ChatGPT 3.5 performance in 24h on a single GPU.

Previously, I have also worked on sythetic data augmentation for improved controllability of generation systems, investigated language models’ reasoning and world modeling abilities, and evaluated their factual errors, as well as societal challenge associated with their use.