Jason Wei

A 24-year-old man presented to the emergency department with rectal pain and lesions of 3 weeks’ duration that were progressively worsening. He had a medical history of poorly controlled HIV, cerebral toxoplasmosis, and genital herpes, as well as a social history of sexual activity with other men. He had been diagnosed with HIV 7 years prior and had been off therapy until 1 year prior to the current presentation, when he was hospitalized with encephalopathy (CD4 count, <50 cells/mm3).

A comprehensive metabolic panel collected from our patient 1 month earlier did not reveal any abnormalities. Serum methylmalonic acid and homocysteine were both elevated at 417 nmol/L (reference range [for those aged 2–59 years], 55–335 nmol/L) and 23 μmol/L (reference range, 5–15 μmol/L), respectively. Serum folate and 25-hydroxyvitamin D were low at 3.1 ng/mL (reference range, >4.8 ng/mL) and 5 ng/mL (reference range, 30–80 ng/mL), respectively. Vitamin B12 was within reference range.

By Jason Wei (OpenAI) Note: Mistakes & opinions are my own, and not of my employer. Why do large language models work so well? Manually inspecting data provides clear intuitions about model usage. Observing data is akin to training a biological neural network. Intuitions gained can be valuable (e.g., annotating a lung cancer image classification dataset led to several papers). Next-word prediction is massively multi-task learning. Learning relationships can be cast as next-word prediction.