Beam Notes
Dispatches from a radiation oncologist's AI notebook. New resources, practical insights, and the occasional unsolicited opinion.
No spam. Unsubscribe anytime.
Past Issues
Beam Notes: Yesterday's Model, Tomorrow's Patient
A landmark Harvard study just showed an LLM outperforming ER attendings on real cases. The model they tested is o1 — frontier in late 2024, two model generations old by the time it appeared in print. Meanwhile design got cheap, the on-prem stack got serious, and the evidence layer is the slowest layer in the whole pipeline.
Beam Notes: The Floor Is Not the Ceiling
JAMA called the gains from AI scribes 'modest.' That number is a floor, not a ceiling. Meanwhile Anthropic just locked its most powerful model behind a closed cybersecurity consortium, every clinician is becoming a developer of one, and the playbook for an AI-orchestrated hospital breach is now public.
Beam Notes: The Skill File
Utah is letting AI prescribe drugs and regulating mental health chatbots in the same legislative session. Prompts are out. Context engineering is in. The skill file is eating the prompt.
Beam Notes: The Tuesday Shift
I have been paying attention to the wrong thing. We all have.
Beam Notes: The Middle Ground
Open LinkedIn on any given morning and you will find two parallel universes. The truth, as it tends to, lives somewhere in the middle.
Beam Notes: Welcome to BeamPath AI
BeamPath AI has a new home, a new look, and — thanks to a rather opinionated AI companion — a newsletter. Here is how we got here and where we are going.