Migrating recent posts from x. Super excited to share our paper in NPJ Digital Medicine looking at how language models perform when asked to make *clinical calculations*. They are not good (35% wrong), but if the model is given a helpful tool, the error rate decreased by 13-fold! alx.gd/calc

As someone who doesn't check my twitter account very often, it's nice to see how the AI notifications are spot on with my interests. It's time for me to actually start moving to bluesky. I'll be moving some of my recent posts over to here.