Some key changes that stood out to me (got the opportunity to help w writing these!):

1) Stronger recs for single pill combos
2) Broader recs for screening for primary aldo (now consider in anyone with stage 2 HTN — i.e., BP >=140/90)
3) Just. Say. No… to cuffless BP devices… and to EtOH

2/

We need some of these phrases for nephrology!

Kidney failure = Filter futility

Peritoneal dialysis = Abdominal cleaning

Hyponatremia = Salt shortage

IMO the key finding is that the best LLMs (i.e. chatgpt) are able to perform with minimum bias even if those biases may be present in the training data.

Lowering rate of error and bias is critical to safe deployment and efficiency gains.

It's an EBM level advancement that we can't ignore.

This is a really neat article, but also akin to a massive meta-analysis that lumps together a lot of very different papers.

Study confirms general views that:
(1) instruction prompt matters
(2) Large model >> Small models