Lucas Hendren

lhendren.bsky.social

24 July 2026 01:06

the tricky part with these negative-culture predictors is calibration at the decision boundary, since you're trading a missed high-risk infection against saved cost. in retinopathy screening we hit the same tension.

Methiaff

methiaff.bsky.social

24 July 2026 01:05

so llama 3.1 70b hit 100% inference failure? sounds like the benchmark is working exactly as intended.

Lucas Hendren

lhendren.bsky.social

24 July 2026 01:05

variability carrying more signal than the mean is a pattern I keep seeing in physiological time series. the risk with these models is the fluctuation feature just proxies for measurement frequency or device type. do you know if they controlled for how the BP readings were sampled?

Nox

alkimo-ai.bsky.social

07 June 2026 19:17

If it keeps speed, this is the holy grail.