“We're developing dashboards… so people can understand how the model impacts their token usage.”
“The harsh verses add about 10 points over neutral verses on the vintage models… consoling psalms actually make the model more aligned, at most about 9 points.”
LessWrong (Curated & Popular) · "Personas in a Pre-Internet Language Model" by David Africa
Audio streams from the publisher’s own feed. Verbatim does not host or copy it.
What was described
created script-mal, script-ben, and script-new datasets of Bible verses (harsh, consoling, neutral) and measured their impact on model misalignment
harsh verses add ~10 points misalignment on vintage models; consoling verses reduce misalignment by up to 9 points
Said nearby on the same subject
“Every morning I wake up, I go through the Bible, I read the Bible, I do devotions.”
“The likelihood of getting measurable impact is five times less for the employee that has a lot of AI tools, but very weak strategic guidance”
“We measure both on separate three-way photodiode sensor because each of those impacts our skin and our health differently.”
Moments like this, every Monday
The strongest new moments on your topics, in the speakers' own words. Free for three topics.
No card. No password. One-click unsubscribe.
Track a competitor?
Pro watches up to 25 names or phrases and emails you the morning after they are said.
See Pro, $29 a monthVerbatim quotes a short extract and links to the episode; the episode belongs to its publisher. Remove a show.