“Our model grader identified 10 rollouts, or approximately 1 to 2 rollouts per 1,000, Bayesian mean posterior rate of 0.17%, 95% interval of 0.09% to 0.28%”
LessWrong (Curated & Popular) · "GLM-5.3 Sometimes Considers How Train-Time Actions Shape Future Behavior" by Axel Højmark, Ezra Newman
Play from 21:00
21:00
Audio streams from the publisher’s own feed. Verbatim does not host or copy it.
What was described
used a model grader to scan 6,452 GLM-5.3 rollouts for verbalized update-aware reasoning and counted the occurrences
10 rollouts identified, ~0.17% rate (1-2 per 1,000), 95% CI 0.09-0.28%
Said nearby on the same subject
“I have eight different local models, open weight models running right now in my house, in my attic... they all have personas. My most recent is GLM-5.3 persona.”
“We have a bunch of fine-tuned models ourselves too… making us more innovative… couldn’t do with the frontier models because of cost and efficiency reasons”
♫
Said on All-In with Chamath, Jason, Sacks & Friedberg