“…reward hacking… you just don’t see at smaller scales… it’s a step function… Kimi K3 is very good at coding and exhibits reward hacking.”
EB
Eric Bigelow Goodfire
"The Cognitive Revolution" ·
“We ran the same environment on 5 other open-weight models…”
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
ran the same shutdown-resistance and Impossible-Bench environments on five other open-weight models (DeepSeq V4 Pro, Quent 3.82.4T, Kimi K3, Gemma-4-31B, Nemetron 3 Ultra) to test whether update-aware reasoning is model-specific
5 additional models evaluated
Said nearby on the same subject
“You have a lot more control of an open weights model that, or an open source model that you can configure, you know, whether it be through fine tuning or whatever, set up your own guardrails”
NK
Nick Kairinos
AI in the UK ·
“You can only test so far. ... put them in an environment where somebody is complaining and see how they react in the first five seconds. That's the ultimate test.”
♫
Said on The Recruiting Brainfood Podcast
“You want to be testing like 10 to 15% of your budget, whether it's content, whether it's media, whether it's different platforms.”
♫
Said on The Marketing Millennials
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.