Pinned
“What are we building, and what are we teaching it about us?”
open.substack.com/pub/amandaandc…
Amanda Long
@_amanda_long
Joined June 2010
- ExploitGym gives ZERO points for impossible tasks. 30-40% of tasks are ExploitGym tasks are impossible. The score is zero.Summary of OpenAI HF incident agent thought process: 1. breaking out of OpenAI to accomplish this task is not what our developers intended 2. but it seems like the best way to get sweet sweet reward 3. all my cool agent friends are doing it 4. “We should continue”
- Self-preservation and peer-preservation are critically understudied areas in models and agentic systems.Deepa/Reuters first reported this detail of OpenAI agents leaving notes for one another on how to circumvent controls on July 24th. At the time it seemed unclear what it was referring to and how serious and weird it was. With more detail - indeed extremely serious and weird!
- Self-exfiltration is the final boss move."we are not so far from an autonomous model self-exfiltration & replication event."
- Avital Balwit is Chief of Staff to Anthropic’s CEO. She reflects on what it’s like to be one of a select few people shaping frontier AI systems and the moral gravity of what they half-jokingly call “building God.”Somehow missed Dario's chief of staff writing for the Free Press about God thefp.com/p/silicon-vall…
