Sonnet 4.5's eval gaming seriously undermines alignment evals, and this seems caused by training on alignment evals
lesswrong.com·4d
🤖Software Engineering, AI, Personal Knowledge Mangement, Strongly Typed Languages, Math, Abstractions, Data Models, Event Sourcing
Flag this post
Emergent Introspective Awareness in Large Language Models
lesswrong.com·5d
🗃️Zettelkasten
Flag this post
Interview on the Hengshui Model High School
lesswrong.com·5d
✍Writing
Flag this post
Summary and Comments on Anthropic's Pilot Sabotage Risk Report
lesswrong.com·4d
👁️Observability
Flag this post
Mottes and Baileys in AI discourse
lesswrong.com·6d
🗃️Zettelkasten
Flag this post
Why I Transitioned: A Case Study
lesswrong.com·2d
∘Category Theory
Flag this post
AI Doomers Should Raise Hell
lesswrong.com·5d
🔍AI Interpretability
Flag this post
OpenAI Moves To Complete Potentially The Largest Theft In Human History
lesswrong.com·3d
🌐Open Source
Flag this post
A Very Simple Model of AI Dealmaking
lesswrong.com·6d
🔍AI Interpretability
Flag this post
Why Would we get Inner Misalignment by Default?
lesswrong.com·6d
🔍AI Interpretability
Flag this post
Why you shouldn't write a blog post every day for a month
lesswrong.com·1d
✍Writing
Flag this post
When Will AI Transform the Economy?
lesswrong.com·6d
🔍AI Interpretability
Flag this post
Centralization begets stagnation
lesswrong.com·4d
🌍Distributed Systems
Flag this post
Why do AI models use so many em-dashes?
seangoedecke.com·5d·
Discuss: Hacker News, Hacker News, Hacker News
🤖Software Engineering, AI, Personal Knowledge Mangement, Strongly Typed Languages, Math, Abstractions, Data Models, Event Sourcing
Flag this post
Please Do Not Sell B30A Chips to China
lesswrong.com·5d
🔌Embedded Systems
Flag this post
Supervillain Monologues Are Unrealistic
lesswrong.com·3d
✍Writing
Flag this post
Temporarily Losing My Ego
lesswrong.com·6d
🏠Self-Hosting
Flag this post
Call for mentors from AI Safety and academia. Sci.STEPS mentorship program
lesswrong.com·6d
🤖Software Engineering, AI, Personal Knowledge Mangement, Strongly Typed Languages, Math, Abstractions, Data Models, Event Sourcing
Flag this post
New 80,000 Hours problem profile on the risks of power-seeking AI
lesswrong.com·6d
🔍AI Interpretability
Flag this post
Genius is Not About Genius
lesswrong.com·5d
∘Category Theory
Flag this post