Extreme AI Risk W05

Show Notes

In this week's newsletter, we explore the topic of modern large models’ alignment and examine criticisms of extreme AI risk arguments.

Join the Alignment Jam hackathon this weekend to get experience in doing ML safety research! https://ais.pub/scale

Opportunities

• Join Ought in a senior role: https://ais.pub/ought 
• Be the communications expert of the Fund for Alignment Research: https://ais.pub/farcomm 
• Refer a cool friend to the Redwood Research summer internship! https://ais.pub/referrr 
• Or apply for it yourself:  https://ais.pub/internrr  
  
  

Sources

• LLMs are implicitly topic models: https://arxiv.org/abs/2301.11916
• Condition predictive models: https://arxiv.org/abs/2302.00805
• Goal alignment: https://arxiv.org/abs/2302.00813
• Alignment with human representations supports robust few-shot learning: https://arxiv.org/abs/2301.11990
• MusicLM from Google very good: https://arxiv.org/abs/2301.11325
• Nuno Sempere’s criticism of high AI risk: https://nunosempere.com/blog/2023/01/23/my-highly-personal-skepticism-braindump-on-existential-risk/
• Neel’s quickstart guide: https://www.alignmentforum.org/posts/jLAvJt8wuSFySN975/mechanistic-interpretability-quickstart-guide
• Memorization → generalization double descent phenomena: https://www.alignmentforum.org/posts/KzwB4ovzrZ8DYWgpw/more-findings-on-memorization-and-double-descent
• Inner alignment in Simulators: https://www.alignmentforum.org/posts/FLMyTjuTiGytE6sP2/inner-misalignment-in-simulator-llms
• Action at a distance in the loss landscape: https://www.jessehoogland.com/note/spooky-action-at-a-distance-in-the-loss-landscape
• Luciano Floridi on ML Street Talk: https://youtu.be/YLNGvvgq3eg
  
  

His article about superintelligence: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4222347 

What is ML Safety Report?

A weekly podcast updating you with the latest research in AI and machine learning safety from people such as DeepMind, Anthropic, and MIRI.