"Correct Answer Features" Cannot Explain Multiple Choice Capabilities
TL;DR: I present theoretical and empirical evidence that LLMs cannot be (exclusively) using a "correct answer feature" as the main mechanism by which they perform multiple…
Latest
The Once And Future Fable #5
We, or at least ‘more than 100 American institutions,’ got Mythos back this week. What we the people do not have is Fable or Sol. While…
Cluelessness: Summary of the argument, why it matters, and counterarguments
(I wrote this post partly to help orient those interested in participating in the EA Forum’s Cluelessness Critiques Competition. The competition closes August 14th.) I’d like…
Is it ethical to work on general-purpose robots given the risk of cyberhacking?
One potential risk of developing general-purpose robots is that they could greatly reduce the friction required to establish a totalitarian regime. If there were millions+ of…
That Which Cannot Be Poked With A Stick Is The Mind-Killer
For part one of the aspirant sequence - which may or may not be arranged into some totally different order when I'm done with it, because…
The Name is Not The Model
AbstractA safety evaluation has to trust something to know what it is testing: the model's name, its version string, and the reasons the model gives for…
Connect to your past selves
(Introduction: Your four-dimensional body)Some people use meditation to learn valuable things about their deepest selves; some use therapy; some use psychedelic drugs. I use time travel.Why…
Unjournal tool hub (finding & assessing high-impact research questions, cruxes, papers, etc.)
Disclaimer: I used an LLM to help draft this post and it likely contains >10% AI-generated text, but I’ve edited/rewritten it extensively endorse it.The Unjournal has…
Preliminary investigation: KL penalties in RL can increase CoT unfaithfulness
Authors: Satvik Golechha, Sid Black, Joseph BloomWork done as part of the Model Transparency team at UK AISI. We consider this to be a small set…
Structural Proxies
Lately I've been thinking a lot about what work would help with actually winning and getting to good worlds. In the spirit of that I decided…