2+ hour, 25+ min ago  (1246+ words) When playing this red-team'blue-team game, you need to select some set of affordances to give to both teams; it's pretty confusing to know what affordance set to pick. Here is one common approach to choosing blue team affordances: This approach…...

5+ day, 9+ hour ago  (1433+ words) We've just released BashArena, a new high-stakes control setting we think is a major improvement over the settings we've used in the past. In this post we'll discuss the strengths and weaknesses of BashArena, and what we've learned about how…...

6+ day, 2+ hour ago  (746+ words) Epistemic Status: Entirely ripped from Chapter 3 of "Causal Inference: The Mixtape" by Scott Cunningham which is graciously provided for free here and is generally a very good textbook. I hope that by writing this I provide a slightly shorter and…...

6+ day, 3+ hour ago  (382+ words) Very Slightly More In-Depth feedback: I have been thinking "someone(s) should make a 'Universal Paperclips', but a bit more realistic or framed around today's situation." I think I would be initially somewhat confused about what's going on. (Universal Paperclips was…...

1+ week, 3+ hour ago  (390+ words) Fellows will spend the first week picking out scoped projects from a list provided by our team or designing independent research projects (related to our overall'agenda), and then spend seven weeks working on that project under the guidance of our…...

1+ week, 19+ hour ago  (209+ words) Given the aversion to touchy-feely AI discussions that I expressed then, the fluffy, emotion-laden musings of the present blog post will perhaps come as a surprise. But here we go. I am honestly unsure about how to handle these conversations,…...

1+ week, 1+ day ago  (17+ words) Published on December 15, 2025 11:17 PM GMTDiscuss Introducing Lunette: auditing agents for evals and environments...

1+ week, 3+ day ago  (700+ words) Published on December 14, 2025 2:10 AM GMTTL;DR We release a replication of the model organism from Auditing language models for hidden objectives'a model that exploits reward model biases while concealing this objective. We hope it serves as a testbed for evaluating…...

1+ week, 5+ day ago  (626+ words) This is the abstract and introduction of our new paper." Links: " Paper, " Twitter thread, " Project page, " "Code LLMs are useful because they generalize so well. But can you have too much of a good thing? "We show that a small…...

1+ week, 5+ day ago  (222+ words) Epistemic status: I'm mostly confident about the evidence, having read the literature for the last months. The arguments below are my best guess." A survey on encoded reasoning and messages See the taxonomy for that survey"here. Preventing Language Models…...