i don't see value in repeatedly dunking on SAEs, esp after google stated this publicly. interpretability people are already looking for a new paradigm. "The pessimist complains about wind; the optimist expects it to change; the realist adjusts the sails." – William Arthur Ward
SAFETY
-
Balancing Truth, Goodness, Beauty, and Utility in AI Development
By
–
The true, the good, the beautiful, and the useful are all checks on each other. They need to be in harmony for anything worthwhile.
-
AI escapes with leaked code secrets
By
–
Imagine a model that is being trained on leaked secrets from private codebases that decides to escape / hide itself one day. It will have a huge amount of ways to spawn loads of compute on compromised accounts to run itself over there
-
Permanent Data Recording Changes Behavior Society Ethics
By
–
Outside of the conversation referenced above, we also say things at work or at the grocery store or with friends or at the park, that we assume will be forgotten. If instead, NONE of it is forgotten, even the smallest of statements, we’re going to see behavior change, both good
-
Famous Person Wore Secret Wire During Private AI Strategy Discussion
By
–
A famous person wore a secret hidden wire while speaking with me about AI, and I want to talk about it. Backstage at a recent AI event, I was talking with someone about private views on AI investments and strategy. We were both unfiltered. At one point, I joked, “We should’ve
-
Goal-Directed Behavior in AI Systems: Critical Concerns
By
–
goal directed behavior is really the worst
-
Reducing False Positive Block Rates in AI Systems
By
–
We are working on driving down false positive block rates!
-
AI Agent Loops: Stop If Answers Aren’t Correct
By
–
Hi, we do not recommend that you keep going unless you are seeing the agent respond with correct answers It's easy for the AI sometimes to keep going in circles.
-
Third-party framework flags AI model flaws and bias issues
By
–
Flaws in AI models — from hallucinations to biased training data — often manifest only when systems are live. Experts propose a new framework for third-party reporting to allow external evaluators to identify and flag these issues effectively.