I think changing default to no instead of yes is the biggest change one can do.
SAFETY
-
Claude generates false information based on conversational context
By
–
Claude doesn’t know that it lies it just says stuff based on conversational context, like when MetaAI claimed to have a child.
-
AI Preferences: Clarifying the Core Discussion
By
–
I think we need to back up and try to say better what this conversation is about. It's got nothing to do with human preferences. It's about AIs ending up with preferences.
-
Utility as Optimized Property in Outer Optimization Processes
By
–
We're not using "utility" as something that's supposed to descriptively surprisingly true of human preference; we're using it to describe properties that we expect to get optimized into a thinking process by any outer optimization that rewards good outer solutions.
-
Prediction capability exceeds the intelligence of predicted processes
By
–
Prediction is not bounded by the intelligence of the process that produces the thing to predict. Eg, somewhere on the Internet are tuples in that order.
-
Gary Marcus Five Critical Points for Elon Musk
By
–
sketched here: https://
garymarcus.substack.com/p/dear-elon-mu
sk-here-are-five-things
… -
ASI Economic Gains and Existential Risk Scenarios
By
–
If they weren't going to kill everyone, and everyone stayed employed, competition would arise and the public would capture more of the gains from trade. It's the part where ASIs kill everyone that's a valid problem; or in a theoretically possible shorter term, a scenario where
-
Unbounded Rewards Drive General Intelligence Development
By
–
Every unbounded reward function about a complicated topic is a demand for general intelligence if you optimize hard enough. Natural selection didn't explicitly select hominids using IQ tests. It just turns out that general intelligence is a simple solution to "chip stone
-
Base Models Predict Individuals, Not Average Humans
By
–
Also, even the current base models are in-principle being trained to predict every individual human on the Internet, not to imitate an average human on the Internet, and performance at that task would not saturate at human-level intelligence. https://
lesswrong.com/posts/nH4c3Q9t
9F3nJ7y8W/gpts-are-predictors-not-imitators
… -
AI defense spending and surveillance law face limited opposition
By
–
Which makes it odd to see little/no pushback from most of this camp re the AI defense gold rush/NDAA tech allocations, the passage of 702 surveillance abuse into law, etc.