The use case is if you have a lot of people who all want to verify that the same known large expensive AI model said a thing. Another use case might be if there's some way to have a fixed model public signature without revealing its weights.
@esyudkowsky
-
Fermi Paradox: ASI Expansion Remains Likely Despite Existential Risk
By
–
We are almost certainly past the entire Fermi filter / all Hansonian hard steps, at this point; by far the most likely way for us to die (ASI) still leaves a giant expanding bubble chewing through the galaxy at a fraction of light speed.
-
Leaders Misrepresent Policy Positions on Nuclear Data Center Strikes
By
–
I see your leaders have been lying to you about which policies I have ever advocated. (There is no reason to ever drop a nuke on a data center.)
-
Chris Olah unable to utilize billion-dollar interpretability funding
By
–
I told open philanthropy a long time ago that if Chris Olah could think of a way to spend $1 billion on interpretability, they should give it to him, but apparently Chris could not think of a way to spend $1 billion.
-
Conscience-driven employees departing OpenAI
By
–
Surprising; everyone I personally knew to have a conscience has left OpenAI, but I guess there's some left anyways.
-
Building Complex Systems Without Understanding Underlying Algorithms
By
–
Oh, now that's just ridiculous. You found a way to build things indirectly without humans understanding the circuits that get found. Of course there's interesting algorithms inside there! You just don't know what they are.
-
AI Boyfriends and Gender Bias in Media Coverage
By
–
I expect AI boyfriends to be a similar actual problem to AI girlfriends, but I expect the MSM to cover it much less. Vibrators and Fleshlights get very different coverage.
-
Counterargument to LLM Niceness Implying Easy Alignment
By
–
I agree. I only counterargue the argument "LLMs nice therefore alignment easy".
-
Human-Competitive AI Boyfriends/Girlfriends Likely Soon
By
–
That said, if you want me to call what seems obvious, it presently seems quite likely that we'll get human-competitive artificial boyfriends/girlfriends, because this is a domain where LLMs just need to match style and can screw up 5% of the time.
-
LLM Alignment Challenges: Public Misunderstanding of Technical Complexity
By
–
This state of affairs itself puts the mockery to the idea that LLMs are a case exhibit of alignment being easy. But the average person exhibiting them as a triumph cannot grasp the distinction, so all you can show them is the outward misbehavior.