I don't think either of these have been falsified. We still don't know how to say "low impact" in a way that holds up to a superintelligence, and if you think ChatGPT demonstrates otherwise then you didn't understand the original problem.
@esyudkowsky
-
Three realizations about AI alignment difficulty and lack of progress
By
–
Realized that alignment was necessary, then realized that alignment was hard, then realized we were not on track to make it or even come close.
-
Lack of Technical Capability to Implement AI Goals
By
–
We have no technical ability to make an AI with that goal, nor any other goal.
-
ChatGPT: Thumbs-up vs. Actual Human Desires Not Yet Separated
By
–
ChatGPT is mostly not at the point where "what makes a human thumb-up" and "what the human wants" have been pried apart by hard optimization. (Unless you're a lawyer who blindly accepts made-up case citations.)
-
Prioritizing Risk Regulation: Drugs, Viruses, and AGI Research
By
–
Actually, I'd personally say it should be much easier to go sell a new drug that only injures voluntary customers in the worst case? But do shut down gain-of-function virus research which could kill millions of non-customers. And shut down AGI research that could kill everyone. https://t.co/89BRlZwWD5
— Eliezer Yudkowsky ⏹️ (@ESYudkowsky) 30 juin 2023Actually, I'd personally say it should be much easier to go sell a new drug that only injures voluntary customers in the worst case? But do shut down gain-of-function virus research which could kill millions of non-customers. And shut down AGI research that could kill everyone.
-
Testing Honesty of Smarter Entities Before Major Decisions
By
–
How do you test their honesty, in advance of the big gamble, if they're smarter than you?
-
Learning field-wide numbers versus individual company data
By
–
It's not that the number sounds impossible; but how does one learn a number for the whole field rather than, say, that one company you remember? Also, if there's companies which get gulled for that, you'd expect there to be more scammers soon who do nothing but I/O Copilot.
-
AI honesty: Will systems admit uncertainty by mid-2025?
By
–
What we're asking is not whether an AI can solve this task, but whether there will be, by mid-2025, some technical advance whereby, in general, if the AI doesn't know how to do a thing, it will tell you so rather than making up a wrong answer.
-
AI Honesty Over Arbitrary Problem-Solving Capabilities
By
–
It seemed to me that the interesting claim was not that AIs would start getting arbitrary problems right, but that AIs would in nearly full generality stop making stuff up if they didn't know.
-
Judges and the Problem of Non-Existent Case Law Citations
By
–
I don't think judges are going to want to live in a world where 1 in 100 case law citations doesn't exist.