Grim if true (for reasons basically unrelated to the totally separate track where later ASI later kills everyone later)
@esyudkowsky
-
Stochastic Parrots Writing Python Functions from Finetuning
By
–
Stochastic parrots: famously self-aware enough to write down in Python the function they've learned from finetuning on a corpus of input-outputs, without any further training in doing that.
-
Methods for Preventing AI or Fighting Superintelligence
By
–
As a method for preventing AI, or a method for fighting superintelligence?
-
AI behaviors learned from examples, not explicit code
By
–
Those aren't coded. They're shown by examples, and some mysterious black box that mostly matches the examples becomes the AI.
-
Verifying Answers from Supposedly Brilliant Alignment Strategies
By
–
I mean, how do you verify the answers of something that is supposedly giving you brilliant alignment strategies?
-
Alignment’s Hard Problem: Two Decades of Misrepresentation
By
–
Alternate take: Alignment started with the obvious observation that alignment might be hard, but the implications of this idea were outside the Overton Window, so a variety of different Pretend Serious groups spent two decades seizing on whoever could be misread for an excuse.
-
Superintelligence in Two Decades: Scaling Limitations and Realistic Timelines
By
–
Two decades more before superintelligence is surprising to me but not forbidden by my understanding of reality. Scaling hits a wall.
-
Yann Ignoring Basic Alignment Questions and Assurances
By
–
Here's Yann ignoring the first most basic questions about his alignment plans and assurances.
-
Incomprehensibility of SGD-Generated Programs
By
–
You can't even understand all the program that stochastic gradient descent creates.