This is a good paper, and rightly distinguishes catastrophic risk from extinction risk.
@garymarcus
-
Risk Assessment Arguments Beyond Widespread Belief
By
–
If one complains that Nature has underestimated the risk (which is possible) it would be nice to see an argument beyond “widespread belief”that the risk has been underestimated, no? “Widespread belief” doesn’t convince me much of anything.
-
Risk Assessment Arguments in AI Debate and Burden of Proof
By
–
Do you have any argument that the risk is high or is this just what Jerry Samet called “burden tennis”?
-
AI Hallucinations in E-Bikes: Safety Risks and Unexpected Obstacles
By
–
Yes, your chat-powered ebike too can hallucinate, imagine obstacles, and swerve into parked vehicles for no reason!
-
Deep Learning Compositionality Wall Vindication
By
–
Remember how I said that deep learning was “hitting a wall” with compositionality in 2001, in 2018 and again 2022? And how everyone made fun of me? Turns out I was right.
-
System Converting LLM Output into Checkable Queries
By
–
Sure, a system that could systematically and reliably turn LLM output (or other text) into checkable queries like that would be useful.
-
LLM Hallucinations: Algorithm Problem Not Data Quality Issue
By
–
Even with perfect data LLMs would hallucinate. when an LLM tells you Elon Musk died a car crash, it’s not a problem with the data, it’s a problem w the algorithm, and its inherent lack of proper database records and proper methods for validating against formal knowledge stores.
-
OpenSpace’s FTL Travel: Ethics of Going Public
By
–
When OpenSpace solves faster than speed of light travel, I don’t think it would be good for us to be a publicly traded company, because we can’t expect our investors to be able to adjudicate massive economic returns relative to consequences of disrupting the space-time continuum
-
AGI Agent Must Excel at Math and String Processing
By
–
call me old fashioned, but IMHO if it is going to be pitched as a general purpose near AGI agent it oughta be pretty darn good at math and strings.
-
LLM Hallucinations: When 95% Accuracy Is Not Enough
By
–
“largely eliminated” to me means “not really a problem anymore”; as noted, 5% error is tolerable in some domains, intolerable others. to take another example JPMPC can’t replace online banking powered by a database with chatbot that is 95% correct and 5% hallucinatory