They are of course vastly different. As Marvin Minsky observed, nuclear weapons are not really dangerous because nuclear weapons are not self-replicating. To this I would add that artificial pathogens are not really dangerous because they're not smart.
@esyudkowsky
-
@esyudkowsky — 2023-09-22
By
–
What do you think are the ultimate limits of what, in your personal faith, technology is ever allowed to do? Can there be solar-powered self-replicating factories that are each a couple of microns across?
-
Superintelligence Deception vs Robot Training: Adversarial Problem
By
–
Fooling a superintelligence is a more adversarial problem than training a robot. In the latter case intelligence works with you; in the former, against you. "Can learn from" and "cannot distinguish" are different orders of requirement.
-
Skepticism About Superhuman AGI Alignment Feasibility
By
–
Where the hell are you getting the idea that anyone anywhere will be able to align a superhuman AGI and use it for things? What the heck is wrong with people's brains that they skip over this step in their story? Does your brain decide that this step is boring so it's okay to
-
Image Models Limitations: Computational Constraints and Reduced Capability
By
–
Images are hundreds of thousands of pixels, so nobody can afford an architecture that runs a large model over every pixel. The resulting small models are in fact much stupider than GPT-4 and have trouble following even slightly complicated instructions.
-
Open Source AI: Inevitable Capabilities and Alignment Risks
By
–
Open source is all you need to make inevitable any feat that one group can do and no one else can stop afterwards. This has both good and bad effects on society. The extinction issue, and more generally the present impossibility of aligning good ASIs to counter bad ones, makes
-
AI Capability: Paying Internet Users to Execute Tasks
By
–
Just to be clear, people do get "The AI can pay Internet randos to do whatever."
-
Value Misalignment vs Pseudo-Epistemic Problems in AI
By
–
This is not what value misalignment looks like. This is a pseudo-epistemic problem and it would be far more straightforward to solve than value alignment, which is about the equivalent of the utility function. Better architecture or maybe just more compute solves this. Any
-
Cognitive Design vs Natural Selection in Nanosystems Problem-Solving
By
–
– Read Drexler's _Nanosystems_.
– Why would you expect any of those problems to be harder to solve for a cognitively designed system rather than a mutated-selected one? -
Continuous prediction learning from webcam visual environment data
By
–
If you walk around with a webcam you can be constantly predicting what you'll see next, and checking if those predictions match reality. So the environment is generating data to learn from, but not humans as such. Open environment, open-ended data gathering.