The promise of Microsoft Recall is that extremely early AGIs will have all the info they need to launch vast blackmail campaigns against huge swathes of humanity, at a time when LLMs are still stupid enough to lose the resulting conflict.
@esyudkowsky
-
Harmless Supernova Fallacy: Bounded Therefore Harmless
By
–
Harmless supernova fallacy, "bounded therefore harmless". https://
arbital.com/p/harmless_sup
ernova/
… -
Understanding Giant Tensors and Writing Goodhart-Proof Objectives
By
–
It doesn't need to be representing new physics, for us to have trouble understanding what the giant inscrutable tensors mean — or writing airtight Goodhart-proof objectives that operate over them, a much higher requirement of proficiency than current interpretability efforts.
-
Building Objectives for Chess Search Trees with Known World States
By
–
We can easily build objectives for chess search trees, because: the complete state of the world is represented by a type known at compile-time; and we can handwrite a function to say exactly what states of the world we want, as described by that known representation. An AGI
-
AI Objectives: Learned vs Human-Coded Implementation Types
By
–
It's also described in my "Creating Friendly AI" from 2001. What sort of objectives do *you* have in mind? Learned? Human-coded? What's their type signature?
-
Intentional AI Development and Gradual Capability Emergence
By
–
– People are trying to build it on purpose
– ChatGPT didn't need to start talking "suddenly" to start talking at some point -
Chollet’s Debate Participation and Availability Questions
By
–
Maybe it turns out Chollet doesn't like debates. Has he done any others? When did he sign up to be in debates and not just (end the world via) building AI frameworks? I'm up for it if he is, but don't want to be the guy claiming he's got to provide that service for free.
-
Unconstrained Solution Search in Natural Selection and SGD
By
–
Solving some outer reward or loss via a process that looks around for solutions and isn't much constrained or much legible in which solutions it finds. Natural selection and stochastic gradient descent are both examples.
-
Bunker construction vs AI safety approaches and their feasibility
By
–
In real life, yes, I'd be far more cheerful about trying to tackle the structure for an apocalypse bunker than for an artificial superintelligence. Just, like, not with the level of thinking or kind of techniques that safetywashers currently go for.
-
Gathering empirical data on apocalypse for apocalypse handling
By
–
Obv. they're causing the apocalypse to gather empirical data for how to handle the apocalypse