Bob Givan

AI professor. he/him/his. Views my own.

[Bob Givan @rgivan1]

@parakeetnebula https://t.co/5VR18FBT1I

[Bob Givan @rgivan1]

@gwern @stuartbuck1 @GaryMarcus @ylecun There remains a question is whether gradient descent can find such a model. But yes the required signal is present in the loss function to prefer a predictor that models the world over one that doesn’t.

[Bob Givan @rgivan1]

@christiandean_ @liron @dwarkesh_sp Not sure we want the unpredictable child in the nuclear launch control room.

[Bob Givan @rgivan1]

@CruellaUllah @ChessifyMe 1. Qe1 does not lead to mate in one after the response 1. … b6

But 1. Qf1 b6. 2. Qa6#

[Bob Givan @rgivan1]

@Algon_33 @ESYudkowsky One could say a single starting idea is “local” to anything you can realize in a small number of easy steps from it and further away from something that require a lot of reasoning steps to reach.

Processing a problem a lot adds definitions and lemmas that contract this distance.

[Bob Givan @rgivan1]

@Algon_33 @ESYudkowsky I think another way to say it would be that most attempts to think about a solution, wherever they begin, lead quickly by reasoning to failure. If you have thought about a complex problem enough you can have identified short reasoning paths to failure from every starting point.

[Bob Givan @rgivan1]

@gbrlvv @HypoGrazer @RokoMijic @slatestarcodex I don’t think the numbers here are prison sentences but reward payouts. In the Wikipedia article similar numbers are used but they are punishments so this is confusing you to label the example as wrong

If you see the numbers as rewards then you see they do the best tight lipped

[Bob Givan @rgivan1]

@s_batzoglou @BoWang87 @GaryMarcus @sama Of course I remember Kostas. Will see if I can find time to read his thoughtful work.

[Bob Givan @rgivan1]

@Melvinouk @EvidenceProf @SellLowest @AppreciatesNick It’s significant not for the truth of the claim but for its impact on his credibility, where it is of decisive significance.

If he admits presenting police input as his own then we must doubt all his statements and to allow that any information police had was provided to him:

[Bob Givan @rgivan1]

@MarouenAhmed @RandolphCarterZ @AISafetyMemes @robertskmiles If in fact we could “hard code” the goal as stated, then barring unintended consequences of that goal (bugs in the goal), it would have no motive to change its code to change the goal.

But that is a big “if”. We don’t know how to “hard code” binding goals in large neural nets.

[Bob Givan @rgivan1]

@perrymetzger Do you think consciousness is a prerequisite for deceptive behavior?

The superhuman poker programs learn to bluff through self play. They are not programmed to deceive, they learn it on their own. And surely are not conscious.

I think consciousness is irrelevant here. Safety.

[Bob Givan @rgivan1]

@perrymetzger I said nothing re Whisper or consciousness

Thought experiment: what would we get if we simply trained AIs to try not to be shut down? (No other goal)

If this could be done we would get very useful AIs during the period when they lacked power to stop shutdown forcefully. Then?

[Bob Givan @rgivan1]

@s_batzoglou @pmarca This essay is built squarely on unjustified dismissal of real and major risks.

[Future of Life Institute @FLI_org]

How would a nuclear war between Russia and the U.S. affect you? Who would survive?

Watch and share our most realistic nuclear war simulation yet:

[Scott Niekum @scottniekum]

I want to address some misconceptions that I've been seeing in the discourse about catastrophic AI risks. None of this particularly novel, but these points are typically scattered in long-form pieces that most people won't read, rather than collected and stated briefly. 🧵

[Bob Givan @rgivan1]

@mmamags @lee_kovarsky They haven’t been shown otherwise. So they are as necessarily innocent as you are (making a presumption that you haven’t been shown to have committed this crime).

[Bob Givan @rgivan1]

@mmamags @lee_kovarsky Yes, if you mean “some erroneously charged people”

[Bob Givan @rgivan1]

@allen_brutus @ESYudkowsky @tszzl Since we don’t know how to set the goals (let alone do it without bugs or missing pieces), the goals of the system will in some sense be arbitrary.

An arbitrary system doesn’t turn out to be an Apollo rocket.

How intelligent the system is is orthogonal and doesn’t help.