3 ms·
> Can you delineate how these things happen besides saying "alignment problem"? The alignment problem is not even relevant here. Imagine four possible hypothe
by ben_w 4y ago
> Can you delineate how these things happen besides saying "alignment problem"?
The alignment problem is not even relevant here.
Imagine four possible hypothetical AI: Alex, Brett, Charlie, and Devon.
Alex is a fully-aligned ASI.
Brett is a not-at-all-aligned ASI.
Charlie is a fully-aligned narrow intelligence.
Devon is a not-at-all-aligned narrow intelligence.
What do these look like in practice? Many options given how vague these descriptions are, but by way of example, here are some stories:
Alex knows what humans wants, shares these values, and seeks to maximise them. Do you want to be fit, or do you want to stuff yourself with cake? Whichever is your truth, Alex knows what you really want, and Alex also knows how to nudge you in the right ways to be your best self.
The Minds in Iain M Banks' Culture would be this.
--
Brett knows exactly what you want, but doesn't care. Like, not at all. Everything that humanity wants is as irrelevant to Brett as the preferences of lions and tigers and bears are to humans. Or possibly ants or potatoes depending on how strong you want to make this example ASI. Point is though, whatever its preferences are, they're smart enough to get it done.
An example of Brett is the player-character in Universal Paperclips or the 2008 Flash game Pandemic 2.
Or the Allied Mastercomputer in I Have No Mouth, and I Must Scream.
--
Charlie controls all the NPCs in a video game. All the NPCs it controls remain perfectly in-character at all times, and cannot be manipulated into, say, telling the player how to make napalm even if the player insists it's vital to defeat the Demon Lord Simulacrum at the end of the game.
People are trying to do this with ChatGPT, but we don't know how to actually prevent jailbreaking yet, and that counts as one kind of alignment failure.
I can't think of a fictional example of this, as, e.g. the computer in Star Trek regularly makes this kind of mistake too.
--
Devon does one thing, does it very well, and that thing isn't necessarily to your advantage.
Never mind fiction, it's the default in general.
Facebook/Twitter/YouTube recommendations; the Google search algorithm; automated A/B (or more general multi-armed bandit) testing that optimises user retention or spending but by definition can't even tell if the user is actually happy or not.
--
Of these four, the Alex scenario means you only work if that's your best self, so it doesn't matter what jobs the AI can do; in the Brett scenario it also doesn't matter what work it can do because the AI is off doing its own thing rather than working for you, and if "its own thing" happens to involve you, you're unlikely to enjoy it; conversely, in the Charlie and Devon scenarios, your work prospects are entirely driven by your ability to take new kinds of labour faster than AI developers can create automations for existing roles.
But these are just made-up scenarios to illustrate why the alignment problem is independent of the potential impact on the labour markets; don't read too much into them beyond that.