this post was submitted on 14 Jul 2023
61 points (95.5% liked)

Showerthoughts

29578 readers
862 users here now

A "Showerthought" is a simple term used to describe the thoughts that pop into your head while you're doing everyday things like taking a shower, driving, or just daydreaming. The best ones are thoughts that many people can relate to and they find something funny or interesting in regular stuff.

Rules

founded 1 year ago
MODERATORS
vatlark@lemmy.world
forkball@lemmy.world
quinten@lemmy.world
Thekingoflorda@lemmy.world
61
The problem with AI alignment is that humans aren't aligned (lemy.lol)
submitted 1 year ago* (last edited 1 year ago) by preasket@lemy.lol to c/showerthoughts@lemmy.world
29 comments fedilink hide all child comments
 

I'm sure there are some AI peeps here. Neural networks scale with size because the number of combinations of parameter values that work for a given task scales exponentially (or, even better, factorially if that's a word???) with the network size. How can such a network be properly aligned when even humans, the most advanced natural neural nets, are not aligned? What can we realistically hope for?

Here's what I mean by alignment:

  • Ability to specify a loss function that humanity wants
  • Some strict or statistical guarantees on the deviation from that loss function as well as potentially unaccounted side effects
you are viewing a single comment's thread
view the rest of the comments
[–] preasket@lemy.lol 6 points 1 year ago (1 children)

Sure, tbh, I think ChatGPT is overhyped. It can be useful, but it's nowhere near AGI. I even have a controversial opinion that the rate of progress will not be exponential - it will be logarithmic, because, I think, the data will be the constraint.

[–] Zo0@feddit.de 4 points 1 year ago (1 children)

I'm not gonna go too deep into it because I'm not qualified to, but I think the issue currently at hand, is that we're throwing stuff at the wall to see what sticks. Most of the AI models currently used in different branches are being used because they showed promise in the original problem they were designed for. All these tools you see today were more or less designed over than 30 years ago. There's a lot of interesting stuff being done at an academic level today but we (understandably so) don't see those in an everyday conversation

[–] preasket@lemy.lol 4 points 1 year ago (1 children)

The idea of backpropagation and neural nets is quite old, but there's some significant new research being done now. Primarily in node types and computational efficiency. LSTM, transformers, ReLU - these are all new.

[–] Zo0@feddit.de 2 points 1 year ago

Haha reading your other replies, you're too humble for someone who knows what they're talking about