Have people who think “exponentials” will get us to agi thought about the fact that every exponential curve eventually turns into an S curve?
Akash Wasil
 @AkashWasil
Apr 19
Dario Amodei speaks to NYT: – Says AGI (ASL-4 systems) could be achieved between 2025 and 2028. ⏳ – Says that AI ultimately should not be in the hands of private actors 🌎 – Compares wielding AGI to being "a king" 👑 – Expresses concerns about concentration of power 😮

Apr 19, 2024 · 8:33 PM UTC

16
3
1
59
I do expect significant progress in the next decade, but it’s far from obvious that further scaling up transformer based models is going to do the trick.
1
17
Replying to @binarybits
With the exception of the expansion of the universe?
1
Is that exponential?
1
more replies
Replying to @binarybits
yes, but no one knows where we are on it. not to mention, it's not one S curve but a succession of them from model architectures to hardware advances.
1
4
Yes I do think we will likely have a succession of s curves based on different models, and if we zoom out enough it might look like a smooth upward curve. But it’s easy to imagine a few years when transformer-based models hit a wall and we have to wait for something new.
3
Replying to @binarybits
I have an article on this next Tuesday. It is a belief reeking with the ignorance of scientific history
1
2
Looking forward to it!
1
Replying to @binarybits
Have those think about “exponentials turning into s-curves” thought about the possibility that it’s actually hyperbolic?
1
4
Can you unpack that for me? I’m not sure what that would mean.
1
1
more replies
Replying to @binarybits
Is anyone really defining AGI the same way? 😂
Load more