| 1 |
It is very hard to judge. While LLMs and generative computer vision models are crazy useful tools, those people calling for halting their development, etc., are needlessly paranoid. Those models have no agency of their own, the models themselves are harmless (however, spreading fake news and such [s]will be[/s] already is much easier with them and that is a big threat, but that won't be slowed by halting their development). All of it is just a bunch of statistics applied to a huge pile of data.
|
1 |
It is very hard to judge. While LLMs and generative computer vision models are crazy useful tools, those people calling for halting their development, etc., are needlessly paranoid. Those models have no agency of their own, the models themselves are harmless (however, spreading fake news and such [s]will be[/s] already is much easier with them and that is a big threat, but that won't be slowed by halting their development). All of it is just a bunch of statistics applied to a huge pile of data.
|
| 2 |
\n
|
2 |
\n
|
| 3 |
Now, their strength depends on how many task can be "solved" with pure statistics; I bet a lot of people thought this approach can't get us that far in language generation, but it did. ChatGPT and others certainly have some interesting emergent capabilites - will it be enough to correct scientific articles in, let's say, 2-5 years, or will we need something stronger than transformers? If it understands how code works and can write unseen code to deal with new tasks, why couldn't it grasp some physics concepts?
|
3 |
Now, their strength depends on how many task can be "solved" with pure statistics; I bet a lot of people thought this approach can't get us that far in language generation, but it did. ChatGPT and others certainly have some interesting emergent capabilites - will it be enough to correct scientific articles in, let's say, 2-5 years, or will we need something stronger than transformers? If it understands how code works and can write unseen code to deal with new tasks, why couldn't it grasp some physics concepts?
|
| 4 |
Time
will
tell.
I
personally
think
something
a
bit
stronger
will
be
needed
(
pure
text
can
only
get
us
so
far
and
model
capacity)
but
we
might
be
on
the
right
track.
|
4 |
Time
will
tell.
I
personally
think
something
a
bit
stronger
will
be
needed
(
pure
text
and
model
capacity
can
only
get
us
so
far)
but
we
might
be
on
the
right
track.
|
| 5 |
\n
|
5 |
\n
|
| 6 |
What I thought was interesting was that reinforcement learning (AlphaGO - destroying that GO master, AlphaStar - beating the strongest StarCraft player, ...). But after having a (very short, potentially wrong) look at it, it seemed to me that it is just this old Agent/Graph search paradigm enhanced by some neural networks - it feels different from how humans think; we will need imho a different technique to create a model that can be as general as humans are and that can get good in games/driving/ThisTypeOfTask only after a few trials. Still, it found its use in protein identification, perhaps there are other niches where it can be as effective as well.
|
6 |
What I thought was interesting was that reinforcement learning (AlphaGO - destroying that GO master, AlphaStar - beating the strongest StarCraft player, ...). But after having a (very short, potentially wrong) look at it, it seemed to me that it is just this old Agent/Graph search paradigm enhanced by some neural networks - it feels different from how humans think; we will need imho a different technique to create a model that can be as general as humans are and that can get good in games/driving/ThisTypeOfTask only after a few trials. Still, it found its use in protein identification, perhaps there are other niches where it can be as effective as well.
|
Thoughts on AI?