0:00
So why the future of AI should not be
0:04
centralized and why control may be the
0:06
real threat? I mean is collective
0:10
intelligence versus centralized AI kind
0:13
of the focus and one of the core ideas
0:16
is that AI shouldn’t be controlled
0:19
guided by a collective intelligence.
0:22
What what all does that mean? Okay. So
0:25
there’s two at least two but we’ll
0:27
contrast two different approaches to
0:29
building the next generation of
0:31
intelligent AI systems. So the standard
0:33
approach that most people are following,
0:35
open AI and Google and Anthropic and all
0:38
the major AI leaders are primarily
0:42
taking the approach that if we want a
0:44
smarter version of chat GPT or a smarter
0:47
version of Gemini or cloud or any of
0:49
these models that people are familiar
0:51
using, the way to do it is to have
0:54
bigger and bigger data centers with more
0:55
NVIDIA GPUs in there and train them
0:58
train the model on more and more data.
1:00
So more data, more computing power gives
1:03
you just a smarter AI. And this has been
1:06
true. So it’s been true for the last 10
1:08
years. You know, each time they put more
1:10
GPUs in and train it on more data, the
1:12
model comes out smarter. And that’s why
1:14
every few weeks it seems like somebody
1:16
comes out with, hey, now my model’s a
1:17
little bit better. And then another one
1:19
comes up.
1:20
So that’s the standard approach. The
1:22
problem with that approach is that
1:24
nobody knows how the models represent
1:27
the information in their little model
1:29
brains. It’s a, you know, it’s a big
1:32
black box and it makes it very difficult
1:34
to predict what they’re going to do. So,
1:36
we mentioned in the early days of AI,
1:38
the first 30 years, you actually
1:40
programmed the rules in. Well, when you
1:42
were programming the rules in, you knew
1:43
exactly what rule you programmed in, and
1:45
you kind of know how the AI was going to
1:47
behave. But once you went to machine
1:49
learning, it was a double-edged sword.
1:51
The good part was you saved all that
1:53
effort of programming in the rules. Wow,
1:55
that was great. But the bad part is you
1:57
had no idea what the AI was really
1:59
learning. And so you get this giant
2:01
black box. And that’s why these models
2:03
hallucinate. And there’s a huge amount
2:05
of effort. The current state of AI
2:07
safety is you build a model, you’ve
2:09
trained it, you don’t know. It’s a black
2:11
box. You don’t know how it represents
2:13
information. And then you have an army
2:14
of humans ask it to do bad things. You
2:17
say, “How do I build a bioweapon?” And
2:19
if it tells you, then the humans say,
2:21
“No, you can never tell someone how to
2:23
build a boweapon.” And it says, “Oh,
2:25
okay. My mistake. I will never tell
2:26
somebody, right? I mean, we’ve all had
2:27
those interactions.
2:29
The problem is it’s like a game of
2:30
whack-a-ole. It’s impossible to think of
2:32
all the bad things and tell it not to
2:35
say them. And the reason that we’re in
2:37
that situation is because we haven’t
2:39
designed it to be safe from the
2:41
beginning. We just allowed it to sort of
2:43
train itself. Then we have the black
2:45
box. Now we’re trying to figure out what
2:47
the blackbox learned and what it might
2:49
do. And after the fact, we’re trying to
2:51
put these guard rails and safety on. So
2:54
that’s the current approach. And
2:55
needless to say, probably from the way
2:57
that I’m describing it, it’s pretty
2:58
clear that I don’t think that’s the
2:59
right way.
