-
0:04:32 · watch on youtube.com↗
Their wordsAnd my take is that this performance increase that you're observing, it is actually orthogonal to intelligence. You are not really measuring intelligence because your benchmark can be hacked purely by preparing for it, by memorizing things in advance.
- 2 min later
-
0:06:29 · watch on youtube.com↗
Their wordsWhat they're primarily memorizing is functions, programs. And these programs do generalize to some extent.
- 4 min later
-
0:10:49 · watch on youtube.com↗
Their wordsthe kaleidoscope hypothesis is this idea that the world in general and any domain in particular follows the same structure, that it appears on the surface to be extremely rich and complex and infinitely novel with every passing moment. But in reality, it is made from the repetition and composition of just a few atoms of meaning.
- 5 min later
-
0:15:40 · watch on youtube.com↗
Their wordsThere was this pattern recognition shortcut available, and this shortcut would be taken every single time. You could not learn generalizable discrete programs via deep learning.
- 3 min later
-
0:18:34 · watch on youtube.com↗
Their wordsAnd a curve is, you can represent many things with a curve, but it's a very bad substrate to represent any sort of discrete computation.
- 6 min later
-
0:24:13 · watch on youtube.com↗
Their wordsthey're very good at giving you things that are directionally accurate, but not actually accurate. So if you want to use them well, you need this post facto verification step.
- 11 min later
-
0:34:50 · watch on youtube.com↗
Their wordsAnd where intelligence comes into play is the ability to adapt to things that are way out of the distribution, because the real world It's not a distribution, right? Every day is new. Every day is different, but you have to deal with it anyway.
- 2 min later
-
0:36:50 · watch on youtube.com↗
Their wordsAnd I'm actually quite confident that at some point in the future, maybe in the near future, we'll be able to create a system that can actually address this issue of novelty, that can actually take what it knows and recombine it in truly original ways to address completely new problems.
-
0:37:11 · watch on youtube.com↗
Their wordsFor instance, one of the things you cannot do with LLMs today is develop new science, right? Because the best they can do is speak back to you some interpolation of something they've read online, right?
- 9 min later
-
0:46:30 · watch on youtube.com↗
Their wordsEmpirically, I think intelligence probably peaks in your early 20s. That's when you're the most efficient in acquiring new skills.
- 1 min later
-
0:47:07 · watch on youtube.com↗
Their wordsSo if you want to acquire any kind of skill that deals with low-level sensory motor primitives, like you want to get really good at playing an instrument, you want to get really good at singing, you want to acquire a native accent in some language, you should do it before you're 15, typically.
- 13 min later
-
0:59:47 · watch on youtube.com↗
Their wordsI think we don't really understand enough about the human mind for that understanding to be a useful guide when it comes to creating AI.
- 4 min later
-
1:04:01 · watch on youtube.com↗
Their wordsI think ultimately to measure intelligence, you're going to want not a static dataset. You're going to want a task generation process, and you're going to ask it for new tasks.
- 6 min later
-
1:10:20 · watch on youtube.com↗
Their wordsI'm a big believer in the idea that the solution to the problem of intelligence must be co-evolved with the challenge, the benchmark.
- 1 min later
-
1:10:59 · watch on youtube.com↗
Their wordscoming up with that program is as hard as coming up with AGI. It is in fact the same kind of thing. You basically need AGI to create the challenge that AGI is a solution to.
- 5 min later
-
1:15:57 · watch on youtube.com↗
Their wordsI'm really a big believer in the fact that no cognitive process in the human mind is pure System 1 or pure System 2. Everything is a mix of both.
- 14 min later
-
1:30:22 · watch on youtube.com↗
Their wordsmy guess is that what makes human intelligence special is that it's not at either end of the spectrum. It's somewhere in between. You have access to a very large, very rich bank of abstractions, of ideas and patterns of thought, but you're also capable of recombining them on the fly to a very meaningful degree.
- 7 min later
-
1:36:59 · watch on youtube.com↗
Their wordsmany folks, even folks who are machine learning experts, they have this misconception that you can only overfit if you are directly training on something, if you're using this training data. That's not the case.
- 4 min later
-
1:41:17 · watch on youtube.com↗
Their wordsAnd in general, I think The research direction that is the most promising to me is combining deep learning with discrete program search.
- 7 min later
-
1:48:36 · watch on youtube.com↗
Their wordsAnd this is why LLM modulo some external verifier is so powerful. It's because you are cutting through the combinatorial explosion problem that would come with trying to iteratively try every possible solution.
- 2 min later
-
1:50:47 · watch on youtube.com↗
Their wordsthe way I see it, intelligence is a tool that is used by an agent to accomplish goals. But it is related to, but it is separate from your sensorimotor space, for instance, or your ability to set goals.
- 7 min later
-
1:58:13 · watch on youtube.com↗
Their wordsYou are downloading these very rich reusable mental templates, like really mental building blocks. And then you can, in your own brain, you can recombine them. You can reapply them on new problems. It makes you more intelligent, like literally more intelligent.
- 6 min later
-
2:04:20 · watch on youtube.com↗
Their wordsThe way the deep learning community as a whole is finding new things is by trying everything else and eventually hitting the thing that works.
- 4 min later
-
2:08:40 · watch on youtube.com↗
Their wordsYou cannot really build good mental models, good world models passively. purely by observing what's going on around you with no goals of your own. That's not how it works. Goal setting is a critical component of any intelligent agent.
- 2 min later
-
2:10:26 · watch on youtube.com↗
Their wordsSo I believe language and thinking are separate things. Language is a tool for thinking.
- 16 min later
-
2:26:37 · watch on youtube.com↗
Their wordsI think it is not strongly tied to substrate. So in principle, you should be able to implement consciousness using the right functional dynamics in silicon.
- 7 min later
-
2:33:44 · watch on youtube.com↗
Their wordsPersonally, I see creating a GI as a scientific problem, not a religious quest.
- 3 min later
-
2:36:21 · watch on youtube.com↗
Their wordsApplied intelligence is grounded in experimental data. And if you are very intelligent, then experimental data becomes a bottleneck. So it's not like you're going to see a runaway intelligence explosion.
- 2 min later
-
2:38:31 · watch on youtube.com↗
Their wordsSo the difficult part, the dangerous part is not the intelligence bit. It's more like the goal-setting and action space bits.
- 3 min later
-
2:41:16 · watch on youtube.com↗
Their wordsI also think that the regulation proposals that I've seen so far are not really satisfactory. They are more leaning towards harming innovation. than protecting the public.