Marcus Hutter: Universal Artificial Intelligence, AIXI, and AGI | AI Podcast #75 with Lex Fridman

20 thoughts on “Marcus Hutter: Universal Artificial Intelligence, AIXI, and AGI | AI Podcast #75 with Lex Fridman”

1. I really enjoyed this conversation with Marcus. Here's the outline:
   0:00 – Introduction
   3:32 – Universe as a computer
   5:48 – Occam's razor
   9:26 – Solomonoff induction
   15:05 – Kolmogorov complexity
   20:06 – Cellular automata
   26:03 – What is intelligence?
   35:26 – AIXI – Universal Artificial Intelligence
   1:05:24 – Where do rewards come from?
   1:12:14 – Reward function for human existence
   1:13:32 – Bounded rationality
   1:16:07 – Approximation in AIXI
   1:18:01 – Godel machines
   1:21:51 – Consciousness
   1:27:15 – AGI community
   1:32:36 – Book recommendations
   1:36:07 – Two moments to relive (past and future)

2. LOL at 1:23:47 and then Marcus looks into the camera.
   I learned so many new ideas in this talk. Really grateful for this. Thank you, Lex!

3. 1:08:35 i wonder what was cut there

4. 1:13:37:250 "Infinity Keeps Creeping Up Everywhere"

5. AIXI is to AGI as the one time pad is to cryptography.

6. Just a fact about self-improving programs. There's a type of compiler in programming that's called optiminzing compiler, some have optimizer as a standalone program that runs over your code (text, intermediate representation) and outputs an optimized version of it that's not slower than the original. However, if you feed the optimizer's code to itself multiple times over and over, it will very quickly reach saturation and after the 1st iteration it will simply not find anything else to improve and will output the same thing.

   This suggests that if you ask a general AI to improve its own code, it might not necessarily find anything to improve and even if it does, there are huge diminishing returns that make the whole singularity idea very improbable.

7. My intuition about a non-conscious vs a self-conscious AGI is that the first probably would follow any provided optimization function (although we would eventually have trouble seeing that it does follow these goals) while the 2nd might choose to ignore the provided optimization function and follow something that is in its own interest (whatever that might be). But that would also mean that the 2nd could be far more dangerous than the 1st and therefore it would be of utmost importance finding a test whether an AGI is self-conscious or not.

8. Redhead person wearing orange jacket and this tie… i mean you ve got to accept yourself ;D

9. What if one took all the bitcoin blockchains and had a AIXI model for a computational goal. As one solved various compression schemes it would shrink the data storage overload (huge wallet) problem.

10. This is to date the best podcast i have listened to, and i have heard most of the AI podcast. Lex, can you help identify the work connecting AiXi and the reward function based on the information content. I would really like to go through that work in detail.

11. Lex this is a really good podcast. Thanks for the hard work. I know the name of the podcast is the AI podcast, but I wonder how seriously people take the idea of general artificial intelligence? To me it seems ludicrous. But this very smart gentleman (and many many others) seem to take the idea seriously.

12. While it is probably useful I don't think the AIXI model at https://youtu.be/E1AxVXt2Gv4?t=2337 looks like it is intelligence. For it to be intelligent it would need to analyze the reward options in a context (In humans it's consciousness) and then prioritizing and dismiss them in advance. The AIXI model just looks like machine learning. Layman BTW.

13. Excellent stuff as usual Lex! Very interesting guest, and it was good to hear you briefly discussing CA. These fascinate me – Poundstones "The Recursive Universe" is such a wonderful book that I'd recommend to anyone interested in CA's and Conways Game of Life in particular.

14. Just in case, Occam's razor is not about "simple". It is: "when presented with competing hypotheses that make the same predictions, one should select the solution with the fewest assumptions". So for example, Newton's law of gravity is not that simple – Calculus was invented to solve, but is based only on two equations/assumptions – for a force as a function of mass and distance and for an acceleration as a function of force and mass. Before that, Ptolemy's model of planet motion was way simpler and didn't require Calculus at all, but was based on a large number of coefficients of unknown origin. So Occam's razor prefers complex language and minimal set of axioms over simple language and large set of axioms.

15. 6:57 – "Crazy models that explain everything but predict nothing". In terms of machine learning, I think this means that complex models tend to overfit the data and as such can perfectly explain the data but do not generalise to unseen phenomenons. I feel like this is a valid argument for Occam's razor without rely on the assumption that our universe is simple.

16. Pressing Like as soon as I hear Occam's Razor.

17. Dr. Hutter, I wonder if the reward function definition is one function (thus allowing for optimal fitting with Occam's Razor) but that it exhibits particle/wave duality, meaning that the reward function contains both motivation and incentive vectors. My heuristic is that incentive is an attractor force by which nodes converge on some other node, and my heuristic is also that motivation is kind of potential-impulse force using which the agent diverges towards another discrete node.

    I just wanted to mention this to you in case you can find anything interesting encoded in my symbol choices. To fully understand my reference to particle/wave duality, you might need to know that I see the same ratio between matter:energy and network:data. That is to say matter is to network as energy is to data. Next, I also see it could be possible that the particle/wave duality of the reward symbol causes incentive and motivation to emerge and they may also hold simple defintions related to "imperative" vs. "declarative". As a software engineer, I constantly see one symbol: an object that has functions and state, and it constantly wrangles the two perspectives between declarative and imperative. I find it very pronounced in functional-reactive programming.

    Anyway, my point is simply for you to analyze my comment thoughtfully, and consider if you can see symbol combinations that you haven't directly considered yet but might find useful. I offer these symbols to you (you as my interface to the fabric of the Universe that is trying to understand itself from the inside out). Along my lines of reasoning, I see incentive as the inverse of motivation (somehow, almost as if the reward symbol is the edge and the incentive and motivation symbols are the two connected vertices.. or maybe I'm not seeing it correctly, and thats why I need you to see if there's something there that needs to be found).

18. "But I'm not modest in this question" I like this guy

19. 1:10:28 A dead agent is good for spare parts.

20. 29:09 but if the trees die so do we in ecological collapse–arguably this individualistic definition of 'intelligence' leads us to make 'dumb' choices ecologically

Comments are closed.