Decentralizing GPT into Cultural DNA with Matt Pirkowksi
https://www.youtube.com/watch?v=l0keANU0xx0
Transcript:
(00:00) there's a couple times that you've referred to like sort of the centralization of the generative models as either Iris or I would say definitely active inference model generative models applied to this as well that you know they're sort of run by one set of people one thing that you know we haven't gotten a talk chance to talk about and we'll probably have to talk about more in a future conversation is the notion of how do you decentralize the compressions of uh the knowledge that is integrated by those models into you know
(00:34) some sort of distributed data store because I've always viewed it more as that there would be many many different irises people often ask me why there isn't just one and I mean I think you know exactly why there shouldn't be just one and that there would be some energy bound uh distributed data store that is taking these compressed tensors or compressed represent presentations of knowledge as they are evaluated by or used by many different models and stored in a secured way uh that many future models can build the representations
(01:16) from I I I I view it more like like there is a DNA strand that is the collective cultural ethos and World observations and that these models are more like the molecules that do the correction when there is the replication of a DNA strand and that ultimately is sort of the model that you know even you and I are have not gotten much work into and really needs to develop is that there needs to be sort of an evolution or a merging of these blockchain structures or just distributed data stores and the knowledge in its very dense form into
(02:02) that Beyond just you know like saving a copy of it you know something more fundamental yeah I mean that like literally so the conversation I mentioned earlier that I had yesterday uh like that exact idea you that was something that I was I was I was speaking with um you uh uh Jordan stuff about when we were discussing you know especially the role of something like blockchains uh the role of something like Bitcoin when it comes to and not Bitcoin specifically but proof of work having something that is highly resilient that ends up being a kind of
(02:39) centralized repository for highly compressed encodings uh that can represent as a kind of um DNA over the space of emergent systems that have models right because DNA itself was the way in which like you're saying these molecular mechanisms and all of the emergent structure above them in its exploration of function space yeah created a feedback process that compressed down into molecular space representations of what works that can then be back unpacked again and experimented with into function space that's this DNA
(03:14) function right but that was all in the in the being a model world but in the having a model world we have to actually because all of this becomes explicit we have to become self sufficiently self-aware that we actually create uh or Ste the creation of a similar kind of um and this is weird because there's this Duality here it's a if you did it on a blockchain like like something like a Bitcoin you want it to be capture resistant so it could be highly centralized but because it's capture resistant that centralization actually
(03:48) can become a tool that increases the capacity for decentralized coordination because if every decentralized system can contribute to it and take away from it to some extent then it they could focus on their attention on it without having to necessarily be tightly coupled with it except when they want to contribute to it or or or you know decode from it right because the decoding can always happen for free the encoding comes with a cost if you want to inscribe yeah but to the extent that anyone can benefit from decoding like
(04:19) think about like what would happen if a a an advanced model were basically leaked and its weights were put onto the blockchain right like now this may or may not be desirable at this point like we might argue that we need a better container for that before it actually makes sense or or could be advantageous and that leaking a powerful model and and making it irreversibly accessible on something like a blockchain with respect to its weights or a compressed version of its weights um or a program that could generate those weights if that was
(04:52) a valid compression um you know it's unclear whether that's desirable or not desirable right now but my point is that these kind of mechanisms are almost with 100% certainty going to be the kind of um places where we encode these highly compressed distillations of our Collective experimentation over over this space of of of computation Ai and also complex coordination via systems um such as uh the entire other world of like proof of stake and exploration of Distributing governance decentralized economics all
(05:33) of that kind of uh you know that space where we're seeing a massive amount of experimentation right now um we're seeing a lot of failure we're seeing a lot of parasitism we're seeing a lot of craziness but we're also seeing you the seeds of of potentially interesting or useful patterns you know to the extent that you know much like uh mutations to the extent that those come into the world and actually provide value um how do you re-encode them into a space where anyone can then replicate those patterns in the world right so if a particular
(06:02) Dow finds a set of configurations and weights over their mechanisms that actually gives rise to really positive capacity for a group to collaborate coordinate and and and be generative as opposed to parasitic or extractive um that's a kind of pattern that should be encoded and shared right and it should be accessible with very low cost if not free um in something like a blockchain that cannot be easily captured or corrupted uh for for local interests of of any other entity on Earth so yeah I mean I think exactly like the these the
(06:38) the degree to which something like an iris is is centralized or decentralized you know it's an interesting question because to some extent it's a question of you know adoption and evolution and and how all these different threads of that um system um come into usage locally right I mean obviously you you want to spread it far and wide and have those um sort of seeds cast upon the Winds of potential throughout human uh interest and experimentation such that you know they land and take root and grow in fertile soil um obviously because we're
(07:13) humans people will try to use it for abuse as well and all these other things that we do but um but yeah me it's again I really do see it as this recapitulation of this natural process at the level of having models as oppos to being models and we are trying to figure out how to you know what it looks like to create the kind of social mechanism that is the analog of of DNA right yeah and there I I would add just in summary to try to link it all back together that um there definitely needs to be this could very
(07:51) easily go wrong where it's it's trending towards reducing that uncertainty in our knowledge representations right I I've already seen over like two years years with GPT first there was a very wide space where you could explore the long tail you could talk about things that aren't well established and more and more it's like there's a regulation into this is the way things are and because they are controlling when that update happens it's harder and harder to do research to that so there whatever compression goes into that distributed
(08:21) data store it needs to have that preference towards the exploration of the uncertain uncertainty like you were you were speaking of to um variational free energy versus what was it expected free energy expected free energy like there needs to be uh a tendency not in our compression of knowledge to say well this represents truth in its totality and rather this is a representation that is inherently mutating and changing over time um and something that is ex existing to expand our capacity over time um to exist and Thrive where
(09:10) our neural mechanisms of encoding memory of encoding models um are just a little bit too lossy for long-term yeah and then there's kind of like sedimentary Dynamic there as well where you to some extent the more the longer that certain aspects of that model have remained stable the more they kind of migrate down to a more you know a deeper and more stable structure that doesn't mean that that that layer can't be changed it does it does get more difficult to change the more energetic constraints on that yeah yeah the more that you have energetic
(09:44) constraints the more that you have dependencies the more that you have a history of the world U being congruent with that representation it does get more difficult to change as it migrates down that uh stack but uh but the understanding is that the entire stack all the sedimentary layers even though proportional to their level they will be like the top layers will be more volatile proportionately than the the bottom layers but they are all subject to change if you have a uh you know observations or empirical uh experiences
(10:18) of sufficient surprise right and we you know and that reflects what we are like as well right like we have you know we we can be shocked to such an extent that our entire world models uh it is revealed to us that our entire world models down to their core need to be recalibrated uh if we encounter experiences that are sufficiently outside of our model although unfortunately we increasingly are leaning into the tendency to to lean into confirmation bias and to try to say you know you know I'm going to ignore the signals the world is giving to me
(10:53) I'm going to try to force my model on the world again it's like there's always a certain element of interplay between like whether it's beneficial to try to like put your model onto the world or allow the world to transform your model but it always has to be in dialogue and it becomes inherently totalitarian when you decide that no matter what I'm not going to change my model and I'm going to push all of the responsibility for transformation out onto the world as opposed to taking any of that uh in your own [Music]
(11:28) model [Music] [Music]