00:04[Music] Language models encode knowledge. I have said it before, I will say it again: language models encode knowledge. And if you've seen the output of GPT-3, I think this should be readily apparent. And to be clear, the audience for this video is people who have already seen what GPT-3 can do and understand its potential, so if you don't know what that is, go look it up. But if you do, the question is: how can you utilize this technology to make society a better, more functional, more integrated system to live within? Because the knowledge that these models encode, they can be just about anything, and that's incredibly powerful.

00:51So to start, with some background. Three years ago I wrote a manifesto called the Cognicist Manifesto. The purpose of this manifesto was to propose a new socio-economic system driven by language models that encode collective belief. And I've done a series of videos on this topic, including one video specifically on GPT-2. But with this video I want to cut to the chase, chop chop chow chop. If you want more details, more information, if you want to know the how and the why, you can watch that series of videos, I'll include a link in the description below. If you want even more details, I'll even include the link to the manifesto. Wow. Ah fancy, two for the price of one.

01:40But this video, the purpose of this video is to as succinctly as possible describe how to utilize GPT-3 to create a socio-economic system that solves the problems that we face with how society does sensemaking in the modern era. How to move past the post-truth era. It's how to fix the information ecology. It's how to fix the internet. It is a means to very literally hold people accountable to their word. Imagine a means to hold politicians to their promises, to find the people who most contribute to collective well-being and elevate their position in society. My belief is simple: those who predict the future best understand the truth the best. Many approaches attempt to make use of this concept: prediction markets, idea markets, futarchy, even science itself just based around prediction. But all fall short.

02:29And so what I'm about to discuss is how to achieve a system in a way such that society directs itself towards a better and brighter future that is increasingly aligned with the truth as time moves forward. And more specifically, in such a way that collective well-being is increased over time. Won't you like that, huh darling? Don't you like that, increased collective well-being, everybody happier over time as time moves forward, wouldn't that be neat, wouldn't that be grand? You start with a language model. That's GPT-3. The model needs to be able to read text that's input and output text that resembles what was input. That's GPT-3.

03:17It needs to be able to distinguish between sources of information, accounts. Those sources of information need to be able to indicate to the model how aligned with their perception of truth the output is. That's a voting system. The model needs to be able to distinguish the date of statement of the information being input. Not hard, a timestamp. The sources of information need to be able to share predictions with the model, and the model needs to be aware of the distinctions between a prediction and a statement. Very easy, a checkbox, which has existed since like 1995, but also like GPT-3 can pretty much do it automatically. So the model needs to be able to learn to associate which sources of information input predictions that most align with the future representations encoded by the model.

04:05It needs to learn who predicts stuff best. The sources of information need to be able to iteratively interact with the model in such a way that as information and sources input their perspectives on the state of truth, the model outputs statements and predictions representing the average collective perception of truth. And that's like the hard part, but GPT-3 can do it out of the box. The problem is the average collective perception of truth right now contains a lot of falseness in it, falsitude, wrongness, there's a lot of misinformation. Yeah, we, we got a bit of a misinformation problem going on right now. There's a lot of, there's a lot of craziness in the what the people believe, in the brains and in the hearts.

04:55The model needs to learn over time to output statements formed from the content of information sources that more frequently post information on the lines of future representations encoded by the model, and needs to learn to listen to the best predictors more over time. The model should distinguish between sources of information that post information that the collective disagrees with at the time of posting, which are later validated by the collective. If someone says the sky is falling and everyone disagrees with them and then the sky falls, the models need to know to listen to that person more. If someone says there's going to be a pandemic, we need to do something about it, and then a lot of people don't listen to that person, once the pandemic gets out of control the model should listen more to the people who said it was gonna be a problem.

05:44And the representations learned by this model need to be written to an immutable distributed database. Um, what would that sound like? Blockchain. Blockchain. The rankings of these models need to come to represent a similar value as money in this new society. That's like, um, bitcoin but like more actually associated with social well-being and like social good and like people being happy. That's it. Supplement the monetary system with a secondary token that represents the ranking by a collection of distributed models learning to identify which sources of information are most aligned with the future, the future, the future. Now this sounds precise and pedantic, I think that's because it needs to be, because you know like complex problems usually require like some complexity in their solutions.

06:31Uh, maybe, I, I, I really don't think you're just gonna magically fix everything that's going on right now with like one law or by like voting in a different dude to like the position of power, or like, you know, voting for a different group of people. Yeah, no, you got to build new models that replace and supplant the old ones. But it is a, or possibly probably, the solution. And I think that if I can, you know, predict GPT-3 three years ago, maybe this is a pretty good prediction too. Even if it doesn't seem like it to you in this moment, and then in three years when you're looking back and it's incredibly obvious that what I was saying was true, except for you have no idea who I was because you never watched this video.

07:19It goes on. You know what, actually it's really simple. Let the people talk to GPT-3. Let them share predictions. Let GPT-3 output the average vision of the future. Key point: let the people vote on the likelihood of that output vision. Let the model learn to listen to those who are validated to be correct in the end. So the model encodes a moving representation of what society thinks the future will be like and helps society to steer itself towards that positive future and away from negative ones. What? Reward those who are right with a token that is like money but isn't.

08:07And for the cherry on top, also let the people vote on whether they want those futures to happen or not. Let the people state actions taken to achieve those futures, and let those whose actions which are associated with achieving those positive futures be rewarded more. Now I know what you're thinking: can GPT really do this in three whole years? It seems so far away, how is it gonna be able to do it? I saw that somebody tricked GPT-3 and it was tricked, they tricked it, and it's never ever gonna get better than what it is right now. Will GPT really be able to do this all in three years? Um, yes. Um, yes. Oh yes. Oh my god yes yes yes yes, just believe me, yes oh my god yes Jesus [ __ ] Christ yes yes yes yes [ __ ].

08:57Do a bot remind you in three years. Jesus Christ, yes. Listen to me, I'm pretty sure I uh predicted that language models would be able to do what they can now three years ago. Um, but then everybody said that it was wrong back then, funny how that works. If only there was like a mechanism to amplify my voice for being right in the face of collective criticism. If only when you see that guy online who's saying the crazy thing about the future, and then you say no, it's actually going to be this way, and then you know like time moves forward and then it turns out that you were right, and then you go and talk to that guy and then he's like I don't give a [ __ ] like, and he moves the goal post and there's like no consequence for his words whatsoever. Like, if there's like, I don't know, like, some sort of chain that could like tie together our past voices with our present voices with our future voices, so that we all speak collectively in one voice, so that we can all move towards positive futures and stop living in this [ __ ] crazy insanity, like wouldn't that be [ __ ] useful.

09:44I don't know. Think about it. Make a prediction. Thanks, bye. [Music] You.