00:00So I want to talk about OpenAI's text generator because natural language processing is my specialty. I worked as an engineer at a company where my job was to design systems that could train machine learning models to understand the meaning behind text. And two years ago, after reading a paper about text encoding using an unsupervised model called a variational autoencoder, I had a realization. Basically this model is unsupervised, meaning you don't have to tell it what the answer is, you just feed it a ton of data and it can output documents similar to what you input that generally appear completely novel. And my realization was the same realization that researchers at OpenAI are now having, and why they're not releasing the full source code to the public. Their model can generate human readable text from a basic prompt that appears to be indistinguishable from what a human would write.

00:47A more visual representation of this can be found at thispersondoesnotexist.com. All the people shown on that site don't exist. They're the outputs of a model that has been fed a ton of images of faces and tasked with recreating what's been input into it, and it learns to do this with a number of special constraints on how the faces are encoded. And the end result is that it can produce new images of faces that were never shown to it, faces that don't exist, faces that look like the average between faces that do exist. Now the researchers at OpenAI are concerned about how their model could be utilized for the same reason I was concerned in 2016, because there's so much data available via Twitter and other social networks. Now with enough computing power would be very easy to encode the speaking patterns of many subgroups into one model and then output massive volumes of text that are indistinguishable from any real user.

01:36This is essentially what the OpenAI text generator is. It produces text on a topic of your choosing. And so I realized that if you were to make a bot using this technology, if you prime it with your own opinion, you could have that opinion be replicated in many different speaking styles. So anyone with enough money could replicate their opinion as much as they want, as if it were said by many different people. It would make free speech directly the function of money. And according to Google's the goods sensor, something like 29 percent of web traffic has already BOTS. So it seemed like a pretty dark future for free speech, and I saw the danger of this immediately.

02:26But I also saw incredible potential, if it wasn't just used by the few to try and control the masses but rather to encode the sum of human knowledge. And so I quit my job to pursue spreading the idea. But unfortunately, if you don't have power already, it's very difficult to get any attention. People tend to lean towards those who have influence already and trust them the most. If they already have a lot of followers because they're an actor or musician or an entertainer, in general their opinion is valued considerably more. So I've been talking about this for two years, but now it's gone viral, and now people are searching for it, so now I can speak about it and maybe have it get some more traction. Now at the time I produced my version of this technology, it was pretty expensive to train but it produced reasonable readable output. It wasn't perfect, but it made me feel that someone with more resources like OpenAI would eventually produce models that would output exactly what we see from their new tech.

03:17And my fear was that those in power would use this for one reason: to placate and control the masses, and spread their own beliefs as if they came from the people and not from a position of power. But I also saw it as revolutionary in terms of how we organize society, because I view text in language and speaking as the ultimate representation of internal truth. And if you have models that can read the content of many people and output average representations, then you have this potential to move beyond democracy. See, democracy is about tallying. Before even voting, a smaller set of people have to agree upon what people get to vote on.

04:07And generally a small number of people write up legislation based on their views and people get to decide if they agree with it. Now, seems strange to me that the way we aggregate our opinions is that a small group of people writes out a list of things that were allowed to vote on and then we just count it. And then if that count is manipulated, which isn't that hard if you look at sham democracies around the world, even that isn't representative of the views of the people. I mean, you could send a letter to your representative but they probably won't actually read it. And because your representative needs money to get reelected, it's very unlikely that your singular viewpoint is going to affect them much unless you have a lot of money to donate to them. They're thinking in terms of getting people's votes, not representing the average belief or writing effective legislation.

04:54Because such a small number of people write that legislation, all those in power need to do is hire lobbyists to influence those few people who write what people get to vote on. But with this new technology you could have true democracy without numerical tallying. So to explain how, I want to go into a little detail on how these models work. The model OpenAI uses is a transformer, not a generative adversarial network like the face generator or a variational autoencoder. But they have some things in common, and they're all trying to learn a compressed representation of the data being input that can then make new outputs that to us look real. So I'm gonna talk very generally and try not to get too into the details of any specific model. But if you have a background of machine learning, you'll note that I do tend to lean more towards the variational autoencoder because that's what I'm most familiar with.

05:43You see, if you input human faces into a generative model, it learns a representation where nearby points in that representation output faces that are similar. These models learn a representation of the data such that it can output the most common facial styles by selecting from the center of that distribution, and output less common ones by selecting a point at the edge of that distribution. It's kind of like a hyperdimensional bell curve. And this is an oversimplification of all possible generative models, but this is one architecture that works. In a generative network, the encoder network learns to map the input data, the images of the faces, into a special vector representation. It's a compressed representation of all the input data which makes no sense to us.

06:33But if you take a point of data in that space and feed it to a decoder network, it will construct something that does make sense to us. And in practice this can produce images of faces that were never in the source data, faces that never existed. At the center of this distribution you might find guys that look like this. Anything that's less common would be at the edge. At the edge you might find people with facial deformities or wounds or weird hairstyles. So if you can learn a representation of all human faces and then output faces that appear to be the average human face, why not apply this technology to opinions? Learn a representation of collective human belief.

07:19We're all sharing our opinions constantly, so want to have models almost writing legislation from them in real time. So you have a hub set up where anyone can share any grievance or opinion about the state of society, and that hub is connected to a model that learns a representation where it can output the average views of those who have input their beliefs into it. At the center you might find opinions like dogs are great, sugar tastes good. At the edges maybe something like parasitic worms are great, rotten food tastes good. Now of course I'm oversimplifying this dramatically, but the point still holds. If we organize our society around this, we can move beyond representative government into maybe representational government. We can solve a lot of problems with how our society currently functions.

08:08But that idea needs to spread now through a minefield of opinions being output by BOTS controlled by people who want to maintain their power. And people seem to spread the most basic of opinions right now. If a model was trained, my opinion would probably be on the far edge of that distribution, because basically no one but me has this belief so far, at least as far as I'm aware. And my preferred implementation of this concept is more complex than I'm presenting here, but in the context of what's going viral, I think this is probably the best thought experiment to get your brains firing in the right direction. And I hope that if this concept starts to move closer to the center of the distribution of all opinions, that I can start to share of the subtleties of what I believe would make it most effective. But that's for another video.

08:57So what do you think? Do you think the average viewpoint is really represented by representative government? Do you think I used the word representation too much? Do you think that the average viewpoint could actually be flawed in many ways? Can you imagine ways that you could alter this to make it more functional, to not just evolve into the most base and simple of ideas? And once that viewpoint is aggregated, where would we store it? Maybe some sort of distributed data store. Some sort of chain of information where the sum of daily opinions are encoded into some sort of block. Just food for thought.

09:51[Music] Well I don't need your pride, and I don't need your fear, and I'll find it all out as I spin through this sphere of my… self and another, because we will...