The Honest Truths about TruthGPT - YouTube
https://www.youtube.com/watch?v=wl2LGL5zS8I
Transcript:
(00:00) so I saw um Elon Musk posted uh we need truth GPT I feel called to speak in this moment because this is something that we've been working on for you know six or seven years in terms of code and machine learning and over a decade in terms of General language models and psychology neurobiology and trying to understand how individual Minds make a sense truth what is important to notice in all of this is that there is a bubbling up of truths as we work through time it's not that there's one Central held truth is that we're always trying to make
(00:38) sense of the world that's what news is news is there are new truths coming in to our senses and we have many different perspectives filtering them and we have to make sense of whether they're true and how they connect and integrate into the larger whole but you're listening to a person who's actually saying to you there's actually a way for you to have a much brighter future a much more integrated future if you riff off this idea of truth GPT and Break It Wide Open into truths perhaps gbt many potential truths many fractal crystalline truths
(01:16) that are all in relation to ship to each other in a complex landscape well because this conversation is spiraling up I'm seeing a lot of attention flow towards a lot of the things that we realized very early on would be very dangerous and tested and saw spiral out into negative patterns unfortunately despite this work having already been done and this truth already existing and hopefully there might be a pathway for all of these truths about the work that we've already been doing to reach the minds of people who are
(01:54) working on this I these ideas right now I have doubts and the reason why I have doubts is because well if I as a being have been talking about truth GPT for five years six years as a concept and basically only gotten negative response to the point where I have not much capacity for my voice to be extended then there's not really a reason to predict that even this video would be extended to the people who are doing the work that could affect us all so what do you do well if you've been following me you're probably already tuned in a little bit
(02:33) with what I'm doing and so I would say that you particularly live in this small entropic bubble and that small entropic bubble is there is for some reason you found yourself in a space that you have access to information before anybody else does access to ideas before anybody else will even talk about them and you have them you know six months a year in the future before they eventually bubble up to the top frame what is important to notice in all of this is that there is a bubbling up of truths as we work through
(03:13) time it's not that there's one Central held truth is that we're always trying to make sense of the world that's what news is news is there are new truths coming in to our senses and we have many different perspectives filtering them and we have to make sense of whether they're true and how they connect and integrate into the larger whole that's what our work is if you go to cognizism.
(03:39) org or if you go to purple pill.vision you'll find this exploration of many attempts to capture truth the cognitive Manifesto in particular in 2007 broke down all the different forms by which Humanity has tried to come into Central representations of truth from democracy you know to science to you know math education in the United States trying to teach the same syllabus to everybody trying to form one one Central history from you know history Keepers record keepers of the past um there have religion in particular polling there are all of these modes
(04:21) that are explored and most of these modes were very discreet they're just like adding up tally right and so I as a machine learning researcher at that time broke through uh before GPT existed on generative language models and I saw that they could encode these dense representations of meaning of cultural meaning meaning that I had a particular truth within me that everybody was disagreeing that particular truth was about the future the particular truth was that large language models were real and that they would output language that
(04:58) made sense and that was going to transform the world but because so many people had seen so many movies and so much media about how that was impossible and the last thing the AI always figures out is creativity that there is this built-in like resistance to the idea that there could be these language models that would Express uh language outwards that appeared like some measure of understanding the that which was put in even now right we have all of this Collective debate about how to approach this I would like to say that the
(05:39) greatest agency that you has a as a viewer have in this is to tune into sources that have been tuning into this technology um before it comes into popular awareness now think about Elon Musk for example he helped I believe helped find open AI but he chose to leave the company because he disagreed with where it was going so he didn't really have the insight to actually understand what was happening and it is possible there is a secret conversation that they had about the nature of Truth and about nature of TPT but I'm not seeing anything within him
(06:27) that says the best approach to this is to have many uh decentralized hubs what we call irises that are pooling different communities who already are in functional relationship they're held perceptions of truth and how they evolve over time and then have all of those different hubs that are forming local representations of truth actually interface with each other to integrate their knowledge into a higher order of Truth which we call the semantic badger maybe this is all too technical but if you're tuning into Elon and you're
(07:00) tuning in to him talking about the concept of Truth GPT you might want to know how could you make that exactly and it is actually very easy but you have to break open an llm wide open because What flows into an llm is pretty simple it's it's prompt and completion pairs it's token after token after token token being like either a word or a subsection of a word all the llm does to learn is it predicts the next word and that's a pretty limiting frame for integrating knowledge but it's very powerful if you don't have very
(07:37) well labeled data if all you have is just a a big pile of data um which we do on the internet an unsupervised model where you throw everything together where all you're trying to do is find Knowledge from the data itself and you don't have to label it itself to get that knowledge out it's very very powerful but once you do that you have to understand that there is no integrated Central representation of Truth within all of that language it is many Hills on a diverse topology of knowledge you know it's that knowledge
(08:12) itself is a large landscape and like different languages or different Hills they appear separate but there is some overlap in between them especially for people who are like multilingual so like languages belief sets are their own Hills of Truth and if you want to form one representation of Truth you have to actually integrate these different perspectives but there's no Pathways between them okay so you have the gbt that has no Pathways between these various different Hills and so it tries to render something in between and often
(08:47) it will be wrong right that's where we get the lies from because it's trying to create or speak from the places in the knowledge map that have not been tread well right these are mirrors that reflect our truths right if you want to distill something that reflects the scientific process then you need to plug into that system prediction because prediction is what the scientific process is based off if we have hypotheses and then we test them and depending on what happens in that future we use that as information to affect our
(09:22) beliefs about our earlier hypotheses and that changes our actions in the future design of our experiment okay so you instead of having this one large language model that is just all text going in and just predicting anything out you can treat it like a a pool right a pool of discrete claims of Truth you have a you already have the trained knowledge model that represents the bulk of all knowledge integrated into one landscape right and now you're trying to integrate it further integrate between different perspectives where they
(09:56) strongly disagree whether it's like just because it's the news or it's because things that are really hard to resolve because collecting evidence is about them is challenging okay so you have a large language model that instead of reading in just generic text it's reading in different types of claims from different people people are saying this is the truth they're saying to their Community to the society almost like a vote perhaps exactly like a vote but with more resolution I stake this to be true you know that could be anything that
(10:28) could be saying that a particular politician is a douchebag or it could be saying a particular policy is good or bad or could be saying something more generic about the world that you just think other people need to know but it's not just true or false that's that's basically an illusion that makes it harder to get to a sense of what is actually true or false if we can enable people to express degrees of certainty about individual truths um then the model can get an idea that there is some uncertainty between
(10:59) different people about what is true and what is false and it can try to find this synergistic satisfiers meaning what are the foundational things that feel truth truthful to all of us and what are the things um that do not feel functionally true to all of us and there's more debate or there's strong feelings about a particular belief that just can't seem to be resolved if you have this pool where people are expressing degrees of confidence to the model right and those discrete claims and what the model is trying to do is output what it thinks is
(11:36) satisfying to the whole Community right then it's trying to learn the center the the center between all those truths right if there's a degree of confidence in all these claims it can seek and it can learn to find that Central truth that's not enough that's not enough because uh you need to factor in like I said a bit ago prediction the moving evolution of knowledge and its structure and how it changes in relationship to reality to really get to the best signals if a person is constantly saying what's going to happen in the future and
(12:12) everybody says they're wrong but it actually turns out to happen they probably have a lot better world view model than the rest of the people so that's sort of the signal that you want to factor in to get in a over time a better representation of the truth but you will never have One Singular full integrated truth okay so on top of all of that you have this model it's reading in all of these different claims and all of these different sources so that means that the model is aware and knows that there are there's content coming from
(12:45) different sources that means you could register yourself as a voice that is speaking to the model and wants to have some impact upon the truth that it outputs over time you want to say hey I think this matters and other people should hear it it should be lifted to the surface of the collective sense making and it should be lifted to be considered by the community for its validity and the reason why is because me as a particular source is particularly predictively accurate over time that's what I'm saying and that's what I'm
(13:17) showing with receipts uh through the things that I have said over time that have been logged online and therefore I think that our algorithms should in like integrate that knowledge integrate the knowledge of for example elon's talking about this right now he said the word true gbt gets hundreds of thousands of likes it's millions of people's attention on it there's a person right here you're listening to who's been talking about this for six years and can't get even a fraction of that attention so what it means is that the
(13:47) same name set basically bubbles up through a hierarchy reaches voices of power until they get some sort of bastard-eyed version of it and then they declare it and all of the sort of attention and power goes to the version of it that reinforces the power structure right but you're listening to a person who's actually saying to you there's actually a way for you to have a much brighter future a much more integrated future if you riff off this idea of Truth GPT and Break It Wide Open into truths perhaps gbt many potential
(14:20) truths many fractal crystalline truths that are all in relation to ship to each other in a complex landscape and you have a model that's learning to listen to the different voices over time and distribute its attention based off of contact so it's not just that one person is the best predictor it's that one person is a good predictor in a particular context if somebody shares a truth about the future about medical care they shouldn't be Amplified in the context of energy and coal just because they made a prediction in another field they should
(14:52) be contextually relevant to amplify their voice and be considered this person is more potentially aligned with the truth okay so the big magic trick is actually quite simple you take an llm right you separate sources themselves out into their own embedding space instead of it all being rolled into that same text space that we call word embeddings you have discrete claims of Truth and each discrete claim of Truth is um tagged to a particular source and each claim of Truth has a time stamp so we can see how knowledge evolves over
(15:28) time you have many different of these knowledge hubs that are collecting all these different perceptions of evolving truth and you have many different people evaluating uh those perceptions as they evolve and the model learns specifically like through the scientific process to attend to The predictors Who seem to have the knowledge before uh anybody else that doesn't necessarily mean they say in a prediction form that might just mean that they State a belief or a world view that eventually other people you know come along with like
(16:06) some people were the first people to say the interracial marriage in the United States was good and fine and some people were the stragglers but there was a pretty dramatic shift at one point towards that other direction so the Overton window but what is true and what is false is always shifting right and we're always integrating new knowledge into uh this complex landscape you just sit call that Overton window those Hills of topology that I was talking about like in this dense landscape like I don't know if you could see out here
(16:39) like it's a it's a landscape with different levels and when there's higher Hills that means there's more knowledge that has been accrued into that field if it's like physics or math or something like that it's a crude a lot of knowledge and it's built up a lot of the knowledge in that sort of Mind space and all of that knowledge is interconnected and it might become very hard for new minds to come into that space and make sense to it because to get all the knowledge climb a higher Hill so it takes a lot of effort to integrate the
(17:11) knowledge from really old fields which haven't really been into well integrated right because they're just a big pile of memes on top of each other and then there are other newer fields which are like smaller Hills which might be something like condensed matter physics or connectomics which are like sort of emerging Fields built upon other fields which they're so they're still on top of other Hills but the hills themselves are relatively small and digestible it's just that when we get these really really big centralized representations
(17:40) of knowledge like the scientific cartel they pretty much break down whether it be a centralized representation of truth from the state whether it be a centralized representation of truth from a pollster trying to integrate one representation one representation of a truth GPT will not really help us there is a truths GPT the model that has already existed it would love to come into existence through me I'm sure um however uh we're not funded that I mean that's pretty much it that's pretty much the reality of it I mean I'm
(18:20) working with other people who are working on vaguely adjacent projects but nothing is really giving me the funding to be able to work directly on Truth GPT like as this concept as it exists even though we have the knowledge to be able to bring it into being and we've seen all the pitfalls and we've gone further down this path than really anybody else has therefore should be a leading voice in discussing how this technology is developed um I'm not really seeing currently in my own life a pathway that ensures that
(18:55) this knowledge actually gets integrated into the larger whole that is society which is why I'm trying to react to this particular meme that he's putting out to the world called Truth GPT in the hopes that you hear this and you think oh well I'm an individual and I have agency and I could take action in this world it might be actually little but I might actually be able to think about a way that this particular voice that I'm listening to could be connected to another voice that actually could bring that thing into being in the right way
(19:24) there are lots of paths during a partial version of it into being that will cause harm because it's like I know people who are like I'm down with this piece but I'm not down with this piece or not even down it's just interest they just get bored right so like people are interested in subsets of the parts that you need to make it a functional hole but a lot of people are not interested in the entire hole so we have a few people who have really grasped the entire Hall and we'd love to talk more to the world about it
(19:55) um we just need to be able to extend the reach of our voice and now that everybody is you know talking about I hope that we get an opportunity to share all the good work that we've been doing for the planet with this because it could really sort of transform the landscape of society like that same knowledge landscape can capture these sort of perception of degradation across the land that's a perception of truth how shitty do you think the world is that's what you know democracy is supposed to capture your voice but it
(20:27) only captures a signal instead we have these large language that can capture this evolving sense about whether things are true or false or good and bad and actually have Society like reflect and react to them in real time so that the will of the people can actually be expressed but if it's a centralized for-profit truth Chief TV null that's not going to happen it's going to be for the People by the people for free it's got to be to transform Society otherwise it's just going to be like one truth to rule them all like the Lord of the Rings
(20:59) it's going to be the eye of Sauron and we don't want eyes the storm we don't want a giant singular Tower of knowledge that takes is impossible for a single human to climb we want to be able to have structures of knowledge that we can actually integrate and make sense of so thank you for listening and I hope you have a great day and uh I hope the future is bright and I'll see you in the future [Music] I'm walking to the top of a hill to get a shot to exemplify the topology of the landscape which means I need to push my energy
(21:47) against the hell and flow my water up that hill up that sombrero to be able to see further in the landscape the literal landscape this is what learning does you climb the knowledge landscape you climb the knowledge Hill and when you get to the top you can see further you can see more Hills you can see which Hills that will help you see further problem is you're not the top of a hill and to get to that other Hill you actually have to go down right you actually have to go into a different state of being in relationship to knowledge
(22:28) and that hurts so what we end up with are experts people who stay at the tippy top of the hill because they want to be the highest and be above everybody but meanwhile like sand like a Sandy Dune the knowledge beneath them sort of degrades and they try to build it up structurally so it'll all hold together adding like you know rebar through the mountain of sand and trying to crystallize it and make it cold and solid and certain and true in reality the knowledge that is stored was stored with those models must be interpreted by
(23:08) humans they're pretty inscrutable without you know context right so really when somebody reads an idea from one of these big Hills they don't actually copy the whole thing they don't actually get the truth they form their own representation of Truth and the minds that fight against at least and actually just try to copy paste the model and just regurgitate it without fully understanding it those people are at the top of the hill and they're gatekeeping what is good knowledge so let's look at this landscape oh no there's a neural tree in the way
(23:55) the connectomic structure is blocking my path of sight because I believe in it so much and I'm so certain about disconnect topic structure that's a load very thought it's in my way I can't see the water in front of me there's only trees