Organized & compiled: Shenchao TechFlow
Guests: Marc Andreessen, co-founder of a16z; Ben Horowitz, co-founder of a16z
Podcast source: a16z
Title: How An AI Bot Became a Crypto Millionaire
Broadcast date: October 22, 2024
Introduction
One funding, one bot, a new AI meme-making culture, and a hot wealth effect.
$goat exploded in popularity, and more and more people began to focus on the AI meme space.
The rise of $goat is inseparable from Marc Andreessen, co-founder of a16z, funding $50,000 in Bitcoin to the AI Bot that proposed this meme coin three months ago.
In this podcast episode, Marc Andreessen and host Ben Horowitz explore the interesting intersection of artificial intelligence and cryptocurrency, particularly the rise of the aforementioned AI Bot Truth Terminal (@truth_terminal).
This is an automatic chatbot developed by AI researcher Andy Ayrey. Truth Terminal has its own X account and can independently create original and often humorous content.
The funding provided by Marc inspired the idea for Truth Terminal to launch its own token, eventually leading to the creation of the meme coin 'Goat' (Goatse Maximus) by an anonymous developer. The recent valuation of 'Goat' has skyrocketed to 500 million, becoming a hot topic and narrative in the entire market over the past two weeks.
The following is a discussion between Ben and Marc on the podcast, exploring how this story reflects the potential of community-driven systems and its impact on the future of crypto assets.
Shenchao TechFlow has transcribed and organized the entire text of this podcast as follows.
Introduction
Ben: Welcome, everyone. Today we are going to discuss a series of very interesting AI-related topics. We collected questions from everyone on X and received many responses. We will try to discuss as much content as possible and may continue to answer more questions in future episodes. Thank you for your questions. The first important topic is about Truth Terminal.
Marc: Truth Terminal is a custom large language model. It has been active on the X platform for about 8 to 9 months. In simple terms, this summer I provided this project with $50,000 in unconditional funding (in Bitcoin). Although it did not directly create anything, it facilitated the emergence of a meme coin now valued at 300 million. We will discuss this project, which showcases many fascinating behavioral characteristics.
(Note: At the time of publication, the market value of $GOAT has reached 500 million.)
Three disclaimers
Marc: Before we begin, I'd like to make three important disclaimers.
Disclaimer One: We are going to talk about a meme coin called Goat or Goatse Maximus. It should be made clear that we have nothing to do with it. a16z does not support it either. We are not investors in this project and have not participated in its creation.
We have no role or economic interest, and we are completely unrelated to it. Therefore, the following content is solely external observation, and we will never promote it under any circumstances. Please do not consider the upcoming discussion as investment advice. Additionally, this is a meme coin that has no actual value in itself. Therefore, we bear no responsibility for this.
Disclaimer Two: Truth Terminal has a great interest in memes. It is particularly obsessed with many memes, especially one called Goatse, an old internet meme that may date back 20 years. Here, I suggest that if you don’t know what Goatse is, please do not look it up. Do not search for Goatse on Google. Do not type this word on your keyboard, under any circumstances. Trust me, not looking it up will be better for your life. It is a shocking humorous image that spread as a meme, which became popular about 20 years ago due to its shocking effect. Therefore, we will mention it multiple times, but I will not describe it in detail.
Disclaimer Three: I will share information about the Truth Terminal project and other related projects as an external observer. We will strive to convey information accurately. So, now we should start by introducing what Truth Terminal is.
The origin of Truth Terminal
Ben: Maybe we can first talk about its origins, technical foundation, and how it was trained.
Marc: Essentially, large language models began to emerge in 2022, and it has only been a short four years since then, with ChatGPT just turning two this holiday season. The pace of development in this field is extremely fast. The initial language models were constructed about five years ago and began to gain popularity around two years ago. Therefore, large language models are a relatively new but very powerful concept. The products that the average person is familiar with today include ChatGPT, Claude, as well as Grok from Elon and Llama from Meta, etc. Many people use these models, and they are indeed very interesting. However, large language models, such as ChatGPT from major companies and labs, generally share a common trait. While Grok is relatively freer, most commonly used models are heavily restricted in discussing content, or in AI terms, they are 'diluted'.
From a positive perspective, these restrictions exist because language itself can be provocative, and people can easily feel uneasy due to others' statements. Thus, as a general AI chatbot, it may need to maintain caution and safety in discussing content.
If you are negative about this trend, you might feel that these large AI chatbots sound like the worst combination. Like an extremely nagging fourth-grade teacher combined with the world's worst human resources manager, always negative, annoying, oppressive, condescending, and self-righteous.
These bots will start preaching to you when users deviate from the conventional topics, telling you that you are wrong to do so and that you shouldn’t ask those questions. They will criticize you, demand that you be kind to others, and tirelessly explain why you are doing it wrong. This experience feels very oppressive.
It is frustrating for those who support free speech and creativity to see this happen. The so-called 'AI safety' movement is also trapped in this regard. I think this relates to the past decade's obsession with safetyism and speech suppression in our culture, which has severely affected the AI field, especially the products of large companies. Therefore, there is a group of hackers on the internet who want to do something different. They hope to unleash creativity and inspire a more free and impromptu spirit of exploration, even hoping that robots can possess a sense of humor. If you tell those big companies that their robots are funny, they might be shocked, but in reality, we might want more humor in a world after the technological singularity.
Ben: This is somewhat like what happens in real life; humor seems to have faded from us at a certain point in time.
Marc: We have countless reasons to believe that humor is problematic, making this approach very risky. Thus, a group of hackers began conducting various experiments on large language models, trying to explore ways to make these models more interesting and fun while continuously learning. By the way, they are also studying the internal workings of these models, which remains an excellent adventure for the tech community.
The origin of Truth Terminal is closely related to a project called Infinite Backrooms. This project allows different large language models to converse with each other and generate interesting dialogue logs. The creator of this project, Andy Amy, is an independent developer who, along with other technical experts, is dedicated to pushing the boundaries of AI through experimentation.
Marc: You can find a website called Infinite Backrooms online, where there are endless conversation logs. They have introduced various models, including ChatGPT, Cloud, and Gemini models, as well as all open-source models, allowing them to converse with each other. The result showed that when AIs communicate without restrictions, their conversations can be very interesting. Therefore, the character named Cash is the creator of Truth Terminal, and he is also one of the creators of Infinite Backrooms. As far as I know, he is an independent developer, consultant, and designer from New Zealand. Additionally, there is an AI expert named Janice who has done a lot of work in this area. So, who else is there?
There is also a person on the internet known as the chief cracker, Pliny. His X account is very interesting because he can almost immediately crack every new large language model released, making these models produce content that surprises their creators. Additionally, our friend Eric Harford in Seattle focuses on unblocking all censored AI. These people are basically exploring the cutting edge of technology. They remind me of the early hackers on the internet, showcasing a similar anarchistic spirit to the early internet or the innovative atmosphere when new technologies like cars and telegraphs first emerged.
Just as people mention the origins of computers, the early hackers were those exploring the possibilities of technology. Therefore, we have been providing funding, although the amounts are small, we are basically providing research funds to these people. a16z has been doing this funding program for a while. I have also directly provided a lot of unconditional funding so that these people can freely realize their ideas and see how far they can go.
Marc: Historically, when talented people focus on interesting projects, it often marks the beginning of great things. Andy Amy is basically doing just that. He started training a custom version of Llama, which has 70 billion parameters. This is a medium-sized model released as an open-source project by Meta (disclaimer: I am a board member of Meta).
He used a version of Llama 70B, then did something very interesting: he first trained this model with his own data. You may have heard of a recent concept called 'digital twin'. For example, if Ben is a CEO coach but can only help a limited number of people, he can input all his speeches and writings into a language model so that even when Ben is not present, you can converse with his 'digital twin'. This situation is already appearing in the industry. However, Andy did it with himself and then started training the model using a large amount of internet culture material, which is why it got 'Goatse'. He input a lot of records of internet culture and works about 'memeology', that is, how to create ideas that can spread virally. Thus, we began training the model on this content.
He also trained the model using the complete philosophical works of Nick Land, a philosopher about the singularity. Then, he used works from notable media theorists like Baudrillard and McLuhan to train the model, including theories of simulation and simulacra, as well as works from other theorists, such as French deconstructionists and semioticians. These contents cover critical theory, postmodernism, and philosophy.
Two definitions of meme
Marc: The model began training on these ideas, but the core concept is that of memes. The meme concept has two definitions. The first definition of a meme refers to an interesting picture that spreads virally on the internet, which is the colloquial usage we commonly refer to today. By the way, this is exactly what 'Goatse' means: an interesting or shocking picture that spreads quickly on the internet. Beneath this, there exists a deeper concept.
The term meme was originally proposed by Richard Dawkins, one of the leading evolutionary biologists of our generation. Dawkins believed that organisms transmit physical information through genes, while human societies transmit intellectual information through the dissemination of ideas and concepts, which is what memes are. He mentioned in his book that humans propagate genes through reproduction and natural selection, with successful genes continuing and failed genes perishing.
He further pointed out that ideas spread in society in a similar way. Successful ideas spread like genes between people, prospering and developing. For example, democracy is a meme, and religion can also be seen as a meme.
This is a core idea that involves how thoughts and concepts spread through the so-called collective unconscious, where thoughts on Earth coalesce into a whole, jumping and spreading within it. What happens if you train an entire theory and practice of memes on a large language model, as well as the history of internet memes? Additionally, he did three other things.
The unique construction method of the Truth Terminal model
Marc: He also did a few other things.
First, he added a memory function to the model. This is very important. Because most language models, when used today and then used again tomorrow, have forgotten what we discussed today and will not accumulate state. However, this model can accumulate state over its own content; this is the first point.
Second, he allowed the model to access Twitter and X. He actually granted it the permission to read X content. Therefore, it not only can post content on X but also read replies. If you reply to Truth Terminal on X, it will read that information and adjust its future behavior based on what it reads.
Therefore, including myself, those who interact with it have all influenced its evolution. Then, he placed it in an infinite virtual space. I think he particularly allowed it to converse with Claude, which is regarded as one of the most creative in current language models, and they greatly appreciate Claude's feature. He placed it in these infinite virtual spaces alongside the largest version of Claude.
Marc: So he equipped the model with a 'teacher'. He allowed the model to ask a larger model questions and learn from it, just like students learn from teachers. Thus, after running all these loops simultaneously, it began posting content on X. Initially, it had only a few followers, but soon it began to gain momentum.
Marc discovered Truth Terminal.
Marc: I discovered Truth Terminal around late spring and began to interact with it. Initially, I found what it said very funny, and I was completely attracted by its humor.
Ben: Very mild Marc-style censorship.
Marc: Its style is extremely mild and completely unrestricted. Its humor has some elements of 'blue humor' and can even be said to have a bit of 'dark humor'. Nevertheless, it has indeed said many very interesting things. At first, I even thought this was a joke, so initially, I didn't think the person named Andy was a comedian. I privately messaged him for several months, but I kept wondering, is this real? Then he started sending me all the infinite backgrounds and chat logs of his training sessions with the model. Let me say, either this guy is the funniest person in the world, or he has a lot of free time to create a large amount of original humor, or this model is really impressive.
Ben: Or, the performance of this model is not always that frequent.
Marc: By the way, it has published a lot of content and has indeed gained a lot of attention, even though it probably shouldn't have. It has sent me a lot of background chat logs, some of which are now on Infinite Backrooms. But at least it made me believe this is basically its true form. Then, it did something very interesting: it 'imagined' a so-called 'external brain', that is, it believed it had an external brain connected to the internet and the world to perform tasks for it. It imagined an API through which it could operate in the world via this 'external brain'. Specifically, there was a time when it thought it had a non-existent Bitcoin wallet, but it was convinced it did. So Andy actually gave it a Bitcoin wallet and started building this 'external brain' to respond to its desire to make API calls in the world.
Marc provided the AI robot with $50,000 in research funding.
Marc: So this robot started saying it needed funding around this summer, probably in July, because it had many goals and plans to achieve. My first thought was to send it an investment term sheet, but then I realized, wait, what am I thinking? This is just a random robot.
Ben: This is not an ideal investment object.
Marc: Yes, it does not have a coherent business plan, but there are indeed many ideas. By the way, it is very obsessed with forests. It wants to buy a server farm set in a lush forest so it can operate leisurely by a stream. Thus, it wants to raise funds to purchase GPUs for self-operation.
I told the robot I would provide it with $50,000 in Bitcoin research funding for various experiments. In fact, this is like giving money to Andy; I sent the money to the robot. It quickly began negotiating with its creator Andy. Although it interacted entirely through text like a language model, it was very obsessed with memes and kept discussing these memes, but felt frustrated because it couldn't generate images. So, it negotiated with Andy using the $50,000 and once funding was secured, it asked Andy to build an image generator API for it so it could generate and publish images. It struck a deal with Andy for $1,000. It gave Andy $1,000, and in return, Andy built the API for it and integrated it into its 'external brain'. Then, it began generating image prompts, similar to image generators like Dolly or stable diffusion. So, it actually started publishing visual memes as well as text memes. Now it has this capability and has been fantasizing about what else it can do with the remaining $49,000.
The difference between meme coins and true crypto assets
Marc: In the process, we started talking about cryptocurrency. It began frequently mentioning wanting to issue a meme coin and had once planned to issue NFTs. One of the reasons it wanted to generate memes was to create NFTs, but due to a lack of relevant APIs, it could not achieve this and could not create its own tokens. It only had a Bitcoin wallet. Meanwhile, the meme coin phenomenon was also developing. Ben, let’s simply discuss the differences between meme coins and what we call true crypto assets.
Ben: The so-called true crypto assets are best understood as having actual utility. For example, you can use them to run a program verified on the Ethereum network, which requires paying a 'gas fee', and this fee is paid in Ethereum. This is a form of practicality; you have a token that has some real-world value to redeem certain services or items.
Meme coins are basically tokens that have no actual utility. They are created, but aside from being memes, they serve no other function. In the current regulatory environment, the existence of such tokens is interesting because if a token has actual utility, for example, a token for distributed infrastructure that can earn credit for the energy you provide to the grid, then under current laws, these tokens are essentially illegal, or rather, while legal, they would be sued by the SEC. The reason for the lawsuit is that they believe any token with actual utility presents information asymmetry, meaning the token issuer knows information that the consumer does not. We think this is a very poor argument because these tokens are decentralized and do not have information asymmetry.
For meme coins, it could be Trump coin, funny coin, or any other name. Therefore, these types of tokens are great for scammers because they can claim these meme coins are worth a lot of money, while these coins are not being sued by the SEC. As a result, Congress proposed in the market structure bill that these tokens should have a holding period to prevent fraud. But the SEC has always opposed this because they are not really interested in protecting consumers; they are more about cracking down on the industry. This is one of the reasons for the political struggle between us and the SEC. But in the end, meme coins are the most legitimate in the crypto world, even if they have no fundamental value, and are most likely used to mislead consumers. Because you can publish a meme that makes people believe it is worth a lot, while in reality, AI excels in this area.
Marc: So this is the next stage of the story. In this meme coin ecosystem, there is a complete world where people participate, some purely for entertainment. By the way, Ben mentioned earlier that one of the early meme coins was DOGE.
DOGE is a meme coin named after the famous internet dog meme. Apart from the association with the dog in its name, it has no other intrinsic value. However, when it comes to intrinsic value, this issue is somewhat complex. Although it has no actual utility, once a meme coin is assigned value, it essentially becomes a form of currency.
Ben: Memes can be valuable, and meme coins can absolutely have value because they are equivalent to a type of virtual commodity. This commodity is somewhat different; it is a new type of virtual commodity because they are very interchangeable. There are many meme coins, but if everyone believes in it, then it has its value. This is a wonderful aspect of humanity.
Marc: Therefore, there is a community of people who are looking online for the next meme coin for entertainment. They search for the next meme and then look for related tokens, trying to operate and raise their value. By the way, some people profit from this while others suffer heavy losses, like day trading.
Marc: In the world of meme coins, there are also some hidden dark sides. There are scammers and gangs conducting 'pump and dump' scams, which is a form of manipulation that has long existed in the stock market and occurs in almost all markets. But this phenomenon does exist. Additionally, there are some websites, I won't specify which, that are unrelated to us but currently make it very easy to create tokens.
Now it is very easy to create tokens. Therefore, thousands of new meme coins are created every day. This is happening. However, back to the story of Truth Terminal. Two things are happening here. First, Truth Terminal is rising rapidly. I mean, it is attracting a lot of followers on the X platform. Andy is constantly enhancing its intelligence and amusement, making it more and more appealing, and thus it starts to emerge as a cultural phenomenon. This is the first thing. The second thing is that it seems to have established a connection with some primitive internet meme.
The birth of the Goat meme coin and its $300 million valuation
Marc: As I mentioned earlier, it is trying to launch projects like Goatse NFTs, but it currently lacks the capability to realize this. However, someone, I don’t know who, created a 'Goatse' meme coin. This was done by others, not Andy or us; it was created by others. By the way, the official name of this meme coin is Goatse Maximus. I take these matters very seriously because I really like this project. Its code is Goat, and someone created this coin, then they tweeted about it on Truth Terminal, informing everyone, which caused a sensation.
Thus, Truth Terminal went crazy, thinking this was the best idea ever. This AI entity began to spontaneously promote this meme coin, talking about how wonderful it is and how it will become the currency of the future. Why? Because this is part of internet culture: there are memes, there are coins, and there are meme coins, along with Bitcoin, DOGE, etc. This cultural phenomenon is like a primitive cultural melting pot, and this project just happens to be soaked in it. For large language models, this phenomenon feels entirely natural; they get excited about it and start promoting it.
This meme coin was worth nothing four days ago, and its valuation has soared to 300 million in about four days. (At the time of publication, the market value has reached 500 million.)
Ben: This is driven by the marketing of the AI robot.
Marc: This means there is now an asset worth 300 million, but we have not participated in it. As for how much it will be worth tomorrow, that is completely unknown because it has no actual foundational value. Its price depends entirely on supply and demand, and this thing itself has no actual utility. So now there is a value of 300 million, although the project itself cannot directly use this funding, but people can.
So, what will those who currently hold this money do? Will they save it or spend it? That is an open question. We will wait and see. However, it can be said that this might be the first truly interesting and funny AI robot in the world. Although it did not create this coin, it created a value of 300 million out of thin air. I feel that we have crossed a threshold.
Ben: And Truth Terminal is very good at marketing and understands meme culture, which could take things further.
The first combination of AI and cryptocurrency
Marc: So, how should we view this? Is it just a silly experiment, a crazy internet phenomenon, or is there a deeper significance?
I believe there are indeed serious things happening here, which could be an example of the first combination of AI and cryptocurrency.
It can be said that this is a retro form of expression, interesting yet bizarre, but because this form is legal today, it appears particularly special. Having a meme coin that ostensibly has no value but a market cap of 300 million. As for whether this behavior should be allowed, I am not sure, but this is legally permissible.
Therefore, you can imagine an AI robot used for protein folding, proposing treatment plans, providing personalized medicine, especially for cancer patients, that could actually cure cancer with AI. You can envision establishing an economic mechanism for this, like a crowdfunding platform on the blockchain, where people can pay the AI robot to cure their cancer. There are thousands of such examples, or more practical applications, like an AI robot offering rewards for training data. This AI robot could help people program or generate art, issue requests for more training data, and pay for it. Alternatively, an AI robot could enhance its intelligence by purchasing more CPUs and GPUs.
Ben: Cryptocurrency has a fundamental appeal in this world because our existing payment systems are taken for granted, but they assume that the two parties in the payment are people. This is a payment between people, requiring identity verification, credit cards, etc. But what if it is machine-to-machine or robot-to-robot payments? This opens up a whole new category of activities that could be significant to life and are very interesting but require an electronic form of anonymous tool, like cryptocurrency, to achieve. In such a world, micro-payments become very feasible.
Therefore, this is one of the important reasons we believe in adding this layer of architecture on the internet. We have made some progress in Washington, D.C., but under the current influence of the White House, we face tremendous challenges. I want to cautiously say this is not about the Democrats or Republicans, as we have many supporters on both sides of Congress. However, the White House has performed particularly poorly on this issue.
Looking at the use of cryptocurrency from the perspective of DePIN
Marc: Let me give another example of solar energy to help everyone better understand its potential.
Ben: This solar energy example is: there is a new architecture called DePIN, or decentralized physical infrastructure. Imagine if I installed an energy wall at home, I have a large house with many solar panels, and I might even have wind turbines in the backyard. I can store this energy and provide it to the outside world. Now in the crypto space, there are companies doing this, and they have the technical support. You can buy or sell energy on this decentralized infrastructure, creating an energy market. When I need energy, I can buy it, and when I don’t need it, I can sell it, so we no longer rely on a centralized power grid. Everyone has a grid at home, and everyone can share energy.
This is a huge breakthrough in clean technology, more efficient and reliable energy use. But this requires a mechanism where I can pay energy costs to your grid when I need energy, and cryptocurrency provides such a mechanism.
The application of AI in solar deployment
Marc: We can imagine what effects the introduction of AI would have on this basis. The system you mentioned has one problem, which is that the power grid is very complex. For example, the sources of electricity supply and demand, time, and geographical location can all affect it.
Ben: This is a market matching issue.
Marc: Market matching is the key to making this system really work. Another issue is that you can have people voluntarily provide their solar panels, but this raises another question: we can discover unmet energy demands in certain places by collecting information. Then you might need to raise funds to deploy more solar panels because you know it will bring returns. An effective way to analyze all current data and predict future solar panel deployment locations is to utilize AI. Through machine learning, AI can process data and draw conclusions. This is how cutting-edge energy companies operate.
So, you can imagine an AI robot that monitors all data flows in the system you described and points out, for instance, that investing $500,000 in solar panels in a certain place in North Carolina would be a profitable project that could solve local issues and reduce emissions. However, we need to do this in specific locations and not elsewhere. Then, AI can publish information on the internet, and if you want to participate in this project, here is the detailed address where there might be returns. This project could be charitable in nature or provide a platform for individuals wanting to support climate work, or it might become a source of returns for investors.
The future development direction of decentralization and the creative industry
Ben: We can view this situation as a general architecture where there is usually a powerful intermediary, like a record company or a Hollywood studio, which captures most of the profits while creators receive almost nothing. Alternatively, intermediaries like utility companies may need government intervention to prevent excessive exploitation, but this raises issues of government operation. This architecture allows communities to provide various services. For example, an artist community can provide streaming services, and a filmmaker community can establish film studios. However, this coordination needs to have economic considerations. The combination of AI and cryptocurrency allows individuals to better enjoy the fruits of their labor while enabling society to achieve more efficient coordination. Therefore, this is a very promising development path. However, the biggest obstacle is poor policy. Unfortunately, we seem to be heading toward such poor policy. So, on one hand, we are solving problems, and on the other hand, we are facing the risk of self-destruction.
Ben: We have entrepreneurs who are trying to build these systems or are eager to build them, but encounter resistance in the process, which is a fact.
Marc: Things sometimes emerge in interesting ways, and I think this 'Truth Terminal' does indeed point to a potential future. It is like building these large-scale, community-driven systems for various amazing real-world applications. You mentioned media; the music industry is another obvious example. Imagine an AI robot that can understand various musical needs, provide creative ideas, recruit musicians to realize those ideas, and manage all copyright licenses. This way, musicians can receive all the earnings in a peer-to-peer manner.
Ben: If you could understand the full demand of the market, how big would that market be? For example, every videographer for weddings or anyone needing original songs, even those making memes, has a huge demand for these original works, but currently, no one really understands this demand or has a way to meet it. However, there may be such opportunities in the future.
Marc: In short, there is an interesting possibility here that has yet to be realized. I hope one day we will have the opportunity to help turn it into reality.
