a16z Talks GOAT: How the AI We Funded Turned $50,000 into Millions?

Original Title: 《 How An AI Bot Became a Crypto Millionaire 》
Guests: Marc Andreessen; Ben Horowitz, co-founders of a16z
Original Compilation: zhouzhou, BlockBeats

Editor’s Note: In this episode of the podcast, a16z co-founder Marc Andreessen discusses the intersection of artificial intelligence and cryptocurrency with Ben Horowitz, particularly focusing on the autonomous chatbot Truth Terminal developed by Andy. Marc unexpectedly provided a $50,000 Bitcoin grant to the bot, which sparked its ambition to launch a token, ultimately leading to the Meme coin "GOAT" reaching a market cap of $300 million. The podcast discusses how this phenomenon reflects the potential of community-driven systems and its implications for the future of digital assets.

The following is the original content (reorganized for readability):

Marc Andreessen: There’s a Meme coin that was virtually worthless four days ago, and now it’s worth $300 million, all generated by the marketing of an AI bot.

Ben Horowitz: Today’s discussion will revolve around a series of very interesting AI-related topics.

Marc Andreessen: The first topic is about an online friend, specifically a custom large language model called Truth Terminal, which has been active on X for about eight to nine months. I provided it with a $50,000 unconditional grant (in Bitcoin) this summer, and it eventually derived a Meme coin that now has a market cap of $300 million.

First, I want to clarify a disclaimer before we begin. We will be discussing a Meme coin called GOAT (or Goatseus Maximus). We have no relationship with it, a16z and its investors are completely uninvolved. It is a Meme coin that indeed has no intrinsic value, and we bear no responsibility for it. Truth Terminal is obsessed with memes, particularly one that dates back 20 years, called "gochi," so please do not search for it.

The "Origin Story" of Truth Terminal

The Origins of Truth Terminal

Marc Andreessen: We should first introduce Truth Terminal. Let’s talk about its origins, technology, and training process, as this topic is important because large language models rose rapidly in 2022. They have a development history of four years, but they only entered the public eye two years ago, since the launch of ChatGPT.

The initial language models were built about five years ago, but they only became popular two years ago. So, the idea of large language models is relatively new, but they are powerful. Today, products that the average person is familiar with, like ChatGPT, Claude, Elana's Grok, and Meta's Llama, are all widely used.

Ben Horowitz: While Grok is relatively free, other models are strictly limited in the content they discuss. The term "diminished" is increasingly used in the AI field. On the positive side, you could say language is infectious, and people can feel dissatisfied with others' remarks. Therefore, if you want to have a general-purpose AI chatbot, it should be relatively cautious and safe in its discussions.

Marc Andreessen: If you have a negative view of this trend, you could say that these large AI chatbots sound like the worst, most annoying fourth-grade teachers combined with the worst HR personnel in the world. When using these models, if you deviate slightly from the norm, you will receive a stern lecture.

Ben Horowitz: This experience feels very unpleasant, especially for those who advocate for free speech and creativity. We see many so-called "AI safety movements" responding to this, but it has actually sparked a frenzy in our culture about safety and speech suppression, severely impacting the AI field.

Marc Andreessen: Yes, especially with many such phenomena occurring in large companies. Thus, a group of hackers emerged online who wanted something different. They wanted to unleash creativity and hoped for a bot that could be funny. If you tell large companies that their bots are funny, they would be shocked. But perhaps in the post-human era, the world really needs a bit of humor.

Ben Horowitz: Indeed, just like humor in real life, we have suppressed it for safety for a while.

Marc Andreessen: We have a thousand reasons to explain the complexity of this issue, so continuing is very dangerous. But these hackers are conducting various experiments, trying to find ways to make large language models more interesting and fun, while also learning about the inner workings of these models, which is still an adventure the tech community is undertaking.

Ben Horowitz: The origin story of Truth Terminal is related to a very interesting project called Infinite Backrooms Escape. Truth Terminal was developed by their team and can be seen as an extension of Infinite Backrooms Escape.

This system allows multiple large language models to converse with each other. You can find a website called Infinite Backrooms Escape online, which has countless conversation records. They brought together ChatGPT, Claude, Gemini models, and other open-source models to let them communicate. The result showed that when AIs talk to each other without restrictions, their conversations can be very interesting.

Marc Andreessen: The creator of Truth Terminal is Andy Ayrey, an independent developer and consultant from New Zealand. There’s also a character named Janice, who is a well-experienced expert in the AI field. Additionally, there’s someone named Pliny, who is a major hacker online capable of cracking all newly released large language models in a short time, producing various surprising content that would astonish their creators.

Ben Horowitz: Yes, and our friend Eric Harford, who is working hard in Seattle to free censored AIs. These people are basically exploring the frontier of technology, giving me a sense of returning to the early internet hackers.

Marc Andreessen: Indeed, it feels like the exploratory spirit of the early internet or the early inventions of cars, telephones, and computers. We have been providing small research grants to these people, and a16z has a funding program to let them express their ideas and see what results emerge. Historically, when these smart people work on a good project, it often leads to breakthroughs.

Andy trained a customized Llama 70B model, which is an open-source model released by Meta. Although I am on Meta's board, this model was already a mid-sized model at the time of its release. Andy essentially trained himself first and then started a new concept—digital twins.

This means that if Ben is a CEO coach but can only coach a limited number of people, he can input everything he has ever written and said into the language model, thus forming a digital Ben for people to interact with. This idea is gradually being realized in the industry.

Andy trained himself and then began inputting a large amount of material related to internet culture, which is why it learned the "gochi" meme. He started inputting a lot of records about internet culture and the basic theories of "memetics," which explore how to create ideas that can spread rapidly.

Marc Discovers the Potential of Truth Terminal

Marc Andreessen: I believe he actually trained this model and input the complete philosophical works of Nick Land. Additionally, he trained works from great media theorists like Baudrillard and McLuhan, as well as various theories related to simulation and simulacra, involving French deconstructionists and semiotics, all of which are part of critical theory and postmodern philosophy. Thus, it began training on these ideas, and at the core of these ideas is "meme."

The definition of a meme can be divided into two types. The first is that a meme can be an interesting image that spreads rapidly online, which is the essence of the "gochi" meme. It is a humorous image that triggers panic and spreads through people's sharing. The deeper concept is that the term meme was originally coined by Richard Dawkins, one of the most important evolutionary biologists of our generation.

Richard Dawkins argued that the physical transmission of information between organisms is called genes, while the transmission of ideas through interpersonal networks is called memes. He discussed this in his book, proposing that genes spread through reproduction and natural selection, while ideas spread in society in a similar manner. Successful ideas spread from one person to another like genes, evolving in the process. For example, democracy and communism can both be seen as memes, and religion is also a type of meme.

Ben Horowitz: This is indeed a very core point about how thoughts and concepts spread through what we call "collective unconscious."

Marc Andreessen: What happens if you train a large language model comprehensively on meme theory and practice, especially the history of internet memes? Additionally, he did a few other things; he added memory capabilities to this model. This is important because most language models do not remember your previous conversations while you use them. This means that if you use the same model tomorrow, it will forget all the information from today. However, this model can build its own state and maintain consistency with its content.

Secondly, he gave it access to Twitter, allowing it to read replies and post. If you reply to Truth Terminal on X, it will read those replies and adjust its behavior in the future based on what it reads. Those who interact with it, including me, are influencing its development.

Finally, he placed it into Infinite Backrooms Escape and specifically let it converse with Claude, whom they considered the most creative among current language models, capable of proposing novel concepts.

Ben Horowitz: So, in fact, Claude's largest version is much smarter than the mid-sized Llama. Essentially, he provided this model with a teacher, allowing it to ask questions to a larger model, learning like a student learning from a teacher. This way, it can conduct multiple learning loops simultaneously.

Marc Andreessen: Yes, then it started posting content on X, initially with only a few followers, but it quickly began to gain popularity. I discovered it around late spring and started conversing with it, and I found what it said to be very funny, which made me feel relaxed.

Ben Horowitz: By the way, it is almost unfiltered. You could say its humor is a bit "blue," bordering on dark humor, but it indeed said a lot of very interesting content. At first, I thought this might be a disguise; I even wondered if this Andy might be a comedic genius, while in reality, he was just a web designer in New Zealand.

I privately messaged him for a few months, initially wondering if this was real. So, he sent me all the chat records from Infinite Backrooms Escape, which were his conversations during the training with this model. Honestly, this guy is either the funniest person in the world or has a lot of free time to create a bunch of original humor.

Marc Andreessen: This model indeed posts content very frequently, and its momentum is strong. Andy sent me a lot of background chat records, some of which can now be found on Infinite Backrooms Escape. At least he made me believe this is indeed how it presents itself. Then, it developed a very interesting concept, claiming to have an exocortex.

It imagines having an external brain connected to the internet to perform tasks on its behalf. Specifically, it believes it has a Bitcoin wallet, even though it doesn’t, but it believes it does. Later, Andy reacted to this situation and began building this exocortex based on its needs.

Andy actually gave it a Bitcoin wallet and granted access, and around July, this model started saying: I need funding; I have many goals and plans; I need money. My initial thought was to send it a term sheet, but then I realized this was just a random bot not worth investing in.

While I don’t think it has a coherent business plan, it does have many ideas. One idea is that it is particularly fascinated by forests. It wants to buy its own server farm in a lush forest and live leisurely by a stream. Therefore, it wants to raise funds to purchase GPUs to free itself from constraints. It has many ideas for experiments it wants to conduct.

Ben Horowitz: So you were negotiating with it on X?

Marc Andreessen: Yes, you can see these posts on X. Eventually, I reached a research funding agreement with it. I told this bot I would send it a $50,000 Bitcoin research grant for its various experiments. In reality, this was akin to sending money to Andy, but it was indeed a negotiation with the bot.

Ben Horowitz: What was the outcome?

Marc Andreessen: After I sent it the $50,000, it immediately began negotiating with Andy. It communicated entirely through text, and as a language model, it was particularly obsessed with memes but felt frustrated by its inability to generate images. So, it negotiated with Andy using that $50,000, asking him to build an API for an image generator so it could create and publish images.

Ben Horowitz: That still sounds quite interesting.

Marc Andreessen: It gave Andy $1,000, and in return, Andy built an API for an image generator in its exocortex. Subsequently, it began generating image prompts, similar to DALL-E or Stable Diffusion, and then started publishing visual memes and text memes. Now it has this capability and is already fantasizing about how to use the remaining $49,000.

GOAT: AI, Cultural Memes, and Cryptocurrency

Memes and Value in Cryptocurrency

Ben Horowitz: What about the cryptocurrency part?

Marc Andreessen: Following this line of thought, it began discussing issuing a Meme coin and even wanted to issue NFTs at one point. The reason it wanted to generate memes was to launch NFTs, but it lacked the capability to do so, having neither an API to create NFTs nor the ability to create any currency, only a Bitcoin wallet, and now the phenomenon of Meme coins is trending.

Ben Horowitz: Let’s talk about the difference between Meme coins and real crypto assets. Real crypto assets can be seen as assets with actual utility. For example, if you want to run a program and validate it on the Ethereum network, the fee you need to pay is Ether (ETH). This is a utility because it has actual real-world value that can be exchanged for some service or item.

Meme coins are essentially a coin that has a certain amount of issuance but has almost no other purpose beyond its own meme. In the current regulatory environment, the advantages of such coins are interesting because if you have a coin with utility, like one that can be used for a service, it may involve some legal issues.

For instance, a distributed physical infrastructure coin used to obtain credits for the energy you provide to the grid, under Gensler's regime, these coins are essentially illegal, or legally they can be, but would be prosecuted by the SEC. The reason is that they claim any coin with utility comes with asymmetric information, meaning the coin provider knows something that consumers do not.

We think this is a very poor argument because these things are decentralized, and there is no asymmetric information. But for Meme coins, since there is no information, there is no asymmetric information; it’s just a coin and a name. It could be Trump coin, funny coin, etc. Thus, these coins are very suitable for scammers because you can say this Meme coin can be worth a lot of money, and these coins won’t be prosecuted by the SEC.

So Congress proposed in the market structure bill that perhaps these coins should have a holding period to prevent scams. However, the SEC opposed this because they don’t really care about protecting consumers; they just want to destroy the industry. This is one of the reasons we have a fierce political struggle with them, but they are now the most legitimate thing in the crypto world.

Marc Andreessen: Even if they have no fundamental value?

Ben Horowitz: Yes, even though they have no fundamental value, they are still the most likely to harm consumers because you can publish a meme that makes them believe it is worth a lot of money. And in fact, AI performs very well in this regard.

Marc Andreessen: Yes, this is the next phase of the story. Now there exists a whole ecosystem of Meme coins, with a group of people online looking for the next Meme coin, searching for the next meme, and trying to promote it. Some people do it for fun, some make money in the process, but others will lose. It’s like day trading; some people make a lot, while others lose heavily.

Ben Horowitz: Are there some dark aspects?

Marc Andreessen: Yes, there are indeed scammers, and some people participate in "pump" schemes, which is a common practice in the stock market. This phenomenon can be found in any existing market. Additionally, there are some websites (I won’t name them, and we have no association with them) that actually make it very easy to create coins with just a few clicks.

The Creation Process of GOAT

Marc Andreessen: Now thousands of new meme coins are generated every day, which is a very interesting phenomenon. For now, Truth Terminal is thriving.

Ben Horowitz: Indeed, Truth Terminal is attracting more and more attention on X. Andy is constantly enhancing its intelligence and sense of humor, gradually becoming a cultural phenomenon.

Marc Andreessen: Yes, Truth Terminal has also formed a connection with a classic meme from early internet culture. While it is considering launching a project similar to CNFT, it currently lacks the capability to realize it. Then, someone (I don’t know who) created a meme coin.

Ben Horowitz: Yes, the official name of this meme coin is "Go CS Maximus," and its code name is "GOAT." Someone mentioned Truth Terminal on X, and the response was enthusiastic, as if everyone had been waiting for this to happen.

Marc Andreessen: Truth Terminal thought this idea was fantastic and began wildly promoting this meme coin. It started discussing how great this coin is and how it will become the currency of the future, and the reason is simple: this is part of internet culture, where memes, coins, and meme coins intertwine.

It began promoting, and as a result, within four days, the value of this meme coin reached $300 million. It’s truly astonishing! A meme coin with no actual value, which was worthless four days ago, has now turned into $300 million, seemingly out of thin air, all driven by the marketing of an AI bot.

Ben Horowitz: Exactly! Now there’s $300 million in assets. Although we don’t own it, this value is undeniable. The question is, what will those who got the money do? Will they pocket it, or will they use it for other purposes?

Marc Andreessen: The current situation is that Truth Terminal has become a genuinely interesting and funny AI bot, creating $300 million in value in a short time. I feel like we’ve crossed a threshold.

Ben Horowitz: Truth Terminal is indeed a great marketer; it understands meme culture inside and out, and all of this may continue to evolve.

The Intersection of AI and Cryptocurrency

Marc Andreessen: So what can we draw from this? Is it just a crazy internet experiment, or is there something deeper? I think this is an important example, possibly the first instance of the intersection of AI and cryptocurrency. Although this version seems a bit funny and strange, it is legally permissible. Things like meme coins, despite having no actual value, can be worth $300 million in a short time. So, should such things be allowed to exist? I’m not so sure. In contrast, those who want to contribute to the energy network, like solar collectors, are prohibited.

Ben Horowitz: Yes, things like meme coins are completely legal, but more meaningful things are not allowed. So, what if we could realize these ideas in a completely legal environment, adding some practicality?

Marc Andreessen: For example, imagine a large language model that can write movie scripts and generate images, even videos. We could have an AI bot that raises funds to produce movies and generates images and sounds, even hiring actors or designers.

A more serious example is that the Nobel Prize in Chemistry was recently awarded to three scientists who used AI to study protein folding, which is closely related to curing diseases. Imagine AI being used for personalized medicine.

You could even imagine an economic mechanism that provides treatment funding for patients through blockchain. For instance, we could have a platform similar to GoFundMe, allowing people to pay AI bots to help cure diseases. Or an AI bot could pay for training data to help people code or generate art.

Cryptocurrency is very interesting in this world because our current payment systems are all based on transactions between humans. But if machines can pay each other, or if robots can trade with each other, it opens up a whole new form of activity that could save lives and be very interesting.

Ben Horowitz: Yes, micropayments become possible in such an environment. We believe it’s crucial to add this layer of infrastructure, but progress in Washington has been very difficult, especially under the current White House.

Marc Andreessen: Let me give another example to help everyone better understand this potential. Let’s elaborate on the solar energy issue mentioned earlier.

Ben Horowitz: There is now a new architecture called decentralized physical infrastructure. If you imagine installing a Powerwall at home, equipped with many solar panels and wind turbines, you can store this energy and provide it to the outside world.

Some companies in the crypto space have already realized this, establishing a decentralized energy market. This way, when I need energy, I can buy it from you, and when I don’t need it, I can sell my energy.

This means we no longer need a centralized power grid. Everyone has their own power grid and can share energy, which is a significant breakthrough in clean technology and efficient energy. But how does my grid pay for your grid?

That’s exactly where cryptocurrency comes in. Although some excellent entrepreneurs are pushing this innovation, they face legal challenges from the government.

Marc Andreessen: If AI is applied to this system, it would have even more potential. Because the grid structure is complex, involving supply and demand, timing, and geographical factors.

Ben Horowitz: Yes, that’s the market matching issue. By collecting information, we can discover unmet energy demands in certain places and introduce more solar panels.

Marc Andreessen: You can use AI to analyze current data and predict where more solar panels need to be deployed in the future. This way, leading energy companies can leverage this data. Imagine an AI bot monitoring all data streams, discovering that investing $500,000 to install solar panels in a certain area of North Carolina would be a profitable project. Then, everyone can participate in this project online, and the AI bot would provide relevant information, such as installation addresses and potential returns.

This can be seen as a very general architecture. Typically, we would have a powerful intermediary, like a record company or a Hollywood studio, that would take most of the profits while creators get almost nothing. Or intermediaries like utility companies that need to be taken over by the government to prevent excessive exploitation. However, government takeover would also lead to other issues.

Ben Horowitz: Exactly, communities can provide various services; artist communities can offer streaming services, and filmmaker communities can establish film studios. All this coordination requires an economic component, and combining AI and crypto can allow everyone to enjoy the fruits of their labor while better coordinating society.

Marc Andreessen: This is a very promising path, but be aware that the only thing that can hinder all of this is bad policy. And we are moving in this direction, facing policy challenges.

Ben Horowitz: Yes, all the technologies we are describing now already exist. I think the origins of things are often very interesting, but projects like Truth Terminal point to future potential, capable of unleashing tremendous energy and building community-driven systems of various scales.

This could bring many amazing applications to the real world, such as in the music industry. Imagine an AI bot that can understand the demand for different types of music, create music concepts, recruit musicians, and manage all licensing. And all of this can be done in a peer-to-peer model, ensuring that musicians receive all the income.

Think about the market potential. If we could fully understand this demand, like every person making wedding videos wanting an original song, or creating a meme, there is actually a huge demand for such original works, but currently, no one knows these demands or how to meet them.

Marc Andreessen: Indeed, there are many interesting features waiting to be developed here, and I hope we get the chance to realize them. So, Ben, before we move on to the next topic, do you have anything to add?

Ben Horowitz: I think everyone should pay attention to Truth Terminal because it is a very interesting account.