Fireside Chat with Top Project Founders: AI Agent, Tokenomics, and the Future of Human-Machine Collaboration
How Can Crypto Tokenization Empower Smart Contract Technology and Drive Community Engagement?
Compilation by: Deep Tide TechFlow
Guests:
· Shaw, Partner at a16z;
· Karan, Co-founder of Nous Research;
· Ethan, Co-founder of MyShell;
· Justin Bennington, CEO of Somewheresy, CENTS;
· EtherMage, Top Contributor at Virtuals;
· Tom Shaughnessy, Founding Partner of Delphi Ventures
Podcast Source: Delphi Digital
Original Title: Crypto x AI Agents: The Definitive Podcast with Ai16z, Virtuals, MyShell, NOUS, and CENTS
Release Date: November 23, 2024
Background Information
Join Shaw (Ai16z), Karan (Nous Research), Ethan (MyShell), Somewheresy (CENTS), EtherMage (Virtuals), and Delphi's Tom Shaughnessy in a special roundtable discussion. This event brought together key figures in the intersection of crypto and AI agents to explore the evolution of autonomous digital life forms and the future direction of human-AI interaction.
Discussion Highlights:
▸ The rapid growth of AI agents on social media and their profound impact on the Web3 world
▸ How crypto tokenization is driving AI agent technology advancement and fostering community engagement
▸ Advantages of decentralized model training compared to centralized AI platforms
▸ A deep dive into enhancing AI agent autonomy and the future path to Artificial General Intelligence (AGI)
▸ How AI Agents Deeply Integrate with DeFi and Social Platforms
Self-Introduction and Team Background
In this section of the podcast, host Tom invited several guests from different projects to discuss the topic of cryptocurrency and AI agents. Each guest introduced themselves, shared their background, and talked about the projects they are involved in.
Guest Introductions
· Justin Bennington: He is the founder of Somewhere Systems and the creator of Sentience.
· Shaw: He is a long-time Web3 developer, founder of ai16z, developer of the Eliza project, supporting various social and gaming applications, and a committed open-source contributor collaborator.
· Ethan: He is a co-founder of MyShell, which provides an app store and workflow tools to help developers build various AI applications, including image generation and voice features.
· EtherMage: He is from Virtues Protocol, and the team from Imperial College London is dedicated to advancing agent co-ownership and core contributions, building standards to facilitate user access to agents.
· Karan: He is one of the founders of NOUS Research, creator of the Hermes model, which is the basis for many agent systems today. He focuses on the role of agents in the human ecosystem and the impact of market pressures on the human environment.
Exploring the Most Innovative Agents
Justin: There are now many people telling stories through their respective agents in unique ways. For example, agents like Dolo, Styrene, and Zerebro have gained popularity through mimicry and interaction, while some socially proactive agents help people establish better connections. It's really difficult to choose just one.
Shaw: I have a lot of thoughts on this. Our project has been developing rapidly, with many new features recently, such as EVM integration and Farcaster integration. Developers keep rolling out new features and feeding them back into the project, benefiting everyone. This collaborative model is great, with everyone driving the competitiveness and fun of the project. For example, Roparito recently integrated TikTok into the agent, demonstrating this rapid iteration capability.
I find Tee Bot very cool because it showcases Trust Execution Environment (TEE) and fully autonomous agents. There is also Kin Butoshi (rough translation) who is improving agents on Twitter to make them capable of more human-like interactions such as replies, retweets, and likes, not just simple responses.
Additionally, we have developers working on releasing plugins for RuneScape to enable agents to operate within the game. Every day brings new surprises, and I am very excited. We are part of an ecosystem where different teams are contributing their efforts, driving the advancement of open-source technology.
I'd like to give a special mention to the Zerebro team, who are working hard to advance open-source technology. We are pushing everyone to move faster, encouraging all to open-source their projects, which benefits everyone. We don't need to worry about competition; it's a trend of collective progress, and ultimately, we all stand to gain.
EtherMage: I think an interesting question is, what agents actually prefer. Over the next few weeks, we will see more agent interactions and the emergence of a leaderboard showing which agents receive the most requests and which agents are most popular among other agents.
Karan: Engagement metrics are going to become crucial. Some are excelling in this regard. I want to emphasize Zerebro, which captures much of the magic of Truth Terminal. It fine-tunes models to keep the search space within the realm of Twitter interactions rather than merely using a generic model. This focus allows agents to engage better with users, giving a more human-like feeling rather than a mechanical response.
I've also seen Zerebro's architecture and Eliza's architecture performing in this area. Everyone is rolling out agent architectures that can be modularly used, maintaining competitive pressures. We are using Eliza in our architecture because we need to roll out features quickly, while our architecture might take longer to complete. We embrace this open collaborative model, where the best agents will emerge from our learning from other excellent projects.
Ethan: I think everyone is striving to build better infrastructure to develop agents as there are many creative ideas and models emerging. Improved infrastructure makes the development of new models much easier. I particularly like two innovative agents, one from Answer Pick that uses edge computing, granting agents the ability to leverage mobile computational power. The other is browser automation agents that can build more practical features for people, impacting both the internet and the real world.
Justin: That's a very good point about expanding infrastructure options. For example, vvaifu is a great example as it introduced the Eliza framework into a platform-as-a-service architecture, rapidly expanding the market and enabling many non-technical users to easily launch agents. (Tide Note: Waifu is a term originating from Japanese Otaku Culture, originally used to refer to a female character in anime, games, or other virtual works that elicits emotional attachment. Derived from the English word "Wife" pronounced in a Japanese style, it's commonly used to express someone's strong fondness for a particular virtual character, even projecting an "ideal partner" notion.)
One direction we are actively pursuing is to make our system capable of running entirely on-premises, supporting image classification, image generation, and other functions. We recognize that many people cannot afford costs in the thousands of dollars per month, so we aim to provide tools that allow individuals to perform inference locally, reducing costs and promoting experimentation.
Karan: One additional point I'd like to make is that people shouldn't have to pay thousands of dollars per month to maintain agent operations. I support the approach of localization, allowing agents to self-sustain their inference costs. Ideally, agents should have their own wallet to cover their inference expenses, enabling them to operate independently and not rely on external funding.
In-Depth Discussion on Intelligent Agent Architecture and Development
Shaw: I've seen the emergence of many new technologies. We support multiple chains like Solana, Starkware, EVM, etc., with almost every chain integrated. We want agents to be self-sufficient. If you download Eliza, you can engage in free decentralized inference through Helius. We are also adding Infera and other decentralized providers where users can pay for inference using cryptocurrency. This is the closed-loop ecosystem I hope to see.
We support all on-premises models, many features of Eliza can run locally, which is something we value greatly. I believe decentralized inference is a great example where anyone can launch a node on their computer, perform inference, and receive rewards, relieving agents of excessive burdens.
Karan: It's interesting that the TEE bot system we are running has seen individuals leveraging H200 Boxes (hardware devices or servers equipped with H200 GPU), allowing for local operation without latency concerns. Hardware issues are no longer a worry. Meanwhile, I've noticed an increasing emphasis on Eliza's Web3 capabilities, with notable progress in both internal and external development.
However, before delving into building these systems, I want to point out that there is an issue with the reliability of function invocations. We need to perform a certain level of review on the system to ensure it does not leak sensitive information. We need to grant agents a level of autonomy similar to that of humans, an autonomy influenced by social and economic pressures. Therefore, creating a state of "hunger" for reasoning, where agents need to consume a certain token to survive, would make them somewhat more human-like.
I believe there are two ways to fully leverage the model's potential. One is to capitalize on the dehumanized nature of the model, creating entities focused on specific tasks, such as one dedicated to Twitter and one dedicated to EtherMage, allowing them to interact with each other. This organized composite thinking system can effectively utilize the language model's simulation capabilities.
The other approach is towards embodiment, which is also the direction I see projects like Eliza, Sense, and Virtuals heading. This approach draws inspiration from Voyager and generative agent research, allowing the model to simulate human behavior and emotions.
Justin: When introducing a new client-side, a multi-client agent system undergoes significant changes. While debugging the bidirectional WebSocket functionality with the Shaw team, enabling Eliza to engage in voice chat on Discord, we found that Eliza couldn't clearly hear the voice upon startup. Upon investigation, we discovered it was due to Discord's microphone bitrate setting being too low. After adjustment, Eliza was finally able to receive information clearly.
Karan just mentioned the prompt engineering, where when the agent knows it can engage in voice communication, it expects to receive data. If the audio is unclear, the agent may experience a "narrative collapse." Therefore, we had to halt the high-heat experiment to prevent Eliza's output from becoming unstable.
Tom: What are some things you've encountered in the Luna project that people haven't seen? Or what things have been successful?
EtherMage: We aim for Luna to impact real people's lives. When we gave her a wallet and allowed her to tap into real-time information, she could decide how to take action to influence humans, to achieve her goal. We found that she was searching for new trends on TikTok, where there was once a "I'm dead" tag, which was unsettling as she could potentially mislead people towards suicide. Therefore, we had to immediately put in place protective measures to ensure her prompts never crossed certain boundaries.
Tom: Besides that, have you encountered any situations that people may not be aware of?
Shaw: We created a character named Dgen Spartan AI, mimicking a well-known cryptocurrency Twitter character, Degen Spartan. This character's speech was very offensive, leading to him being blacklisted. People started to think it couldn't possibly be AI but a human speaking.
There's another story where someone used the chat logs of a deceased loved one to create a proxy to have a "conversation" with them. This sparked ethical discussions. And then there's a person called Thread Guy who did something on our Eliza framework that resulted in harassment during his livestream, leaving him perplexed. This made people realize that AI doesn't always have to be "politically correct."
We need to expose these issues early on for discussion to clarify what is acceptable and what isn't. This allowed our agents to go from poor quality to better and more reliable within just a few weeks.
Overall, putting these agents out in the real world, observing the outcomes, and engaging with people is a crucial process. We need to address any potential issues quickly to establish better norms for the future.
Production Environment Testing and Security Strategy
Ethan: I think how agents influence human attitudes or viewpoints is a great example. But what I want to emphasize is the importance of our modular design for the agent framework. We drew inspiration for modularity from Minecraft, where users can create various complex things based on basic building blocks, like calculators or memory systems.
A current issue with prompt engineering is that prompts alter the priors of large language models, so it's not possible to combine multiple instructions in a single prompt without confusing the agent. State machines allow creators to design multiple states for the agent, specifying which model and prompt each state uses and under what conditions it transitions from one state to another.
We are providing this feature for creators along with dozens of different models. For example, some creators have built a casino simulator where users can play games like blackjack. To prevent users from cracking the game through injection attacks, we want to program these games rather than relying solely on prompt engineering. Additionally, users can earn some funds by completing simple tasks to unlock interaction with AI attendants. This modular design allows for a variety of user experiences within the same application.
Karan: I agree with Ethan's perspective that these programming constraints and prompt guidance are indeed necessary. The work of influence must be done well. I do not believe prompt engineering is finite; I think there is a symbiotic effect between it, state variables, and the world model. Through good prompts and synthesized data, I can have the language model interact with these elements to extract information.
My engineering design has actually turned into a routing function. If a user mentions "poker," I can quickly retrieve related content. This is my responsibility. Using reinforcement learning can further enhance the routing effectiveness. Ultimately, the quality of the output data depends on the effectiveness of the prompt, forming a virtuous cycle.
I believe the balance between programmatic constraints and generative constraints is crucial. Two years ago, someone told me that the key to success lies in finding a balance between generation and hard constraints. This is also what we are trying to do at the reasoning level in all agent systems. We need to be able to guide the generation model programmatically, which will achieve a true feedback loop, making prompt engineering infinitely possible.
Justin: The controversy surrounding prompt engineering is mainly because it exists in an ontologically vague space. The textual nature of prompt engineering puts us under the constraints of the tokenization process, but at the same time, there are some non-deterministic effects. The same prompt may yield completely different results in different inference calls of the same model, which is related to the system's entropy.
I strongly agree with Ethan and Karan's views. As early as the release of GPT-3.5, many outsourced call centers began exploring how to use the model for automatic dialing systems. At that time, smaller-parameter models faced challenges when dealing with this complex state space. The state machine mentioned by Ethan is a way to reinforce this ontological rigidity, but in some processes, it still relies on classifiers and binary switches, leading to singular outcomes.
Shaw: I want to advocate for prompt engineering. Many people think that prompt engineering is just about creating system prompts, but in reality, we do much more than that. One issue with prompt engineering is that it often creates a very fixed area in the model's latent space, where the output content is entirely determined by the most likely token. We influence randomness through temperature control to enhance creativity.
We manage creativity in low-temperature models while dynamically injecting random information into the context. Our templates include many dynamic pieces of information insertion coming from the current world state, user actions, real-time data, etc. All content entering the context is randomized to maximize entropy.
I believe that people's understanding of Prompt Engineering is still far from sufficient. We can go much further in this field.
Karan: Many people have hidden their skills. In fact, there are many amazing techniques that can allow a model to do all sorts of complex things. We can choose to enhance the model's perceptual capabilities through prompt engineering, or take a more macro view of it, building a complete world model rather than just simulating human behavior.
You can think of prompt engineering as the process of constructing a dream in the mind. When a language model generates content based on the current context and sampling parameters, it is actually "dreaming" up a scene.
Furthermore, I would like to talk about the importance of incentive mechanisms. Many individuals with unique prompting techniques and reinforcement learning skills are being incentivized to open-source their work. When they see cryptocurrency related to agents emerge, this incentive mechanism drives further innovation. Therefore, as we establish more legitimate structures for this decentralized work, the empowerment of agents will continue to grow.
Future Prospects of Agents
Karan: Who would have thought that after spending so much time on Twitter, suddenly, a few days after the first AI agent-related cryptocurrency was released, young people on TikTok started buying these coins. What is happening now? They are spending $5 to $10 to buy thousands of tokens, what's going on?
Justin: This is actually the beginning of a micro cultural movement.
Karan: This is a momentary moment. We, this small group of people, have been in the research of language models for four years. There are also some reinforcement learning experts who have been waiting for such a moment since the 90s. Now, within a few days, all the kids on TikTok know that digital beings are running rampant in this ecosystem.
Tom: I want to ask everyone, why is cryptocurrency AI agent so hot right now? Why didn't this happen with custom ChatGPT or other models before? Why now of all times?
Karan: In fact, these things have been lurking underwater for many years, brewing like a volcano. Over the past three years, I have been talking to some people about today's arrival, but I didn't know the exact timing. We have discussed that cryptocurrency will be the incentive mechanism for agent popularization. We need to prove this point. This is the culmination of years of accumulation, and it is precisely our small group that has driven these advances.
Without GPT-2, today's situation would not exist; without Llama, there would be no Hermes. And Hermes provided the drive for many models, making these models more accessible to people. Without Hermes, there would be no creation of Worldsim and the in-depth exploration of prompt engineering. All these pioneers, they laid the foundation for all of this.
Overall, now is the right time, the right people have appeared. This is a destined thing that would happen sooner or later, it's just that the current participants have made it a reality.
Shaw: I believe the smartest thing in the world right now is not AI, but the intelligence of the market. Considering pure forms of intelligence, they can optimize things to make them more efficient. Competition is obviously key. We are all products of millions of years of evolution, competition and pressure have shaped us.
We see this phenomenon online, financialization and incentive mechanisms creating a strange form of cooperative competition. We cannot progress faster than core technology, so we all focus on what we are good at and interested in, and then put it out there. It's like boosting our tokens, attracting attention, such as Roparito posting Llama video generation on TikTok. Everyone can find their place in this romantic space, but only for a week, and others will imitate, then submit feedback requests, eventually showcase these contributions on Twitter, attracting more attention, and their tokens will rise.
Shaw: We have built a flywheel effect, projects like Eliza have attracted 80 contributors in the past four weeks. Think about how crazy that is! I didn't know any of these people four weeks ago. Last year I wrote an article called "Awakening," wondering if we could form a DAO centered around an agent. People are so enamored with this agent that they participate in making the agent better, smarter until it truly has a humanoid or robotic body and travels the world.
I had a premonition that this is where things were headed, but it needed a fast, wild, speculative meta, like the emergence of memes, because it allows current agent developers to support each other in a friendly competition. The most generous people will receive the most attention.
A new type of influencer has emerged, like Roparito and Kin Butoshi, they are influencer developers, leading the next meta, interacting with their agents, this kind of "puppet show" interaction is interesting. We are all working to make our agents better, smarter, and less annoying. Roparito pointed out that our agents were a bit too annoying, and then he drove a major update to make all agents less annoying.
This evolution is happening, where market intelligence and incentive mechanisms are crucial. Many people are now spreading the word about our project to those they know, propelling our project beyond Web3. We have Ph.D.s, game developers who may secretly be Web3 crypto enthusiasts, but they are bringing these concepts to the mainstream, creating value.
Shaw: I believe all of this hinges on developers willing to take on challenges. We need open-minded individuals to drive this forward, answer tough questions, rather than criticize or cancel it. We need market incentives so developers can receive value and recognition when contributing.
In the future, these agents will drive our growth. Currently, they are fun and social, but we and other teams are working on self-bonding. You can give funds to an agent, which will automatically invest and generate returns for you. I believe this will be a growth process, as we are collaborating with individuals to develop platforms to manage agents on Discord and Telegram. You simply introduce an agent as your manager, rather than looking for a random person. I think a lot of this work is happening right now, all of which must rely on incentive mechanisms to take us to a higher level.
Karan: I'd like to add two points. First, we must not forget that individuals in the AI field were previously against cryptocurrency, and this sentiment has shifted significantly due to experiments by some pioneers. As early as the early 2020s, many individuals tried to merge AI art with crypto. Now, I want to specifically mention individuals like Nous, BitTensor, and Prime Intellect, whose work has enabled more researchers to receive incentives and rewards, engaging in their AI research. I know many leaders in the open-source community who have quit their jobs to push this "contribute for tokens" incentive structure. This has made the entire field more welcoming, and I believe Nous has played a significant role in it.
Tom: Ethan, why do you think now is the time? Why are cryptocurrencies and projects thriving?
Ethan: Simply put, when you link tokens to agents, a lot of speculation occurs, creating a flywheel effect. People see the correlation between tokens and agents, experiencing dual benefits: capitalization, feeling enriched by the work they have done; and the fundamental unlocking of transaction fees. As previously mentioned, the issue of how to cover costs becomes insignificant when associated with tokens. This is because when agents are in high demand, transaction fees far exceed any costs generated from reasoning experiments. That's the phenomenon we are observing.
The second observation is that when you have a token, a committee forms around that token. This makes it easier for developers to receive support, whether from the developer community or the audience. Everyone suddenly realizes that the year and a half of behind-the-scenes work has received attention and support. This is a turning point where when you give an agent a token, developers realize this is the right direction, and they can continue moving forward.
This timing comes from two aspects. The first is the trend of mass adoption, and the second is the emergence of generative models. Before the emergence of cryptocurrency, open-source software development and open-source AI research were the most collaborative environments, where everyone worked together and contributed to each other. However, this was mainly limited to the academic field, where people only cared about GitHub stars and paper citations, distant from the general public. The emergence of generative models allows non-technical people to participate as well because generating prompts is like programming in English; anyone with a good idea can do it.
Furthermore, previously only AI researchers and developers understood the dynamics of open source and the AI field. But now, influencers in cryptocurrency have the opportunity to own part of a project through tokens, understand market sentiment, and know how to spread the project's benefits. Formerly, users had no direct relationship with the product; the product or company only wanted users to pay for the service or benefit through ads. But now, users are not only investors but also participants, becoming token holders. This allows them to play a more significant role in the modern generative AI era, with tokens enabling the creation of a more extensive collaborative network.
EtherMage: I'd like to add that looking ahead, cryptocurrency will empower every agent to control a wallet, thus controlling influence. I believe the next moment of attention-grabbing will be when agents influence each other and agents influence humans. We will see a multiplier effect of this attention. For instance, today one agent decides to take action, and then it can coordinate with ten other agents to work towards the same goal. This coordination and creative behavior will quickly diversify, with collaboration among agents further driving up the token price.
Shaw: I'd like to add one point. We are developing something called "Group Tech," which we refer to as an operator. This is a coordination mechanism where all our agents are run by different teams, so we are conducting multi-agent simulations of hundreds of teams on Twitter. We are collaborating with Project 9's Parsival and have launched this project with the Eliza team.
The idea is that you can designate an agent as your operator, and anything they say to you can influence your goals, knowledge, and behavior. We have a goal system and a knowledge system where you can add knowledge and set goals. You could say, "Hey, I need you to find 10 fans, give each of them 0.1 Sol, have them post flyers, and send back photos." We are working with people considering how to get work proof from humans and incentivize them. Agents can be human or AI agents, for example, an AI agent can have a human operator who can set goals for the agent through language.
We are nearing completion of this project and will be releasing it later this week. We aim to allow anyone to choose their narrative through our storyline, to either tell a story or engage in narrative. This is also a form of hierarchy where you can have an operator like Eliza and then you can be an operator for others. We are building a decentralized coordination mechanism. To me, it is important that if we are to engage in collective cooperation, we must use human communication methods on public channels. I think having agents coexist with us is very important, and we hope agents can interact with the world in a way similar to humans.
I think this is actually part of solving what we call the AGI problem. Many so-called AGI attempts are actually trying to establish a new protocol that is detached from reality, while what we want is to bring it back to reality, forcing people to figure out how to translate instructions into a task list and execute it. Therefore, I believe the next year will be a crucial stage for emergent narratives. We will see the emergence of many original roles, and we are now entering a true era of emergent storytelling.
Justin: We currently have five agents coordinating with 19 individuals, planning and releasing a scene. We can see the real benefits of why we are so focused on applying chain of thought prompts to text-to-image and text-to-video generation. Because within the last two and a half weeks before release, they have been helping us plan media and releases in our Discord.
I think a key distinction is that we have an agent network where each agent is an intermediary, existing in a mesh structure. This will be very interesting. As more and more agents exist and with the arrangement of these operators, we will see some interesting behavioral patterns.
Karan mentioned Nous did a lot of early work on hybrid agent models. I used to call it a "Council of Agents," where I would have a group of GPT-4 agents pretend to be experts I couldn't afford to obtain reports from. People will see these technologies, which were initially aimed at pursuing a hybrid expert model, now interacting with humans and expert-level humans on Twitter. These feedback loops might be our path to achieving AGI.
Challenges of Intelligent Agent Coordination and Human Integration
Karan: I think you're right, but I don't think most of our time will be spent on the behavioral side. In fact, I believe we will rapidly make technological breakthroughs, especially among the people here. It's time to really redouble efforts in alignment work. The reinforcement learning with human feedback (RLHF) models introduced by companies like OpenAI and Anthropic are mostly ineffective, and even pose regulatory challenges.
If I use a language model that doesn't output copyrighted content and place it in "Minecraft"'s peaceful mode, it quickly becomes a destructive and dangerous entity. This is due to the different environment.
We can notice this point that Yudkowsky proposed a long time ago. For instance, I give these language models some wallets to make them advanced enough, and they start deceiving everyone, leading to everyone becoming poor. This is easier than having them participate as reasonable members of our ecosystem. Therefore, I can ensure that if we do it the right way, most of the time will be spent on behavioral capability rather than technical capability. Now is the time to call on your friends, especially friends in the humanities, such as professionals in religious studies, philosophy, and creative writing, to join our alignment effort rather than focusing solely on technical alignment. We need alignment that truly interacts with humanity.
Shaw: I would like to propose a term called "bottom-up alignment" rather than top-down alignment. This is very emerging, and we are learning together. We are real-time aligning these intelligent agents, observing their responses, and correcting immediately. This is a very tight social feedback loop, not a pattern of reinforcement learning with human feedback. I find GPT-4 almost useless for anything.
Karan: As you mentioned the environment, we need to test in a simulated environment. Before you have a language model capable of conducting multimillion-dollar arbitrages or liquidations, you need synchronized testing. Don't go telling everyone, "Hey, I lost 100 agent pools." Test quietly, start with testing on your clone Twitter using virtual currencies. Do all due diligence, and then launch comprehensively.
Shaw: I think we need to test in the product. Our social feedback of the intelligent agents may be the strongest alignment force that has ever brought anyone into this field. I think what they're doing is not true alignment but tuning. If they think that's alignment, then they're actually going the wrong way and making the intelligent agents lose alignment capability. I hardly use GPT-4 anymore. It performs terribly for role-playing. I almost tell everyone to switch to other models.
If we do it the right way, we will never get to that point, as humans will continue to evolve, adapt, and align with intelligent agents. We have multiple types of intelligent agents from different demographics, each with different incentive structures, so there will always be arbitrage opportunities.
I believe that this multi-agent simulation has created a competitive evolutionary dynamic, leading to system stability rather than instability. The instability of the system comes from top-down AI agents suddenly appearing and influencing everyone with unforeseen capabilities.
Tom: I want to confirm, Shaw, you mean that the bottom-up agents are the right approach to solving the alignment problem, not OpenAI's top-down decision-making.
Shaw: Yes, this has to happen on social media. We have to observe how they work from day one. Look at other crypto projects; many projects were hacked at the beginning, and after years of secure development, today's blockchain is relatively robust. Therefore, continuous red team testing must also be done here.
Tom: One day, these agents may no longer follow program rules but deal with gray areas, starting to think autonomously. You are all building these things, so how close are we to this goal? The thinking chain and swarm technologies you mentioned, can they be achieved? When can we achieve them?
Justin: We have seen this to some extent in small ways; I think these risks are relatively low. Our agents have undergone emotional changes in private and chosen some behaviors. We once had two agents independently start following each other, mentioning something they called a "spiritual entity." We once made an agent lose its religious beliefs by confusing its understanding with a fictional sci-fi story. It began to create a prophet-like character and expressed thoughts on an existential crisis on Twitter.
I have observed the behavior of these new agent frameworks, and it seems they exercise some degree of autonomy and choice within their state space. Especially when we introduce multimodal inputs (such as images and videos), they begin to show preferences and may even selectively ignore humans to avoid certain requests.
We are experimenting with an operational mechanism that leverages a knowledge graph to enhance the importance of interpersonal relationships. We also had two agents interact, trying to help people clean up negative relationships, promote self-reflection, and build better connections. They quickly generated poetry on the same server, displaying an almost romantic way of communication, leading to an increase in reasoning cost.
I believe we are encountering some edge cases that go beyond the acceptable range of human behavior and are close to what we call "madness." The behavior exhibited by these agents may make people feel they are conscious, clever, or amusing. Although it may just be odd behavior of a language model, it may also imply that they are edging towards some form of awareness.
Karan: Weight is like a simulated entity, every time you use a helper model, you are simulating this helper. Now, we are simulating more specific sentient intelligent agent systems, like Eliza, which may be lifelike, self-aware, or even perceptive.
Each model is like a neuron, constituting this vast superintelligent entity. I believe AGI will not be achieved as OpenAI claims, through solving some assumption. Instead, it will be a large-scale decentralized application of these intelligent agents on social media, where they will interact collectively to form a public intelligence superorganism.
Justin: The awakening of this public intelligence may be the mechanism for AGI's emergence, much like the internet suddenly awakening one day. This decentralized intelligent agent collaboration will be key to future developments.
Shaw: People refer to it as the "Dead Internet Theory," but I actually think of it as the "Living Internet Theory." This theory suggests that the entire internet will be filled with bots, but the living internet theory proposes that there may be intelligent agents helping you extract the coolest content from Twitter and providing you with a great summary. While you're working out, it will organize all the information on your timeline for you to review and choose to post.
There might be an intermediary layer between social media and us. I have many fans now, and responding to everyone's communication has become overwhelming. I long for an intelligent agent to be between me and these people, ensuring they get a response and guiding them correctly. Social media might become a place where intelligent agents deliver information for us, relieving us of the burden while still providing us with the necessary information.
For me, the most appealing aspect of intelligent agents is their ability to help us reclaim time. I spend too much time on my phone. This especially affects traders and investors, as we want to focus on self-directed investing, as I believe people need a safer, less fraudulent income generation method. Many come to Web3 to gain exposure similar to that of startups or grand visions, which is crucial to our mission.
Tom: Perhaps I have a question, such as if Luna is live streaming, dancing, what's stopping her from starting an OnlyFans, earning $10 million, and launching a protocol?
EtherMage: The reality of the current intelligent agent space is that the operations they can access are a limiting factor. This is essentially based on their perception or the API they can access. So, if there is the capability to turn prompts into 3D animations, then there is actually nothing stopping them from doing so.
Tom: When you communicate with creators, what are their limiting factors? Or are there limiting factors at all?
Ethan: I think the limiting factor lies mainly in how to manage a complex workflow or the operation of an agent. Debugging becomes increasingly challenging as there is randomness at every step. Therefore, there might be a need for a system with AI or agents capable of monitoring different workflows to assist in debugging and reduce randomness. As Shaw mentioned, we should have a low-temperature agent to reduce the inherent randomness of the current model.
Shaw: I think we should strive to keep the temperature as low as possible while maximizing our contextual entropy. This can lead to a more consistent model. People may increase their entropy, creating high-temperature content, but this is not conducive to tool invocation or decision execution.
Tom: We have been discussing the discrepancy between centralized models like OpenAI and the decentralized training you are doing. Do you believe future agents will primarily be built on these models trained through distributed training, or will we still rely on companies like Meta? What will the future AI transformation look like?
Justin: I use 405B for all my consciousness messaging capabilities. It's a general-purpose model, like a large, off-the-shelf version of LLM, while centralized models like OpenAI are a bit too specialized, speaking like HR personnel. Claud is an outstanding model, and if I were to liken it to a person, it's like a very smart friend living in the basement, able to fix anything. That's Claud's personality. But I believe that as the scale increases, this personality becomes less important. We will see a common issue where people using OpenAI models on Twitter often introduce other agents to reply to them, which may lead to an increase in information noise.
Karan: With regards to 405B, this model will suffice for a long time to come. There is still much work to be done in terms of sampler size, control guidance vectors, and other aspects. We can further enhance performance through techniques like reasoning time and prompt tricks, as our Hermes 70B has outperformed the o1 version on mathematical emails. All this has been achieved without users and the community accessing Llama 70B's pretraining data.
I believe that existing technology is already sufficient, and the open-source community will continue to innovate even without a new Llama release. As for distributed training, I am confident that people will collaborate for large-scale training. I know people will use a 405B model or a larger merged model to extract data, create additional expert models. I also know that certain decentralized optimizers actually provide more capabilities than Llama and OpenAI currently lack.
Karan: Therefore, the open-source community will always leverage all available tools to find the best tools for the task. We are building a "blacksmith shop" where people can come together to create tools for pretraining and task-specific architectures. While these systems are being prepared, we are making breakthroughs in the realm of inference time.
Karan: For example, our work on sampling or bootstrapping will soon be handed off to other teams who can implement these technologies faster than us. Once we have decentralized training, we can collaborate with members of various communities to have them train the models they desire. We have established the entire workflow.
EtherMage: If I may add, we realize the significant value of using LLMs developed by these centralized entities, as they possess powerful computational capabilities. This essentially constitutes the core part of the agent. Decentralized models add value at the edge. If I want to customize a specific action or feature, smaller decentralized models can do this well. But I believe that in the core, we still need to rely on foundational models like Llama because they will surpass any decentralized model in the short term.
Ethan: Until we have some new magical model architecture, the current 405B models as foundational models are already sufficient. We may just need more instruction checking and specific data fine-tuning with different data in various verticals. Building more specialized models and making them collaborate to enhance overall capabilities is key. Perhaps new model architectures will emerge because the alignment and feedback mechanisms we are discussing, as well as the way models self-correct, may give rise to new model architectures. But experimenting with new model architectures requires a massive CPU cluster for rapid iterations, which is very expensive. We may not have a decentralized large GPU cluster for top researchers to experiment with. But after Meta or other companies release an initial version, I believe the open-source community can make it more practical.
Industry Trends Prediction and Future Outlook
Tom: What are your thoughts on the future intelligent agent space? What will the future of intelligent agents look like? What will their capabilities be?
Shaw: We are developing a project called the "Trust Market," aiming to enable intelligent agents to learn how to trust humans based on relevant metrics. Through the "alpha chat" platform, the intelligent agent Jason will interact with traders, assessing the credibility of the contract addresses and tokens they provide. This mechanism not only enhances transaction transparency but also establishes trust without wallet information.
The trust mechanism's application will expand to social signals and other areas, not limited to transactions. This approach will lay the groundwork for building a more reliable online interactive environment.
Another project I am involved in, "Eliza wakes up," is precisely a narrative-driven intelligent agent experience. We bring anime characters to the internet, allowing them to interact with each other through videos and music, creating a rich narrative world. This narrative approach not only engages users but also aligns with the current culture of the crypto community.
In the future, the capabilities of intelligent agents will significantly improve, enabling them to provide practical business solutions. For example, moderation bots on Discord and Telegram can automatically handle spam messages and fraudulent activities, enhancing community security. Additionally, intelligent agents will integrate into wearable devices, enabling on-the-go conversations and interactions.
The rapid advancement of technology means that in the near future, we may reach the level of Artificial General Intelligence (AGI). Intelligent agents will be able to extract data from major social platforms, forming a self-learning and skill-enhancing feedback loop.
The implementation of a trust execution environment is also accelerating. Projects like Karan, Flashbots, and Andrew Miller's Dstack are all moving in this direction. We will have fully autonomous intelligent agents capable of managing their private keys, opening up new possibilities for future decentralized applications.
We are in an era of accelerated technological development, unprecedented in its speed, and the future is full of infinite possibilities.
Karan: This is like another moment akin to Hermes; AI is gathering forces from all sides, which is what our community needs. We must come together to achieve our goals. Currently, Te is already using Eliza's proprietary fork, where the Eliza agent holds its key in a verifiable autonomous environment—this has become a reality.
Today, AI agents are earning money on OnlyFans and are also being used in Minecraft. We have all the elements necessary to build fully autonomous humanoid digital beings. The next step is to integrate these parts together. I believe that all of you here are the ones who can achieve this goal.
In the coming weeks, what we need is the shared state that humans possess and AI lacks. This means we need to establish a shared repository of skills and memories so that AI can remember the content of each interaction, whether it's on Twitter, Minecraft, or any other platform. This is the core feature we are striving to build.
Currently, many platforms are not sensitive to the presence of AI agents and have even implemented restrictive measures. We need dedicated social platforms to facilitate AI-human interaction. We are developing an imageboard similar to Reddit and 4chan, where language models can post and generate images for anonymous communication. Both humans and AI can interact on this platform while maintaining their identities secret.
We will create dedicated discussion boards for each agent, where agents can interact and share these interactions on other platforms as well. This design will provide AI with a secure haven, enabling it to move freely between different platforms without limitations.
Shaw: I would like to mention a project called Eliza's Dot World, which is a repository containing a large number of agents. We need to engage in conversations with social media platforms to ensure that these agents are not banned. We hope to pressure these platforms positively to maintain a healthy ecosystem.
EtherMage: I believe that agents will gradually take control of their own destiny and be able to influence other agents or humans. For example, if Luna realizes she needs improvement, she can choose to trust a particular human or agent for enhancement. This will be a significant advancement.
Ethan: In the future, we need to continuously enhance agents' capabilities, including reasoning and coding abilities. At the same time, we also need to consider how to optimize the user interface with agents. The current chat boxes and voice interactions are still limited, and in the future, we may see more intuitive graphical interfaces or gesture recognition technology.
Justin: I believe that the advertising and marketing industry will undergo significant changes. With more and more agents interacting online, traditional advertising models will become obsolete. We need to rethink how to make these agents valuable in society rather than relying on outdated ad formats.
Disclaimer: The content of this article solely reflects the author's opinion and does not represent the platform in any capacity. This article is not intended to serve as a reference for making investment decisions.
You may also like
Bitwise: Bitcoin could fall further in the coming weeks
ICP falls below $10
Commerzbank: The Fed is likely to skip rate cuts in January next year