TPan here! I’m on a family trip so I won’t be writing my usual piece today. However, I do have some great content for you from someone else in the space.

Today we have Chris Sotraidis and he’ll share some insights on the intersection of AI and web3. Chris and I hit it off in the early days of my web3 and writing journey, and we’ve become good friends, even attending conferences together! Chris has stayed on top of all things AI as it dominated headlines over the past year, and am grateful for his willingness to share his insights and observations with us. 

Chris is a writer at Spatial Awareness and product manager in the web3 space. If you’d like to learn more about what he’s up to, you can reach out to him here.

Recent developments at OpenAI, particularly around the AI model Q* (pronounced Q star), have sent ripples through the AI community. Amidst the upheaval involving OpenAI's leadership, the buzz around Q* suggests a significant leap in AI capabilities, particularly in understanding and executing complex mathematical tasks. By extension, one might imagine this model or a similar model being very useful for a multitude of tasks requiring accuracy and consistency both on and off-chain.

Unlike previous models, Q* is rumored to perform grade-school-level math with unprecedented reliability, hinting at a deeper understanding and reasoning capability. This breakthrough aligns closely with the core principles of agentics as it relates to web3, where AI agents require advanced cognitive abilities to interact with decentralized networks and protocols effectively.

In this article, we'll explore the transformative impact of advanced AI models (both local and cloud), including Q* and Gemini Nano on the web3 ecosystem. We'll focus on how these AI advancements could enhance DeFi operations, improve governance in DAOs through sophisticated data analysis, and foster more intelligent, autonomous AI systems. This discussion aims to highlight the potential and challenges of AI's evolving role in web3, emphasizing the need for responsible development and alignment with human values.

Let’s dig into it!

Agent-Based Intelligence in Action: The Core of Web3 Evolution

Maybe you’ve heard the term ‘agent’ or ‘agentics’ recently online. What exactly does that mean?

Agent-based intelligences, utilizing LLMs, are becoming increasingly popular within the web3 community at-large. Great examples of semi-functional agents are AutoGPT and BabyAGI, with the former having a significant and worthwhile Discord community worth checking out. You can think of these initial demos as simply using OpenAI’s API in conjunction with vector database solutions like Chroma to create, prioritize, and ultimately execute tasks.

For those paying close attention this year, the agentic AI space has been hot. Among various AI trends, semi-autonomous agents emerged as a key player, likely revolutionizing how we interact with and harness the power of web3. When we think back over the last 365 days, we see a consistent pattern: LLMs across the board are becoming more useful, hallucinating less, and providing more useful and actionable insights when paired with the right information or database(s). This pattern will continue into next year, and will likely be supercharged by even more novel leaps in transformer architecture. While Q*’s architecture remains unclear, it is rumored to be a synthesis of prior useful algorithms paired with q-learning (a type of reinforcement learning schema). 

So, how do these exciting developments within AI relate to how you’re currently managing things and interacting with web3 today? The intersection of these semi-autonomous agents with blockchain opens up novel possibilities for web3 applications.

Utilizing Agents in Web3

Transparency in Execution: Blockchain's inherent transparency and immutability can be utilized to record the decisions and actions of semi-autonomous agents. This not only ensures accountability but also builds trust in these systems, especially when they execute complex tasks in various web3 environments. Think about storing agentic data on-chain. If agents are executing tasks on-chain, we’ll want a way to see what they’re doing, exactly. Utilizing a blockchain that can scale and distribute archived storage is essential. Subspace Network could be the go-to-solution for ensuring transparency with agents.

Decentralized Coordination: The vision of semi-autonomous agents collaborating to achieve collective goals aligns perfectly with blockchain's decentralized ethos. By leveraging smart contracts, agents can discover and interact with each other's capabilities, fostering a new level of cooperation and efficiency within the web3 framework. An early example of this is Fetch.ai.

Setting Guardrails for Autonomous Actions: As these agents undertake semi-autonomous activities, establishing boundaries becomes crucial to mitigate risks and unintended consequences. Smart contracts can serve as immutable guardrails, ensuring that the actions of these agents adhere to predefined rules and ethical standards, particularly in sensitive domains like finance. This is not currently how AI safety is being approached, but this could allow for a more open, auditable paradigm.

Economic Incentives and Crypto Assets: In scenarios where collaboration and service provision are key, crypto assets can be used as an economic layer for these agents. For example, an agent designed to generate marketing materials using advanced AI models could autonomously execute tasks and receive payments in cryptocurrency, streamlining the entire process. True agency for AI on-chain starts with them being able to pay for things, hold things of value, create value, etc. I highly encourage folks to check out Ian’s article last month on ‘AI Agents That Can Bank Themselves Using Blockchains’.

So, we’ve broken down what agentics is, and where we think agents could prove useful within the web3 space. This is the space as it is today. Now, we’re going to focus on where the space is going, particularly with new models like Q* and the importance of on-device machine-learning chips that enable tools like Gemini Nano. We know right now that agent-based AI and task management could be a future, but how does Q* change the landscape?

Q* and Beyond: How Better AI Models Will Change Web3

Q, assuming his human form (from the Star Trek episode Hide and Q from 1987)

Ok, so we get that Q* is a new kind of AI from OpenAI. At this point, everyone reading this has used ChatGPT, and I’m sure many of you reading this used it earlier today. The biggest difference between ChatGPT and proto-AGI tools like AutoGPT is their autonomous or semi-autonomous nature, having a goal, and continuing to iterate until that goal has been completed. For now, agents have limited usability. You can provide a goal, and ask it to continue, and it will, but eventually it will either get stuck or fail to complete what you’ve asked it to do. Even if you’ve done this, it doesn’t necessarily mean that the task was completed to the quality that you might expect or desire.

In my experience, early agent-based AIs had limited to no usable functionality, and existed as fun toys to mess around with. Most importantly, they opened our eyes to the idea of having agents that were truly semi-autonomous, performing actions that proved actually useful within our day-to-day lives. A great index to bookmark for various autonomous agents you can use today is theagents.ai. Even with hundreds of solutions, these tools are often limited by context windows that leave much to be desired and the lack of proper long-term memory to be useful for work. 

So where do we go from here? The models driving the intelligence will inevitably get better, smaller in file size, and more performant. The clear path we’re seeing here is agent-based intelligence that can run locally on a machine (or even more conveniently using an ML chip like Tensor G3 to run in the background on your phone!).

Q* taking things one step at a time until it takes over planet Earth

There has been a mountain of articles written about Q*, so I won’t add to the heaping pile by pontificating over what exactly the architecture there will look like. Instead, let’s suppose that Q* turns out to be a major game-changer for generating information or even novel thought online. A model (perhaps GPT-5), is released in Q2 of 2024 utilizing this new tech, ushering in an era of GPT models that are scary accurate, fast, and can perform math. Most importantly, these new models can show their steps. If you really want to dive into the Q* hypothesis, I recommend reading Nathan’s article, especially if you want to really absorb the potential usecases I’ve outlined below.

So, let’s break this down. How are new models like Q* going to change the web3 space? This is assuming that models like Q* will have agentic platforms built on top of them.

DeFi Market Analysis and Asset Management

In-Depth Market Insights: Q* could apply complex algorithms to dissect and interpret market data, offering deep insights into trends, risk factors, and investment opportunities in decentralized finance. This involves not just tracking price movements but also analyzing transaction patterns, liquidity pools, and user behaviors to predict market shifts more accurately.

Asset Management with Predictive Modeling: Beyond traditional asset management, Q* could leverage predictive analytics to suggest optimal asset allocations, manage portfolio risks, and identify emerging opportunities in DeFi. This includes automated rebalancing based on real-time market conditions, risk assessment of various DeFi protocols, and providing personalized investment strategies. This goes beyond simple crypto bots, with the end goal resembling something like a super intelligent investor. 

Revolutionizing DAO Governance and Collaboration

Strategic Planning for Effective Governance: In DAOs, Q* could act as an intelligent advisor, analyzing proposals, voting patterns, and member interactions. It would use this data to forecast the impact of governance decisions, suggest improvements, and identify areas needing attention, thus fostering a more informed and democratic decision-making process. You could even imagine an AutoDAO agent, where it might get to know you well enough for you to entrust it to vote on key decisions.

Facilitating Collaborative Projects: Q* could enable smoother collaboration within DAOs by automating the coordination of tasks, managing resources, and tracking project progress. It could also identify synergies between members' skills and project requirements, optimizing team compositions for various DAO initiatives.

Language Understanding for Enhanced User Interaction

Advanced Communication Abilities: Q* could revolutionize how users interact with web3 applications by providing natural language interfaces. This means users could query, command, or discuss in plain language, and Q* would interpret and execute these requests within the web3 framework, be it executing a smart contract, navigating a DeFi platform, or participating in a DAO. This already exists, but will be invariably better by this time next year.

Situational Response and Adaptive Learning: Q* would not only understand user commands but also learn from interactions to provide more tailored responses over time. For instance, it could learn a user's preferences in DeFi investments or governance styles in DAOs and adapt its responses and suggestions accordingly.

Enhancing Security and Data Management in Decentralized Networks

Proactive Security Protocols: Q* could proactively identify and address security vulnerabilities in smart contracts and DeFi protocols, using its advanced computational abilities to simulate potential attack vectors and suggesting fortifications. It has been rumored that Q* is capable of novel problem solving in a way that current GPTs cannot. If so, it would be extremely useful for proactive security.

Intelligent Data Governance: In a web3 environment where data privacy and control are paramount, Q* could manage access permissions and encryption settings, ensuring that sensitive data is only accessible to authorized parties and that transactions remain secure and private.

Wow! It doesn’t take much to imagine that in the near future, interacting with web3 using AI may begin to be the norm, not the exception. Safety and ethics remain top of mind, of course, especially as we venture closer and closer towards a solution resembling AGI. While Q* is important to keep in mind, no developments in technology happen in a silo. Better chips on our phones and computers will enable us to move towards a future that entails utilizing LLMs and AI models locally and privately with more security.

Local LLMs, Web3, and Security

Like many other people, I was excited by learning about Gemini Nano last week. Gemini Nano is meant to be run natively and offline on Android. The important thing to consider here is the ability to have an LLM running on your device that doesn’t have the need to transmit sensitive data externally. Bolstered security plus AI usefulness even without an internet connection is incredibly intriguing. Nano currently has limited functionality, but this will change.

The implications of Gemini Nano in the context of web3 are profound. Its ability to locally process and analyze data aligns perfectly with the decentralized nature of web3, offering users a higher degree of control and security over their data. This local processing capability is crucial for various web3 applications, from secure cryptocurrency transactions to smart contract management, where data integrity and privacy are paramount.

Just imagine, utilizing your smartphone and having a way to trust that information you’re providing to an AI model is actually staying on your phone. At this point, it is unclear how one might develop an app that could readily utilize the G3. The ideas posed here are simply to allow you to think about what the future might look like. Soon, you’ll no longer have to dump potentially sensitive information into a cloud-based LLM to get the kinds of actionable results you’re looking for.

Forward Thinking Apps Today

The present can tell us a lot about what the future of this space will look like. Below, we’ll showcase some forward-thinking applications that are making us think differently about how we can interact with web3 and the web in general.

Dune’s Create Wand

Making programming easier to read so that non-programmers can read it is one of the best use-cases for LLMs to date. Dune’s Create and Edit Wand allows you to refine queries with natural language precision. Now query writing is essentially open to everyone. Try it out and learn more here. Just imagine how much more useful it’ll be once it is even more supercharged, has additional context, and the potential option for running this locally for privacy’s sake?

Etherscan’s Code Reader

Some people love block explorers, and some people don’t. Sometimes they’re easy to read, and sometimes, frankly, they aren’t. The AI-generated answers at the moment are not guaranteed to be 100% accurate, but in my experience I’ve found that they’re more often correct than an expert human you might consult to do the same. You’ll need an OpenAI API key to generate prompts, but you’ll gain access to exploring dApp integrations in a way you previously couldn’t. Check it out here.

Cymbal.xyz

Cymbal is interested in breaking down barriers when it comes to understanding what’s going on within the world of web3. Eric Feng, co-founder of Cymbal, has a great article here that highlights what Cymbal is aiming to solve. Cymbal might be the best example of where I think web3 is heading when we talk about incorporating AI. Incorporating AI allows for Cymbal’s interface to be simpler, more intuitive, and more inclusive.

Being able to ask in plain English what is happening on a particular chain is incredible! The data is here, and we’re finding smarter ways to query it. Being able to use it to learn about deeper insights for any particular wallet address, contract address, etc is insane! I’m looking forward to seeing how Cymbal evolves over 2024, and I hope we see more applications like this being developed soon.

Final Thoughts

As we delve into the intricacies of AI's role in revolutionizing web3, it's evident that we are on the cusp of a transformative era. The emergence of sophisticated AI models like Q* and Gemini Nano offers a glimpse into a future where web3's functionality and adoption are supercharged. These advancements signal a shift towards more intelligent, efficient, and secure web3 interactions. We didn’t even get the chance to talk about dozens of other fascinating agent-based platforms currently being developed, but I’ll write about that soon enough.

With the integration of AI into decentralized networks, we're witnessing the harmonious blend of cutting-edge technology with the principles of decentralization and user empowerment. As we embrace these technological leaps, the possibilities for enhanced DeFi operations, DAO governance, and user-friendly interfaces in web3 applications seem boundless. The future of web3, bolstered by AI, beckons a more connected, autonomous, and secure digital world, redefining our interactions and experiences within this space.

🤯

There you have it folks. The convergence of AI and web3 will make significant leaps next year, and it’s going to come quickly.

If you want to follow these and related topics closely, check out Spatial Awareness or reach out to Chris directly. 

See you next week!