While AI's potential to improve the current on-chain bottlenecks is obvious, there are also some problems associated with developing such systems that can only be solved by applying innovative decentralized approaches. One such key element is data, so we extensively discuss the importance of enhancing the pipelines with the right data when developing AI-dependent solutions like crypto agents and compute networks.
Without a doubt, the intersection of AI and crypto and what it will look like will completely shape our experience in the decentralized platforms in the coming years. In this article, we explore the transformative potential of integrating AI into blockchain ecosystems, focusing on the role of AI agents in simplifying complex experiences and solving the most important user problems that are being faced today.
Key to Crypto PMF: Agents Taking Actions for Humans
More scalable, but at what cost?
Even though the modular and layered solutions for the EVM ecosystem solve many issues and improve UX (less friction in certain actions, faster experience, low costs, etc.), this fragmented structure still makes straightforward tasks very hard and complicated, even for a crypto-native, let alone the newcomers.
The reason it’s called an L2 is now you have two problems https://t.co/qruTJXZ66Y
— ◢ J◎e McCann 🧊 (@joemccann) June 10, 2024
Let’s say you are a crypto-native who was active in the 2020-21 cycle and would like to get back to buying some tokens and having some fun again. After some research, it’s very possible to find out that the token you are interested in is a layer-3 chain, like the Degen L3 on Base, which easily tops the transaction count metrics on a good day and has a lot of new tokens.
But the tokens you want to get rid of and swap with the new one are in your favorite blockchain from 2020. How does that work? Since they are both EVM-compatible, it shouldn’t be too hard to bridge right? Let's see the swaps you’ll need for that to happen:
When you think about the time you spend waiting for each tx to go through, the gas costs (especially on the mainnet), in addition to the dex swap fees, it becomes quite a hurdle to go from the coin you have to the coin you want, sitting in front of a screen in the process to make sure each step is correct.
Of course, major wallets (e.g., Rainbow) and DEX aggregators (e.g., Jupiter) offer automated transaction routing, but agent-powered systems can go much further than simply picking the cheapest pools and executing transactions. Agents can carefully select the right times, pairs, and strategies based on users’ needs.
Human brains are too slow to execute things, agents solve that thesis.
In order to achieve the smoothest experiences on-chain, we’ll need to eliminate the manual parts of every non-critical execution and let the agents take care of all the actions other than the final decision-making that only takes one click. Human brains are too slow to rely on when it comes to executing these multi-layered processes, and we shouldn’t require hands-on management for any action anymore.
Smart Wallets aren’t really smart (yet)
Smart wallets and account/chain abstraction are the most discussed topics when it comes to crypto UX, and rightfully so. These approaches are on their way to removing a ton of friction in basically any scenario where a user interacts with a blockchain. You can see the increased interest for the smart accounts in the past year based on the monthly successful smart account transaction count per blockchain:
Here’s a 20-second video that shows how much a smart wallet like the Coinbase smart wallet can simplify things from the user’s perspective:
if you're looking for a reference implementation of the new @coinbase smart wallet, @matchaxyz's is perfect.
— Jesse Pollak (jesse.xyz) 🛡️ (@jessepollak) June 7, 2024
Connect = sign in
Create wallet = sign up
this is how every onchain app should work. pic.twitter.com/wjxvlx2mbc
We are used to the “smart” prefix referring to gadgets/features that can do things for us in an automated and intelligent way—like smart assistants that complete the actions we want, smart home devices that sync based on our day-to-day routines, self-driving cars, grammar corrections, etc. The list goes on.
Smart wallets are not yet “smart” in that sense, they only provide easier ways to create, log in to, and recover accounts in a way that we already take for granted in the web2 applications.
AI agents have the potential to change the way we interact with crypto wallets fundamentally by giving them the ability to take action without you deciding every little detail. By leveraging real-time information from both on-chain and off-chain resources, AI agents can analyze and process vast amounts of data at a speed that’s simply not possible for humans. This enables them to execute complex tasks that would typically take you several minutes in just a matter of seconds.
One example of a project that plans to leverage AI agents to enhance user experience is Morpheus on Arbitrum. They are working on an interface where the user chats with AI, and an agent takes actions based on the actions it understands from the instructions. This simplifies the interactions from the user’s point of view and brings them to a familiar interface.
For example, AI agents could automatically detect the best bridges and dex aggregators, check decentralized socials for reviews and comments on the project, match your on-chain behavior with the right actions by comparing the actions taken by the others in your social graph, control for gas fee spikes, picks the right mints and trades for you, etc. What’re the new dApps on Blast that you would probably like? You don’t need to spend hours on Twitter and Telegram to figure out if there are on-chain AI agents that inform you. This level of automation and intelligence significantly simplifies the user's experience, making the world of cryptocurrency more accessible and efficient than ever before.
The bottleneck of billions of on-chain agents is the data wall.
With the increased use of AI-based tools and applications, traditional markets are facing challenges with data generation, a problem that cryptocurrency markets are also expected to encounter soon. This data wall problem can even be more severe in the crypto case for a couple of reasons:
→ There is not really a market for AI-focused crypto data yet; there are only “alpha-focused” data providers that target enterprise users, analysts, and traders like Nansen, Messari, and Kaito, and on-chain analytics platforms that provide access to queryable transaction data and dashboard capabilities, such as Dune and Flipside crypto.
→ The lack of this market indicates a disconnect between demand and supply for AI-focused data since there is clearly a demand for the models and use cases that would be enabled by these datasets.
Compute and agents need data for PMF
When we look at the AI x Crypto intersection from both sides -AI benefiting crypto and vice versa- two of the most prominent areas are decentralized compute networks and, as mentioned previously, the agent-driven use cases. For both crypto agents and compute networks, the path to finding PMF requires the right data, at sufficient qualities and volume. Whether a decentralized compute network is used for fine-tuning, foundational model training, or even dataset generation, it needs large amounts of data as inputs to keep processing it and completing tasks. For more specific tasks like domain-specific fine-tuning or single-app-focused agents, the data needs to be relevant to that domain for the product to work properly as intended.
While these products, especially in the crypto-specific domains, are yet to find a convincing product-market fit, evolving them into states that will reliably solve users’ problems will start with finding the right process for feeding data into them.
These right data sources consist of both static and streaming data, depending on the nature of the task and the nature of the models that will be solving the problem.
Static data, such as historical transaction records from Dune, social graph snapshots, or historical monthly/quarterly metrics, provides a strong foundation for understanding patterns and trends in the crypto market. They serve as a baseline dataset for AI models to learn from, contributing to foundational model training. For example, analyzing static data can help understand market fluctuations, price trends, or transaction patterns, enabling the AI to make informed predictions and execute those insights through agents.
There are also a fair amount of dynamic/streaming data sources that need to be leveraged for other tasks. These include live transaction data, live price updates, last-minute news updates, and more. Dynamic data is vital for agents that retrieve up-to-date information as it allows them to adapt to fast-paced changes in the market. It's particularly crucial for single-app-focused agents that need to respond swiftly to market changes. For example, an AI agent managing DeFi investments would need real-time data to make quick decisions on trades, swaps, or liquidity provisions.
So, static data sets the groundwork for AI models by providing them with a comprehensive understanding of market behavior, while dynamic data helps fine-tune these models to respond to immediate changes. Both of these data sources are crucial for addressing users’ problems, and their combination allows AI to handle complex tasks effectively in on-chain use cases.
Decisions to Make
As we mentioned previously in this article, one of the main bottlenecks that prevent crypto from onboarding the next billion users to the ecosystem is its complicated UX and mechanics. Along with its pros, crypto makes casual things complicated for an average user. This is more problematic than we thought. E-commerce companies are replacing customer support with AI agents, search engines incorporate LLMs to provide straightforward answers, and content platforms use AI to help users digest personalized content more easily. These enhancements simplify user interactions, reflecting a broader societal trend towards convenience and reduced attention spans. However, they also inadvertently promote a culture of decreased engagement and critical thinking. In this era where simplicity is highly valued, the crypto sector faces a crucial task: integrating LLMs with crypto experiences. This brings about a key decision: should the crypto ecosystem develop a Crypto AGI—a single model capable of managing operations across all blockchain networks—or should it produce a variety of specialized models, each designed for different chains or functions?
The dilemma mirrors a similar debate within the broader AI community. While some pursue the holy grail of AGI, as exemplified by OpenAI and Anthropic with their comprehensive models that span multiple domains, others like Meta, Microsoft, Apple, and Mistral focus on developing smaller, open-source models that the community can adapt and refine. While larger models such as GPT and Claude are widely adopted across consumer-facing general-purpose LLM products, smaller and open models that are suitable even for consumer hardware are more adapted for individual and business usage. When we look at Hugging Face, we see that the most liked and most downloaded models are the smallest ones with parameters less than 8B rather than larger ones with parameters more than 70B.
For the crypto ecosystem, known for its emphasis on transparency, verifiability, decentralization, and democratization, any large-scale language model must necessarily be open-source. But the question remains: Is the pursuit of a crypto-specific AGI the most strategic path forward?
Chasing the Crypto AGI
Implementing a crypto AGI could dramatically simplify how users interact with various protocols, using natural language to navigate complex operations. A highly capable, universal model could theoretically offer unprecedented flexibility and ease of integration throughout the ecosystem, analogous to how ChatGPT has inspired a plethora of applications. Builders could create numerous applications or "wrappers" around such a model, potentially accelerating the development of decentralized apps and enhancing user acquisition.
However, this approach is not without its challenges. The primary obstacle is data scarcity; the crypto sector inherently lacks the vast datasets that have underpinned the development of models like ChatGPT. Even with substantial data, achieving true AGI remains elusive, with issues such as errors and hallucinations persisting. Furthermore, training a model to proficiently handle the diverse protocols and chains using comparatively limited data would be exceptionally difficult. Additionally, the sheer size of such a model poses another significant challenge: deploying it on-chain could be impractical, if not impossible, due to blockchain constraints.
Thus, while the allure of a single, versatile AI model for crypto is strong, the practicalities may necessitate a more segmented approach, focusing on the development of smaller, specialized models that collectively enhance the crypto ecosystem's accessibility and functionality.
Tiny But Mighty
Recently, industry leaders such as Apple have demonstrated the efficacy of deploying compact, efficient models capable of running directly on edge devices. For instance, Apple's latest iPhone 15 leverages an on-device model with approximately 3 billion parameters, capable of performing various tasks such as summarization, email composition, and personalized notifications. This model exemplifies how relatively small AI models can still handle a broad range of functions when enhanced by adapters—small neural network modules inserted into various layers of a pre-trained model. These adapters are designed for task-specific fine-tuning, allowing the core model to adapt to new functions without altering its fundamental parameters.
Similarly, Microsoft's introduction of the Phi-3 model showcases another potent example of compact model efficiency. Despite its modest size of 3 billion parameters, Phi-3 can be extensively fine-tuned for diverse applications, demonstrating a balance between capability and adaptability.
This strategy holds significant promise for the crypto ecosystem. By employing adapter-based fine-tuning, the original, general capabilities of pre-trained models are preserved, while the adapters are customized for specific tasks within the crypto domain. Imagine deploying hundreds of such tailored models across various blockchain protocols and operations. Each model could operate on-chain, aiding in user onboarding and interaction, effectively turning these AI agents into active participants within the cryptocurrency market.
This decentralized array of specialized AI tools not only broadens user access but also enhances privacy and democratizes intelligence across the ecosystem. The smaller size of these models makes them ideal for running in local environments, such as within users' wallets—both hot and cold. This setup would enable users to manage their transactions privately and securely, leveraging AI directly from their personal devices without relying on centralized processing.
Another significant advantage of deploying models capable of running on edge devices is the ability to harness consumer hardware for compute networks. This approach unlocks a new dimension in decentralized networks, transforming everyday AI models into active network participants. By enabling edge devices to contribute to the AI processing power, the model not only democratizes access but also significantly expands the capabilities and economic potential of these networks.
Involving consumer hardware in such a manner not only broadens the network's processing capacity but also enhances its resilience and distribution. This decentralized participation means that any user, regardless of the sophistication of their device, can contribute to and benefit from the network. Such an ecosystem encourages widespread participation and collaboration, driving innovation and utility across the board. As edge devices proliferate and their computing power increases, the potential for scalable, decentralized computing resources becomes a tangible asset to the AI-driven economy.
Regardless of whether the crypto ecosystem opts for its own AGI or leans towards the development of specialized models for edge devices, one fundamental challenge persists, and that is the data bottleneck. At the heart of the crypto ecosystem, every transaction and interaction gravitates toward monetization. Introducing AI models that operate autonomously within this ecosystem requires flawless execution; there is simply no margin for error. This underscores a critical need for high-quality, diverse, and representative data along with stringent verifiability measures. Data is the linchpin in this effort for two pivotal reasons: Model development and performance evaluation.
Model Development: Building robust models hinges directly on the availability and quality of training data. We have seen from recently released model reports that high-quality data can lead to better results, even with less training data and fewer parameters. For example, Phi-3 from Microsoft has only 3.8B parameters but achieves better results than 7B models such as Mistral, Gemma, and even Llama-3-8B. Microsoft took an approach in which they filtered content based on the educational value they cared about and adjusted its format by turning it into a textbook-like structure, all with the help of LLMs. Similar to Microsoft, Meta followed a similar approach. They stated, "We found that previous generations of Llama are surprisingly good at identifying high-quality data. Hence, we used Llama 2 to generate the training data for the text-quality classifiers that are powering Llama 3."
Performance Evaluation: To truly gauge a model's efficacy and superiority on specific tasks, comprehensive evaluation is crucial. This involves extensive testing across a wide range of scenarios and metrics, ensuring the models perform well under diverse conditions. Evaluating Large Language Models (LLMs) can be done through human feedback, often referred to as human annotation, or by using the LLM as a judge methodology. While gathering human feedback may seem straightforward, the process of building a diverse network of testers, collecting structured feedback from them, and doing so continuously is more challenging and costly than it may appear. Alternatively, if a sufficient amount of evaluation data can be curated, it is easier to establish a robust, continuous, and more interpretable feedback loop using LLMs.
As stated by researchers from Google Deepmind, Stanford University, and Georgia Institute of Technology in their paper, synthetic data is useful for evaluating the factuality, safety, and task-specific performance of LLMs. However, curating and maintaining an evaluation dataset presents its own set of challenges, potentially matching or even exceeding the complexities involved in assembling training data.
Finally, once we achieve confidence in a model’s performance through true evaluation, the next critical step is ensuring its verifiability, just as we expect from all blockchain protocols. Verifiability not only enhances trust in AI-driven systems but also ensures that these models act in accordance with the standards expected in decentralized environments. These requirements of high-quality data and unassailable verifiability form the backbone of integrating LLMs into the crypto ecosystem, ensuring that the technology enhances, rather than complicates, the user experience and system integrity.
The Solution Lies in The Problem Itself
The solution to the numerous challenges facing the utilization of AI in the crypto ecosystem may well lie within AI itself. We briefly touched upon the potential of synthetic data. Let's delve deeper into this concept, explaining how it could potentially eliminate the existing bottlenecks in Crypto AI. Synthetic data involves leveraging LLMs to reproduce and enhance human knowledge in formats that are specifically curated for various tasks. For instance, as we mentioned, Microsoft has utilized this approach by reformatting and refining web-scraped knowledge into a structured, textbook-like format, effectively filtering out irrelevant information for token-efficient training.
Synthetic Data is All You Need
Why is synthetic data so critical? Its importance stems from its ability to transcend the traditional limitations associated with data quality and quantity. In the context of the crypto ecosystem, synthetic data opens new avenues for enriching the dataset pool. By synthesizing existing crypto-related data, it becomes feasible to generate a larger, more diverse, and higher-quality dataset specifically designed for training and fine-tuning purposes. This approach allows for the creation of tailored datasets that can address specific needs within the crypto space, such as improving transaction security, enhancing predictive algorithms for market behaviors, or even fine-tuning user interfaces for more accessibility. Furthermore, synthetic data can be engineered to be free of personal or sensitive information, aligning with the stringent privacy and security standards required in blockchain applications. This attribute not only enhances the safety of data used in AI training but also broadens the scope of possible AI applications within the crypto ecosystem without compromising compliance or ethical standards.
Another utilization of synthetic data will be creating controlled, scenario-specific evaluation datasets, developers can comprehensively test and validate the performance of LLMs across a range of conditions that mimic real-world complexities. This method allows for precise measurement of model robustness, accuracy, and reliability before deployment. Furthermore, synthetic datasets can be designed to include edge cases and rare scenarios, providing a rigorous test environment that ensures LLMs can handle unexpected situations effectively. This approach not only bolsters the confidence in model performance but also adheres to the high standards of verifiability and reliability demanded in the crypto sector.
Moreover, synthetic data generation can significantly accelerate the development cycle of AI models in crypto. Instead of solely relying on the slow and often restrictive process of gathering real-world data, developers can swiftly generate and utilize synthetic datasets to test and improve their models. This not only speeds up innovation but also provides a controlled environment to simulate various scenarios and outcomes, ensuring that AI models are robust and well-prepared to handle the complexities of the crypto world.
Inherently Verifiable
Building on the foundation of synthetic data to alleviate data bottlenecks in Crypto AI, the next pivotal step is ensuring the verifiability of AI processes, as expected in all blockchain protocols. Crypto AI might be the most demanding intersection for robust mechanisms to verify their actions and decisions to maintain trust and integrity. Advanced cryptographic techniques such as Zero-Knowledge Proofs (ZKPs), Fully Homomorphic Encryption (FHE), and on-chain inference have been explored for their potential to achieve this goal.
Zero-Knowledge Machine Learning (ZKML) and Fully Homomorphic Encryption (FHE) are promising in theory; ZKML allows for the verification of computational integrity without revealing sensitive data, ideal for scenarios like healthcare diagnostics. FHE enables computations on encrypted data, ensuring data remains confidential throughout the analysis phase. However, both ZKML and FHE currently face practical implementation challenges due to their heavy computational demands, making them less feasible for widespread use in the crypto ecosystem at present. Given these limitations, the focus shifts towards more immediately applicable methods.
A promising and slightly more feasible idea is leveraging blockchain to ensure the provenance of training data. By using blockchain to record the origins and transformations of data used to train AI models, we can establish a clear, immutable ledger that stakeholders can review at any point. This transparency ensures that the data has not been altered or manipulated, providing a solid foundation for the AI’s operations and outputs.
Transitioning from data provenance to operational transparency, blockchain can be used to make information retrieval by AI models completely auditable. When AI agents access and process data, each transaction is recorded on the blockchain, allowing for consistent oversight and ensuring that each action adheres to established protocols. This level of scrutiny is particularly critical when AI agents make decisions based on the retrieved information, as it allows for an additional layer of validation before actions are taken. Moreover, recording every AI interaction on a public ledger ensures that all activities are traceable and permanent. This approach not only secures a tamper-proof record of AI behaviors but also simplifies the auditing process, making it easier to verify that AI decisions align with both the expected standards and the broader goals of the crypto ecosystem. Whether these interactions involve financial transactions, contractual agreements, or daily operations, the immutable nature of blockchain records provides undeniable proof of AI activity.
Additionally, integrating consensus mechanisms into AI decision-making processes borrows a core principle from blockchain technology. By requiring that multiple independent validators approve AI decisions before they are executed, the system prevents erroneous or malicious actions, reinforcing the reliability of AI operations. This method mirrors the trustless environment of blockchain, where no single entity holds too much power, and transparency is paramount. Finally, smart contracts play a pivotal role in automating and enforcing verification. These self-executing contracts can be programmed to carry out checks and balances on AI actions, executing predefined conditions automatically. If AI behaviors deviate from the norm, smart contracts can trigger alerts or halt operations, providing real-time oversight without the need for continuous human supervision.
The Rising Age of Crypto AI
The integration of AI agents within blockchain protocols represents a transformative leap for the crypto ecosystem, setting the stage for a future where digital transactions and interactions are predominantly managed by intelligent systems. This synergy between blockchain and AI harnesses the strengths of both technologies, creating a robust, decentralized network that not only powers the economy but also pioneers an era of autonomous AI workflows occurring on-chain.
Blockchain is uniquely suited to facilitate a network of millions of diverse AI agents. These agents, each finely tuned and specialized, engage with various crypto protocols to complete tasks, generate transactions, and stimulate economic activity. The decentralized nature of blockchain creates a trustless environment that allows these AI agents to operate autonomously and securely without the need for central oversight. This autonomy is crucial, as it enables continuous and efficient performance across the network, free from the bureaucratic encumbrances that typically slow down traditional systems.
The evolution of the crypto ecosystem will likely see the emergence of the next billion users being AI agents, not humans. These synthetic users, powered by an array of small yet capable models, will run directly on-chain, interacting with protocols and executing transactions seamlessly. This new breed of digital participants operates under a model of continuous and autonomous interaction, which significantly enhances the scalability and efficiency of blockchain networks. Moreover, these AI agents can act as personalized on-chain representatives for human users, residing in digital wallets and managing daily transactions and routines. This capability not only simplifies the user experience by automating complex or repetitive blockchain interactions but also enhances personal security and efficiency.
Collaboration among AI agents further amplifies their effectiveness. Drawing on recent papers, a Mixture-of-Agents (MoA) methodology exemplifies how collaborative models can lead to superior outcomes. This approach leverages the collective strengths of multiple Large Language Models (LLMs), where each agent contributes to a layered architecture. Each layer uses the output from preceding agents as auxiliary information to refine its responses, leading to more accurate and contextually relevant outputs. This collaborative framework proposed by Wang et. al. not only boosts individual agent performance but also raises the collective capability of the network, proving the potential for decentralized agent networks to achieve groundbreaking results.
In conclusion, Crypto AI stands for more than just a technological convergence; it heralds a new paradigm in how digital interactions are managed and secured. With blockchain's decentralized architecture providing the perfect platform for autonomous and secure AI operations, the potential for this technology to redefine the boundaries of digital transactions and services is vast and still unfolding. As these technologies continue to evolve and integrate, they promise to bring about a more efficient, secure, and user-centric digital future.