This episode explores the emerging world of user-owned AI, contrasting it with centralized AI models, and discusses how blockchain technology can facilitate a decentralized, agent-based AI ecosystem beneficial for crypto investors.
Structured and Narrative Organization
Market Downturn and AI Advancements
- The podcast begins with a playful banter about the current market state and the rapid advancements in AI. Illia, a returning guest, emphasizes his focus on user ownership, advocating for individuals to have their own AI, rather than relying on a single, profit-driven entity.
- "I'm in a user ownership camp and like user own AI user own internet...I don't want like a single company to control effectively you know our fate."
Agent Interaction and the AITP Protocol
- Illia introduces a new protocol (AITP - Agent Interaction and Transaction Protocol) designed for agent interaction and transactions, likening it to "HTTP for AI." This protocol aims to facilitate communication, negotiation, and transactions between AI agents.
- The mental model presented is one where AI assistants act on behalf of users, interacting with other AI agents representing services, companies, or even autonomous entities.
- Blockchain is highlighted as crucial for discovery and registry in this agent network.
User-Owned AI vs. AI Meme Coins
- The conversation shifts to distinguishing user-owned AI from the current trend of AI meme coins. Illia clarifies that user-owned AI prioritizes user control over data, assets, and the power of choice, contrasting sharply with centralized AI providers that monetize user data.
- User-owned AI is envisioned as a personal assistant that protects and potentially grows user wealth, keeps data confidential, and aids in decision-making.
- "User owned AI is is effec how do we apply these properties to have an AI model that's on your side."
Challenges of Centralized AI and the Need for Verifiable Training
- Illia highlights the risks of centralized AI, including data monetization, potential code manipulation, and the ability to dictate user choices. He emphasizes the need for verifiable training, confidential inference, and a monetization model that rewards creators.
- The concept of "sleeper agents" within AI models is introduced, illustrating how models can be trained to behave in specific ways based on triggers.
- A decentralized, confidential cloud for machine learning is proposed as a solution, enabling verifiable training and encrypted model parameters.
Leapfrogging Centralized AI with User Data
- The discussion explores how user-owned AI can potentially surpass centralized models. Illia suggests that open-source models, run in a confidential end-to-end mode, can provide a competitive advantage.
- Access to comprehensive user data (email, calendar, medical records, etc.) allows for fine-tuning models specific to individual users, something centralized models cannot achieve due to privacy concerns.
- "By getting all this user data it's also negative for them because it's massive uh kind of liability."
The Quest for "Organic AI"
- The concept of "organic AI" is introduced – an AI model where the training data is fully transparent and verifiable. This is presented as the "Holy Grail" in AI research, ensuring the model is free from hidden biases.
- The challenge lies in creating a coalition of academic institutions and researchers to agree on data sourcing and training methodologies.
- It is acknowledged that creating a completely unbiased model is impossible; the goal is to make the biases inspectable.
Fine-Tuning and Bias in AI Models
- The conversation delves into how users can fine-tune models to align with their own beliefs or preferences. The ability to trace the provenance of training data is highlighted as crucial for understanding and potentially correcting biases.
- Examples of biases in existing models (Gemini, Claude, Grok) are discussed, attributed to the data they were fed during training.
- "...it is possible to if you have training data to actually look back and see which kind of training examples have activated it."
AI Interoperability and Agent Learning
- The hosts discuss the potential for AI agents to learn and fine-tune one another. Illia emphasizes that user-owned AI allows for personalized models, reflecting individual preferences.
- The scenario of AI agents born from smart contracts is explored, where token holders can govern and influence the AI's training.
The Future of AI Meme Coins and Useful Agents
- The discussion touches on the past and potential future of AI meme coins. Illia suggests that the next wave will focus on "useful agents" that can perform tasks on-chain.
- "...the next obvious thing is making agents actually useful and actually doing stuff."
Current Use Cases and Agent Building
- Illia describes various use cases being built, including e-commerce, travel planning, and even a no-code agent builder. These agents utilize APIs and virtual browsers to interact with existing services.
- A "boss agent" concept is introduced, where a single assistant agent coordinates multiple specialized agents.
- These agents are being integrated into platforms like websites and wallets (e.g., Sweat Wallet).
Use Cases for Business Owners
- The potential for AI agents to assist business owners is discussed. Illia expresses excitement about a project manager agent that can track projects across various platforms (Slack, Notion, etc.).
- The platform is designed as a developer platform, allowing anyone to build and host agents.
Monetization Strategies for Agent Creators
- Illia outlines various monetization strategies for agent creators, including:
- Researchers/developers can charge per million tokens when their fine-tuned models are used.
- Agent builders can monetize based on usage (e.g., per token, per session) or restrict usage.
- Commercial agents can charge a percentage on transactions.
- "...payments to become effectively the core rail for monetization..."
Near Protocol and Sharding
- The conversation concludes with a brief mention of Near Protocol's sharding capabilities, highlighting its potential for scaling to accommodate a large number of AI agents.
- Eight shards are already live on the testnet, with plans to launch them on the mainnet soon.
Actionable Insights and Strategic Implications
- Focus on "Useful Agents": Crypto AI investors should shift their attention from meme coins to projects developing agents that perform real-world tasks on-chain.
- Verifiable Training is Key: Researchers should prioritize developing methods for verifiable training and transparent data sourcing to build trust in AI models.
- Agent Interoperability: The development of protocols like AITP is crucial for creating a functional AI ecosystem, presenting opportunities for infrastructure investment.
- Monetization Models: Explore the various monetization strategies for agent creators, particularly those leveraging on-chain payments.
- Scalability Solutions: Pay attention to blockchain platforms with robust scaling solutions (like sharding) to support the growth of the AI agent ecosystem.
Speaker Attribution and Analysis
- Illia: As a key figure in the Near Protocol ecosystem, Illia provides an insider's perspective on the development of user-owned AI and the role of blockchain technology. His tone is optimistic and forward-looking, emphasizing the potential for a decentralized AI future.
- Host: The host guides the conversation, probing Illia for details and clarifying complex concepts for the audience.
Technical Terms and Contextual Enrichment
- zkML (Zero-Knowledge Machine Learning): Not directly mentioned, but relevant to the discussion of confidential inference. zkML enables verifying the output of an AI model without revealing the input data or the model itself.
- Sharding: A database scaling technique that partitions data across multiple machines, enabling higher transaction throughput.
- AITP (Agent Interaction and Transaction Protocol): A new protocol for agent interaction and transactions, similar to HTTP for AI.
- Sleeper Agents: Models can be trained to behave in specific ways based on triggers.
Reflective and Strategic Conclusion
The discussion underscores AI's transformative potential within crypto, particularly through user-owned, interoperable agents. Investors and researchers should prioritize projects enabling verifiable training, practical agent functionalities, and scalable blockchain infrastructure to capitalize on this emerging trend, focusing on real-world utility over speculative meme coins.
