AI is booming, with new models like GPT-4, Gemini, and Llama popping up daily, and applications being deployed across sectors from healthcare to dating. But this rapid growth also brings complexity. What exactly are these models? What do they do? How deep do we need to go? How do we make sense of it all?
The crypto / blockchain industry faced similar complexity, and adopted a “layered” framework to clarify technology and its functions. So why not apply this to AI?
The “3As” Framework for LLMs
Here’s what I’m thinking:
| Layer | Blockchain Usage & Examples | LLM Usage & Examples |
| Layer 1: Archetype | Core blockchain protocol providing infrastructure for transactions and consensus (e.g. Bitcoin, Ethereum, Solana) | Base models trained on massive datasets, providing general language understanding and generation capabilities (e.g. GPT-4, BERT, LLaMA) |
| Layer 2: Adaptation | Solutions built on top of Layer 1 to improve scalability, speed, and cost-efficiency of transactions (e.g. Optimism, Polygon) | Domain-specific models fine-tuned on specialized data to enhance performance in specific areas (e.g. LegalGPT, MedPaLM, FinBERT) |
| Layer 3: Applications | The application specific protocols and dapps built on top of Layer 1 or Layer 2 (e.g. Warpcast/Farcaster, OpenSea/Seaport). | Fine-tuned task-specific models designed for particular use cases (e.g. customer service chatbots) |
Why the “3As” Framework Matters
- Drive Clarity: Provides a clear classification of LLMs, helping everyone understand their roles and capabilities.
- Inform R&D and Decision-Making: Directs research efforts and helps businesses choose the right tool for the job, balancing general-purpose and specialized needs.
- Guide Regulation & Ethics: Clarifies the responsibilities at each layer, promoting responsible AI development and use.
Nuances and Considerations
As with any framework, there are nuances. Here are a few:
- Should there be an L0? We could also call out the core technologies, such as Transformers vs. Stable Diffusion, at layer below. While this increases complexity, it could provide a more complete view of the space.
- Should there be an L4 in crypto? We’re seeing an emergence of application specific protocols (e.g. Seaport for NFTs, Farcaster for Social), so should we separate this layer from the applications themselves?
- Are they really sequential? Layers can overlap — Layer 3 models might build directly on Layer 1.
Despite these nuances, the “3As” framework is a valuable tool for organizing our understanding of LLMs and fostering informed & productive AI conversations.
What do you think? How would you refine it? Let me know!