Introduction
Last updated
Last updated
The emergence of ChatGPT and Large Language Model (LLM) has revolutionized how humans produce and consume knowledge. Within a year, AI-native applications have evolved from chatbots to copilots, to agents.
AI agents would increasingly evolve from supportive tools (akin to Copilots) to autonomous entities capable of completing tasks independently. — Dr. Andrew Ng at Sequoia Capital AI Ascent 2024 Summit
Agents are software applications that can complete tasks on its own autonomously like a human. The agent can understand the task, plan the steps to complete the task, execute all the steps, handle errors and exceptions, and deliver the results. While a powerful LLM could act as the “brain” for the agent, we need to connect to external data sources (eyes and ears), domain-specific knowledge base and prompts (skills), context stores (memory), and external tools (hands). For agent tasks, we often need to customize the LLM itself
to reduce hallucinations in a specific domain,
to generate responses in a specific format (eg a JSON schema),
to answer “politically incorrect” questions (eg to analyze CVE exploits for an agent in the security domain),
and to answer requests in a specific style (eg to mimic a person).
Agents are complex software that require significant amount of engineering and resources. Today, most agents are close-source and hosted on SaaS-based LLMs. Popular examples include GPTs and Microsoft/GitHub copilots on OpenAI LLMs, and Duet on Google’s Gemini LLMs.
However, as we discussed, a key requirement for agents is to customize and adapt its underlying LLM and software stack for domain-specific tasks — an area where centralized SaaS perform very poorly. For example, with ChatGPT, every small task must be handled by a very large model. It is also enormously expensive to finetune or modify any ChatGPT models. The one-size-fits-all LLMs are detrimental to the agent use case in capabilities, alignment, and cost structure. Furthermore, the SaaS hosted LLMs lack privacy controls on how the agent’s private knowledge might be used and shared. Because of these shortcomings, it is difficult for individual knowledge workers to create and monetize agents for his or her own domain and tasks on SaaS platforms like OpenAI, Google, Anthropic, Microsoft and AWS.
In this paper, we propose a decentralized software platform and protocol network for AI agents for everyone. Specifically, our goals are two-folds.
Goal #1: Empower individuals to incorporate his/her private knowledge and expertise into personal LLM agent apps. Those apps aim to perform knowledge tasks and use tools just as the individual would, but also reflect the individual’s style and values.
Goal #2: Enable individuals to provide and scale their LLM agents as services, and get compensated for their expertise and work.
GRT AI is “YouTube for knowledge and skills.”
