0:00 Introduction: The video outlines the speaker's current work on restructuring the understanding and teaching of AI agents, moving beyond the typical definition.
0:00 Context: The speaker mentions a broader project that has kept them busy, explaining the recent lack of YouTube uploads. They also mention the birth of their first son.
0:01 React Agents: The video briefly covers the React agent, highlighting its foundational role in today's LLM-based agents (0:01:32). It's a review of how agents use reasoning steps and tools to answer questions.
0:02 React Process: React agents use reasoning and action steps (ReAct), leveraging external tools (like search) to gather information and formulate answers (0:02:05).
0:03 Example: The agent uses the "search" tool to determine what program the Apple Remote was designed to interact with and finds it is the "Front Row Media Center."
0:04 Iteration: The agent then reasons about what other device controls Front Row. It determines the "query" to be sent to the Search tool, "front row," and then sends it off.
0:06 Final Answer: After gathering information from the tools, the agent uses an "answer" tool that generates "keyboard function keys," completing the React Cycle.
0:07 Limitations of Current Definitions: The speaker argues that the current commonly accepted definition of agents (LLM + tool use) is limiting and wouldn't be used in production, as the speaker's definition of "agent" is broader than that (0:07:12).
0:07 Neurosymbolic Architecture: The speaker turns to academic papers to define a broader definition of "agent," highlighting a paper that describes agents as a "neurosymbolic architecture" (0:08:11).
0:09 Symbolic AI: Symbolic AI involves using handwritten rules, ontologies, and logical functions to achieve AGI. It uses philosophical grammars such as the Sologistic logic from Aristotle.
0:12 Neural/Connectionist AI: Connectionism focuses on building AI systems loosely based on the mechanisms of our brains and what is now referred to as neuron networks.
0:20 Proposed Definition: A proposed definition of AI agents as "neural plus symbolic" allows for broader applications.
0:21 Semantic Router Example: The speaker uses Semantic Router as an example. Semantic Router uses embedding models (neuron network-based) to classify text and trigger different actions.
0:21 Routing Queries: The embedding model has 2 possible paths, a "political route" and an "ask LLM route." Anything that triggers the "ask LLM route" can then be sent to an LLM or an RAG pipeline.
0:23 Modified Queries: The original query can be modified with "system notes" like "use the rag tool" to increase speed and controllabilty.
0:25 Expanded Agentic System: The speaker proposes that the agentic system includes more than just the LLM, even arguing that a system can be agentic without an LLM.
0:27 Alternative Pathways: Even when there are tool calls in a loop, the agent's path (News or Math websites for example) can vary based on the intent identified.
0:27 Conclusion: The speaker is thinking about and approaching agents in a slightly different way than the "standard narrative," which they hope to cover in more detail and structure soon (0:27:34).