Re-Thinking AI Agents: A Neurosymbolic Approach

• 0:00-0:36 The speaker is working on restructuring the way he thinks about AI agents and how he teaches about them. This is part of a larger project, which is why he hasn't been posting on YouTube as frequently (almost 2 months).
• 0:00-0:07 Personal Note: He also had his first son recently.
• 0:01:12 React Agents: Foundational, But Limiting:
  • 0:01:32 The presenter mentions he has code examples using React agents (re: reasoning, act).
  • 0:01:41 React is considered the foundational structure for LLM-based agents.
  • 0:02:14 React agents involve multiple reasoning steps and can call tools for external information.
• 0:02:17 React Agent Walkthrough:
  • 0:02:17 Explanation of the React agent through an example.
  • 0:02:23 Input text (query) is fed into the LLM.
  • 0:02:29 The LLM goes through multiple reasoning steps and can call tools.
  • 0:03:33 "Re" and "Act" refers to reasoning and action.
  • 0:03:58 LLM structures its response using the React methodology, which involves tools for search, queries, and answers.
• 0:07:14 Limitations of Current Agents & Importance of Broader Definition:
  • 0:07:14 The speaker states that React or Open AI tools are very limiting for production use cases. In his opinion, the speaker views what an agent is much more than what most people percieve the definition to be.
  • 0:07:40 The presenter reviewed several papers to better understand what a "good definition" of an agent could be.
• 0:07:43 Neurosymbolic Agents:
  • 0:08:37 The presenter highlights that an agent should have a neurosymbolic architecture (mixing neural and symbolic approaches to AI).
  • 0:08:11 A miracle paper describes their system as a neurosymbolic architecture.
  • 0:09:08 The neuros portion are neural networks, and the symbolic portion are the "handwritten stuff" or coded instructions.
• 0:09:25 Symbolic AI (Traditional AI):
  • 0:09:25 Involves written rules, ontologies, and logical functions.
  • 10:16 Example: syllogistic logic from Aristotle (major premise, minor premise, conclusion).
• 0:12:38 Neural AI (Connectionist AI):
  • 0:12:46 Based on neural networks and the mechanisms of the brain.
  • 0:12:53 Key figure: Rosenblatt and the perceptron.
• 0:17:26 Mixing AI Approaches:
  • 0:17:26 Mixing old (traditional AI) with neural AI (Neuron networks/ LLMs).
  • 0:17:45 Neuron networks almost learn symbols but they are just not handwritten.
• 0:18:14 Agent is Neurom Plus Symbolic:
  • 0:19:30 Building agent system by mixing Neuron networks with runnable code.
  • 0:19:52 The speaker prefers the term "Neural + Symbolic" rather than pure neural network because he doesn't want to restrict himself to only saying LLMs
• 0:20:11 Semantic Router (Example Agentic System):
  • 0:21:45 Uses embedding models (neuron network based) to categorize user inputs.
  • 0:21:51 A demonstration uses the router to identify whether a user wants to do a general search to send it to the appropriate pipeline.
  • 0:23:11 An alternative, the speaker suggests modifying the user query with a system note to tell the LLM what function it should use.
• 0:25:21 Agents & Different Paths:
  • 0:25:21 The presenter uses the example of two potential different tools being triggered given an agents prompt.
  • 0:25:52 The presenter highlights that, with such examples, it is easy to see that viewing Agents as only a "LLM plus some tool calls" would greatly constrict its capabilities.
• 0:27:34 Conclusion & Closing:
  • 0:27:34 He hopes that the video helped provide sensible logic behind the thinking that he has.
  • 0:27:51 Standard narrative on what an agent is, is valid but it's not all that an agent is.
  • 0:27:56 He will cover this with more structure in the future with less rambling.