NEW CORE of AI Agents (MIT, Stanford)

---

---

AI Summary

Summary of AI Agent Core Discussion

Technical Abstracts Overview

• High dimension execution skill for dynamic agents in continuous domains.
• AI-driven review system by Stanford, MIT, Columbia, and Boston Universities.
• Strategic skill learning by LLMS via AB level tree search by Caltech, NC Labs, UC, Shenen University, and RPI.

Previous Work and Videos

• Discussed multi-agent devices for adaptive cyber defense, regret gap, AI game theory, reinforcement learning, and synthetic science with EII.
• Explored coding multi-agent AI systems and self-improvement, self-reflection, and self-learning of agents.
• Touched on distributed AI computing and EI agent design.
• Mentioned the potential of AI to make coders obsolete according to the AWS Chief.

AI Code Editors

• Briefly tested a new AI code editor, cursor.com, and compared it to Microsoft Copilot Plus+.

Methodology for Analyzing Papers

• Vector space model to calculate similarity and retrieve info segments.
• Recognized the limitations of RAG (Retrieval Augmented Generation) in capturing the core complexity of papers.
• Discussed the importance of long-context LLMs for better performance.

AI-Driven Review System

• Prototype system that simulates human academic reviews.
• The system provides structured reviews with ratings and confidence levels.
• Different models may focus on different aspects, highlighting strengths and weaknesses.

Agent Skill Estimation and Strategic Skill Learning

• Importance of estimating agent skill levels and upskilling agents.
• Discussed the integration of game theory and other mathematical approaches for strategic skill development in multi-agent environments.

Methodology for Agent Skill Improvement

• Monte Carlo skill estimation using particle filtering in dynamic networks.
• Tree search methodology combined with an LLM for self-improvement.
• Probabilistic models used for dynamic adaptation.

Challenges and Opportunities

• RAG’s failure to establish shared themes in complex papers.
• The need to upscale agent performance levels for better results.
• The shift from code complexity to design complexity in EI agents.

Essential Building Blocks of an Agent

• Functional capabilities: perception, tool use, knowledge representation, learning, planning, action execution, goal management, and communication.
• Structural capabilities: core AI model that powers various functions.

Joint Estimation of Execution and Decision-Making Skills (JEDS)

• Method to estimate both execution and decision-making skills of an agent.
• Uses Bayesian networks to model the relationship between actions and skills.
• Updates joint probability distribution based on observed actions using Bayes’ rule.

Conclusion

• The complexity of AI agents requires new methodologies for optimization.
• The transition from LLMs to complex agents offers additional functionalities and performance improvements.
• Future videos will explore new solutions for optimizing agent performance.

Detailed Instructions and URLs

• No specific CLI commands, website URLs, or detailed instructions were provided in the transcript.