mickelb.org

AI Agents - A look at free resources to get started

2024-12-12 | Mick Kalle Mickelborg

AI Agents in 2025

AI agents will become an even bigger focus globally in 2025. AI Agents have the capability to break the boundaries of conventional language models beyond text generation into the realm of intelligent agency, decisionmaking and function calling. The amount of resources available is growing every day and at the same time the barrier of entry is diminishing. However, for many, it will also seem daunting with the evergrowing amount of resources out there to get started. I am a firm believer that building something from scratch is the right approach, and in order to do so, you need both the theoretical and practical knowledge to get build AI Agents.

This course (offered free of charge, by the way) by Dawn Song at UC Berkeley delves into hands-on approaches to building AI Agents from scratch. The curriculum encompasses foundational LLM concepts, prompting (zero-shot, one-shot, few-shot, Chain-of-Thought reasoning), planning, function calling, and infrastructure development for agent deployment. It also explores specific applications such as code generation, robotics, web automation, medical fields, and scientific discovery. The course is offered by legendary Dawn Song and consists of lectures from esteemed professionals from Google DeepMind, OpenAI, Anthropic, Meta AI - making the course that much more of a no-brainer to anyone interested in a comprehensive understanding of LLM agents' capabilities and challenges.

Here are the best aspects from the course:

  • Knowledge you can't get anywhere else: The course offers emerging foundations and frameworks for building AI Agents. Instead of skimping over the details, the course also lists research papers that created the foundation for many of the assumptions underlying AI Agent development, such as ReAct, StateFlow, AutoGen, etc. I believe this is important for understanding, thoroughly, what AI Agents entail, rather than just applying them.
  • First-hand projects/labs: The course offers real projects to build your own AI agent, which is an excellent way to introduce function-calling, prompt optimization, reasoning methods, etc.
  • Diverse application domains: Examples include automating coding tasks, synthesizing scientific research, supporting healthcare decisions, and navigating the web to retrieve and analyze information. The course also covers applications either in personal use or scaling to enterprise solutions.
  • Excellent support on Discord: Participants in the course had access to TAs and other peers through the (very active) Discord channel. The help form the TAs and supporting crew was top-notch, particularly considering how many participants were enrolled.
  • Access to AI Agent Hackathon: The course was accompanied by a remote and sponsored hackathon to build your own AI Agent and an opportunity to win prizes. I for one used this opportunity to work on a project of mine (heavily inspired by the StateFlow research paper) in which I build a stateful AI Agent to extract medical research knowledge from research papers through PubMed's API.

Authors Notes

I believe that the potential for AI Agents is growing with the growing body of research being released every day. New methodologies and approaches to agent reasoning, planning, and self-optimizing function calling is something that SHOULD excite anyone in this domain. A deep interest of mine is how AI Agents can be used to further the agenda of AI safety, particularly pertaining to mechanistic interpretability either by opening doors to understanding large dimensional spaces through automated sparse autoencoding research or the analysis of polysemantic and monosemantic neurons for concept extraction.

There will be a new coming course offered in Spring 2025, albeit more focused on advanced implementation of AI agents in mathematics (which I find super exciting.