UC Berkeley Agent/AI Free Public Course Speakers: Dawn Song (Professor at UC Berkeley) Chen Xinyun (Research Scientist at Meta) Special Guests: Including OpenAI's VP of Technology, NVIDIA's Research Scientists, OpenAI's Co-founder, and more **Line 1: Training pipeline, not prompt engineering.** OpenAI's Yann Dubois' lecture is a must-listen. He breaks down pre-training, reasoning RL, and post-training: the difference between chatbots and agents isn't about whether they can use tools, but whether the environment and feedback form a closed loop. NVIDIA's Jiantao Jiao follows up with a talk on verifiable agents—code, browsing, task completion—whether they can be verified is more important than Arena scores. Microsoft's Weizhu Chen gives a more down-to-earth talk about how data, environment, and reward hacking can mess you up when training agentic models. **Line 2: Evaluation is both overrated and underrated.** There's a dedicated evaluation week in the course, where Meta's Sida Wang discusses predictable noise in LLM benchmarks. Nowadays, there are tons of leaderboards out there, and many people treat them as gospel. This course focuses less on leaderboard hype and more on error bars, reproducibility issues, and how evaluation itself can be an attack surface. If you're working on agents and don't address this, you'll just be spinning your wheels later. **Line 3: Multi-agent systems and embodiment.** Noam Brown and Oriol Vinyals' lectures trace the old threads of game theory and coordination into the LLM era. Peter Stone's lecture brings agents back to the physical world and environment—his quote is spot on: "Agents must have an environment." This ties into the current wave of personal assistant hardware, robots, and simulations transitioning to real-world machines. James Zou uses agents for scientific discovery, while Clay Bavor talks about Sierra-like real-world customer service agents—one leans toward scientific automation, the other toward product implementation. **Line 4: Security is not an appendix.** Dawn Song's closing lecture is a must-hear. Once agents can read files, use tools, and access the internet, issues like prompt injection, tool misuse, backdoors, and red teaming go far beyond the alignment stories of chatbots. Her lecture is deeply connected to her own security research—this isn't just a guest appearance. If you're short on time, focus on these five lectures: Dubois, Jiao, Evaluation/Sida Wang, Bavor or Weizhu, and Dawn Song. For multi-agent systems, add Brown/Vinyals; for robotics, add Stone. https://agenticai-learning.org/f25 (Includes YouTube videos and PPT materials)