by Radsym Sendyka
Session 6 - “APIs to LLMs” notebook-based presentation - I go through content included in a google colab notebook shared with participants, explaining concepts and code examples as we go along. Go from a place where most learners have only interacted with LLMs via websites, to a place where they’ve all called a few different models successfully, and have a better understanding of what goes on behind the scenes. Information on what kinds of models are best for what tasks, what criteria are useful to judge “best”. Spare time for setup - Google Colab, GitHub, distributing API keys.
[Soumya’s last session 5 will inform where I start - might include a little more theory of LLMs before the API examples, tbd based on conversation with him and his opinions]
Session 7 and First Half of 8 - First Hackathon - learners in groups of ~4, creating a shared project that uses a few different APIs, to automate some complex task. Using an openrouter API key (one per group, initial spend limit ~10$ per group, flexibility to get more if they are doing something really cool)
Topic: TBD - happy to involve something C2D3-related, or following up on Soumya’s teaching. Can also leave a free-hand option to groups, have them come up with toy ideas, consult with TAs and me/Soumya.
Success criteria:
1) At least 2-3 different models used as part of a workflow, in most cases including at least one non-text model. 2) Some external action(s), like reading / writing files, running python functions, writing code, internet search… 3) Code on GitHub, all group members contributing at least one commit.
TAs to keep short notes on group progress, share with lead instructors.
Note: GoogleColab is not the best coworking environment, so might have to think about how to cooperate. I have ideas.
Second Half of Session 8 - A short introduction to Agentic AI. What is agentic AI? Benefits over the LLM pipelines we were building. Agency Autonomy Intro to frameworks like LangGraph and GoogleADK Introduce additional goals for the hackathon projects - to make them Agentic. Success criteria now include Agency (replacing hardcoded external actions with agent-decided actions), and Autonomy (replacing hardcoded LLM call order with agent-decided order) Last ~15 minutes back in groups to discuss how to apply this in projects