A three-day workshop about understanding, not just doing: how language models actually work, what retrieval really does, what turns an LLM into an agent — and why these systems fail. Every lecture comes with a hands-on lab.
Led by Pavlos Protopapas (Harvard University) and Guillermo Cabrera-Vives (Universidad de Concepción).
Large language models are easy to call and hard to understand. Anyone can wire up an agent in ten lines of code; the hard part is knowing why it works — and why it breaks. This is not a “how to call the API” workshop. It takes practitioners from “I can prompt a model” to “I understand what is happening inside my LLM system, and can design, build and judge one”. It has one through-line: an LLM is not an agent. A model predicts the next token; a useful system wraps that model in retrieval, tools, memory and a control loop — and each layer adds capability, cost and new ways to fail.
Lectures build the mental model: what is going on under the hood, and why. Labs in Google Colab, using the OpenAI Agents SDK, put that model to the test — you change one thing, predict what happens, and see the trade-offs that only appear in code. The goal is to understand, not to copy a recipe. Two industry talks, from CMPC and CDIA (Universidad de Concepción), show how these systems are used in Chile today.
What a language model is, from transformers to ChatGPT, and why prompting alone is not enough. Then retrieval-augmented generation, from naïve RAG to self-correcting variants.
Reasoning models, the agent loop, tools and MCP; then context engineering, harnesses and skills. Labs: the Cheese Quiz agent and a Travel Planner.
Which model and which surface for which job, when more than one agent helps, and the limitations, risks and evaluation of LLM systems.
Engineers, ML engineers, data scientists and technical leads who are comfortable in Python and have used an LLM API. No prior experience with agent frameworks is needed.
Slides and notebooks are posted here before each session. All times are Santiago time.
| Time | Session | Materials |
|---|---|---|
| 9:00–9:10 | WelcomeWelcome and practicalities The idea of the workshop, the three days, the team, Wi-Fi and logistics, and the readiness check. | Slides soon |
| 9:10–10:30 | LectureLanguage models I — from n-grams to attention 1 What is a language model? Tokenization, n-gram models and their limits. 2 Language models with neural networks: word embeddings and feed-forward models. 3 Attention and transformers (continues after the break). | 1 · Language models2 · Neural LMs3 · Attention |
| 10:30–10:50 | Coffee break | |
| 10:50–12:30 | LectureLanguage models II — from transformers to ChatGPT 3 Attention and transformers, continued. 4 BERT and GPT: masked and autoregressive models, and how GPT generates text. 5 From GPT to ChatGPT: instruction tuning, reward models and RLHF. | 4 · BERT & GPT5 · ChatGPT |
| 12:30–13:30 | Lunch | |
| 13:30–15:10 | LabChatGPT lab — from a base model to an assistant See the morning's ideas in code: GPT-2's next-token probabilities, an instruction-tuned model, temperature and log-probabilities, tokens and cost, why the model has no memory, reasoning models and structured output. Seven short practice exercises. Runs in Colab. | Notebook soon |
| 15:10–15:30 | Coffee break | |
| 15:30–17:00 | LectureRetrieval-Augmented Generation Naïve RAG, better retrieval (hybrid search, re-ranking), Self-RAG and Corrective RAG. Optional: fine-tuning with LoRA/QLoRA. | Slides soon |
| 17:00–17:30 | Industry talkIgnacio Becker — CMPC | |
| 17:30–18:00 | LabRAG lab Chunk, embed and search eleven books about cheese, then let the LLM answer from them. Runs in Colab — no Docker needed. | Notebook soon |
| Time | Session | Materials |
|---|---|---|
| 9:00–9:30 | CapstoneCapstone brief — design an agentic system The capstone is a design exercise, not a build. Form teams of 4–5 and pick a scenario; talk it through in the breaks. Hand in an architecture design and a technical design (six slides) by Saturday 9:00. | Slides soon |
| 9:30–10:30 | LectureAgents I Reasoning models, what makes something an agent, how the agent loop works, tools and MCP. | Slides soon |
| 10:30–10:50 | Coffee break | |
| 10:50–12:30 | LabAgents I lab — The Cheese Quiz Build a quiz-writing agent with the OpenAI Agents SDK. | Notebook soon |
| 12:30–13:30 | Lunch | |
| 13:30–15:10 | LectureAgents II Context engineering, the agent harness, skills, single- vs multi-agent systems, and evaluating agents. | Slides soon |
| 15:10–15:30 | Coffee break | |
| 15:30–17:00 | LabAgents II lab — Travel Planner A planning agent with several tools. | Notebook soon |
| 17:00–17:30 | Industry talkLead ML Engineer, CDIA — speaker to be announced | |
| 17:30–18:00 | CapstoneWorked example — designing a pisco sour assistant We fill in the same six slides you will hand in: the architecture decision, the diagram, the technical design, the evaluation plan, cost and risks. Then: write up your design tonight. | Slides soon |
| From 18:00 | Pisco sour reception After the worked example: a pisco sour with the teaching team, the speakers and each other. A good moment to talk your capstone design through. | |
| Time | Session | Materials |
|---|---|---|
| 9:00–9:45 | LectureThe AI landscape 2026 & multi-agent systems Five ways to use AI today, open vs proprietary models, cost levers — then multi-agent patterns and when multi-agent is overkill. | Slides soon |
| 9:45–10:15 | LabMulti-agent lab — The Cheese Quiz Guild An orchestrator and a guild of specialist agents, called as tools, write a cheese quiz together. Build one specialist, then read the trace to see who decided what. Handoffs and a cost guard are stretch goals. | Notebook soon |
| 10:15–10:35 | Coffee break | |
| 10:35–11:20 | LectureLimitations, risks and evaluation Hallucination and sycophancy, bias, privacy, jailbreaks and prompt injection — and how to build an evaluation that tells you something. | Slides soon |
| 11:20–13:00 | CapstoneCapstone reviews and discussion Designs are due at 9:00. Selected teams present in five minutes — problem, decision, diagram, evaluation — then one hard question from us, and a discussion with the room. | |
About 15 minutes. Please do this before Thursday morning.
OPENAI_API_KEY, and run all cells. The last line should print ✅ ready. Every lab works the same way: click Open in Colab, save a copy, run.Instructors, organizers, teaching and support staff, and guest speakers.
Harvard University
Department of Computer Science, Universidad de Concepción. Leads the Day 3 multi-agent lab.
Department of Computer Science, Universidad de Concepción.
Executive Director, Centro de Datos e Inteligencia Artificial (CDIA), Universidad de Concepción.
Workshop support — ask her about logistics.
Supports the labs.
CMPC
Lead ML Engineer, CDIA, Universidad de Concepción
For registration, the venue, payments and anything practical, write to the CDIA team at cdia@inf.udec.cl. Questions about the content go to the teaching staff in the room.
No. You should be comfortable in Python and have called an LLM API at least once. We start agents from first principles.
Lectures and materials are in English.
You use your own OpenAI API key. The labs are designed to cost a few dollars in total.
Yes. Slides and notebooks stay on this page after the workshop. Please do not redistribute them without permission.
Run the readiness check first. If it does not print ✅ ready, bring the output to any member of the teaching staff before the first lab.