Large Language Models (LLMs) are often experienced as black boxes through chat interfaces or APIs. Running a local LLM enables direct inspection of its architecture, weights, and output logits. This provides insight into how the model operates internally.

After a brief introduction to the core ideas behind LLMs, I will demonstrate how to load a small pretrained model, run it locally, inspect its outputs, and fine-tune it to observe concrete improvements in behavior. All examples will run on a standard laptop and will not require GPUs, API keys, or paid subscriptions.

The talk will include live code demonstrations. Attendees may optionally follow along on their own machines, but this is not required. The focus is on building a clear mental model of how LLMs work internally and showing that experimenting with them locally is both feasible and practical.

This session is aimed at anyone curious about how LLMs work and how to explore them beyond black-box interfaces. Basic Python experience is helpful, but no prior AI or machine-learning background is required. Code to reproduce the talk workflow will be available in an open-source GitHub repository.

Outline (45 minutes):

• Motivation and setup, including loading the model
• Core LLM concepts at a high level (tokenization, embeddings, attention, transformers, logits)
• Live demo: running a local LLM
• Live demo: fine-tuning and observing model improvements
• Wrap-up and next steps