No video of the event yet, sorry!

Large Language Models (LLMs) are increasingly described as reasoning systems, coding assistants, or “AI agents.” These descriptions are useful for product positioning, but far less helpful when trying to understand what the system is actually doing.

This talk takes a practical, grassroots view of modern LLMs as they appear in real developer workflows: GPT-class models, Claude-style systems, and tool-augmented coding assistants (Codex-like or otherwise). The focus is not on branding, but on mechanism.

We will walk through the basic pipeline: training on next-token prediction, emergence of high-dimensional probability distributions, and inference as constrained sampling. From this perspective, “reasoning” is not a new capability, it is a byproduct of scale, data, and optimization pressure on likelihood.

We then connect this to observed behavior in real systems: why these models appear to plan, why they fail confidently on trivial edge cases, and why tool use and agent frameworks improve usability without changing the underlying model.

The goal is to establish a usable mental model for engineers building on top of these systems, particularly in the current ecosystem of coding assistants and LLM-integrated tooling.

Attendees will take away:

1. A working mental model of LLMs as probabilistic systems
   
2. A clearer separation between observed behavior and inferred “intelligence”
   
3. Intuition for common failure modes in LLM-based tools
   
4. Context for why agent frameworks help and what they do not change