Series · 7 parts

The LLM Playground

  1. 1

    Building an Instrument, Not a Model: Why My First ML Project Was a Playground

    Why I built an LLM playground instead of training a toy model, and what the architecture teaches before any experiment runs.

  2. 2

    Watching a Model Fail on Purpose: Tokenization and the Decoding Tradeoff

    How text becomes tokens, how probability distributions become text, and what happens when you push both to failure on purpose.

  3. 3

    The Same Model, Before and After: What Post-Training Actually Changes

    A controlled comparison of a base model and its instruction-tuned sibling, testing whether assistant behavior is installed or emergent.

  4. 4

    Who Judges the Judges? Small Models Grading Small Models

    Using small models as evaluators for other small models, and finding out how much you can and can't trust the judge.

  5. 5

    It Works on My Laptop: The Distance Between Running and Shipping

    The gap between an application that runs and one that ships, written from an actual deployment rather than a plan for one.

  6. 6

    The GCP Deployment Runbook: Every Command, Every Fix

    The exact commands, in order, including every failure and fix, for deploying a FastAPI + React app to Cloud Run.

  7. 7

    What Building an LLM Playground Actually Taught Me

    Every claim the series made, in one place, stated plainly and traced back to the experiment that supports it.