Converge
Open the app

Week 20 · Season 6 · Home

Bring it home

You can run an open model on your laptop, explain how bits and floats store numbers, quantize and benchmark it honestly, and check its licence.

Open week 20 in the app

This week

  • Bring the model home: open weights
  • How computers store numbers: bits and floats
  • Squeeze the numbers: quantization
  • Measure, don't vibe: fair benchmarks
  • Read the licence and the model card

How you learn

By hand: write 0.1 in binary until it repeats, round a list of weights to 4 bits with a scale and measure the error, then compute the median speed of a few timed runs before you trust any benchmark tool.

What you build

Private mode: FRIDAY runs on a small open model on your laptop and never calls the cloud while it is on, with a chat held with the Wi-Fi off and a fair benchmark against the cloud model.

Stretch: Benchmark two builds of the same model (8-bit and 4-bit) on your own test set and write down which one you would ship, and why.

What you ship

Project 6: Teach a small model. A small open model fine-tuned with LoRA to answer in your format, with an eval that proves it.

  • FRIDAY answering with the Wi-Fi switched off, and tests that prove private mode never calls the cloud
  • BENCH.md: median tokens per second, time to first word, memory and score, with n
  • MODELS.md: the model, its licence link and your five answers

Lessons

  1. Bring the model home

    After this deck you can explain what open weights are (a model's learned numbers, free to download), how a small program runs them as a server on your own laptop (self-hosting), and switch FRIDAY to that local brain by changing one address.

    12 minFull access

  2. How computers store numbers

    After this deck you can explain how a computer stores a whole number in bits, why a floating-point number like 0.1 is only ever stored approximately, and how to compare floats and handle money safely in Python.

    12 minFull access

  3. Squeeze the numbers

    After this deck you can work out how much memory a model's weights need, explain quantization (rounding every weight to fewer bits, with a shared scale), and choose the build that fits your laptop with room to spare.

    12 minFull access

  4. Measure, don't vibe

    After this deck you can run a fair benchmark of two builds (same prompts, a warm-up you throw away, several timed runs), report speed as the median tokens per second, and pick a build only after it passes a quality bar on your own test set.

    12 minFull access

  5. Read the licence

    After this deck you can find a model's licence on its model card, tell an easy-going licence from one with extra rules (user caps, banned uses), and answer five questions before you build on a model.

    11 minFull access

  6. Go deeper: formats, outliers and memory maths

    After this deck you can read a GGUF build name (Q8_0, Q4_K_M), work out bits per weight for a block format, explain why outlier features make rounding hard and how AWQ protects them, and estimate a model's full memory: weights, plus the conversation's cache, plus room for the runner.

    14 minFull accessGo deeper · optional

Labs

Python in the browser, next to the lessons.

Weekend build

Hand-picked official docs, papers and talks.