Sushant Gundla / Learn
About Sushant ↗

← Fig. 1 — one request, end to end

LLMs

The model layer, and how to choose in it.

Fig. 2 — Model, detail at larger scale

The thing in the middle, and which one to choose.

Key plan — part 3 on Fig. 1
18lessons4parts7min a lesson126min in total

Figure 2 — the parts of this course, as links to their lessons

What the detail shows

  1. The request is put together inside the context window — a bracket with a hard stop at the end of it. Nothing past the stop goes in.
  2. It crosses the line onto somebody else’s machine through a meter, which counts every token in and every token back out.
  3. Three ceilings sit above the line. A request that hits one is turned back and sent again.
  4. Inside the enclosure the weights are frozen, and one attention step runs across the tokens, repeated at every layer.
  5. What leaves the stack is a spread of possible next tokens. Three knobs used to choose between them; on the newest models all three are struck out, because the providers have taken them away.
  6. The same weights can sit in a second enclosure on your own machine, on your side of the line, where nothing is metered and the height of the box is the memory you have to find.
  7. Five candidates arrive at a switch: three struck out on your own cases, the provider, and your own box. One is wired through, and that is the model you get. Throwing the switch is how you change it.

Schedule of lessons

18 lessons

  1. The machine

    6 lessons · 42 min

    1. What a model is made of

      Frozen numbers, one repeated attention step, and two training stages. That is the whole product.

    2. Sampling, and the knobs providers are taking away

      Temperature, top-p, and top-k explained — and why some current models no longer let you set them.

    3. Base, instruct, and reasoning models

      Three different products wearing one word on the label. Picking the wrong one costs money and quality.

    4. Thinking and effort, model by model

      The same empty request thinks on one model and stays quiet on the next. Here is how to find out which, and how to turn it down.

    5. Context windows and their limits

      The window is a hard ceiling, and filling it is neither free nor harmless.

    6. Why models hallucinate

      A model that does not know still answers, because guessing scores better than admitting doubt.

  2. The providers

    5 lessons · 35 min

    1. The big three, and what actually differs

      Anthropic, OpenAI, Google, and the layers you can buy them through. Where the real differences live.

    2. What you actually pay for

      Per-token billing, the levers that move it, and how to see the bill before it arrives.

    3. Caching and batch: the two big discounts

      Pay less for the part of the prompt you repeat, and for the work that can wait. Both discounts can fail without an error.

    4. Rate limits and reliability

      Three kinds of limit, two kinds of 429, and a retry loop that knows when to stop.

    5. Timeouts, fallbacks, and failures that look like success

      How to stop a slow call hanging, what to fall back to, and how to catch an answer that came back 200 and wrong.

  3. Open weights

    3 lessons · 21 min

    1. Open weights vs closed APIs

      What open really means here, what the licence lets you do, and an honest trade that keeps moving.

    2. What fits on a GPU

      Weights, KV cache and overhead: the sum that tells you whether a model fits before you download it.

    3. Running a model yourself

      The tool ladder, the speed trade quantisation makes, and the idle-GPU arithmetic that decides whether it pays.

  4. Choosing

    4 lessons · 28 min

    1. Reading a leaderboard

      A ranking is a measurement with a method behind it. Read the method before you read the number.

    2. Picking a model

      Start from your task and your own cases, run more than once, not from a leaderboard.

    3. When a model is retired

      Every model id has an end date. Find yours before the provider's email does it for you.

    4. Switching models safely

      Why a prompt does not transport, and the recipe for changing model without a bad week.

Notes

What these models actually are, who sells them, what open weights really buy you, and how to pick one without guessing. The layer under everything else you build.

Every figure on this sheet is counted from the lessons themselves. Your progress is kept in this browser only — no account, nothing sent anywhere.