Model choice lab
Which model should you use? The biggest one isn't always the right call.
The idea inside
Pick by your constraints: quality, latency, cost, privacy, context length, tool use.
After this lesson
You can choose a model for a task by weighing quality, latency, cost, privacy, context length, and tool use.
Where it leads
Inside this lesson
That's the real lesson stage, paused. Claim your pass to operate it.
See how AI actually works, end to end.
This lesson is one stop on the full arc. Unlock all of it, and keep it for life.
What you get
- The 34-lesson main path, a finishable route from a word to agents
- Goal tracks for using AI at work and building AI features
- Boss labs that make you apply a whole act, not just recognize it
- Spaced recall that brings each idea back before you forget
- Course memory: every term defined, with links to where it first appears
- A shareable capability card when you finish the main path
- Lifetime access on every device, every future lesson included
Not videos to watch. You predict, operate the machine, then prove it. That is why it stays.
99 interactive lessons and challenges. No videos, no code.
Free launch pass: lifetime access, no card needed
New here? The first lessons are free to try. Start with lesson 0.1
What this lesson shows
Pick by your constraints: quality, latency, cost, privacy, context length, tool use.
The question it opens with
Which model should you use? The biggest one isn't always the right call.
The walkthrough, in the lesson's own words
- Tap a task. The best-fit model lights up; the wrong ones dim.
- Offline on a phone. Pick the model you'd reach for first.
- There is no single best model. You match its profile to your constraints.
- Each card shows the same six axes. Longer bar = better on that axis.
- All ratings here are illustrative, not real model scores, but the trade-offs they show are real.
- Frontier has the best quality, but it lives in the cloud. With no signal it can't run at all.
- A cloud model is a reasonable instinct, it usually answers best. But a hard constraint isn't a preference you can trade quality for: it either clears the floor or it's out.
- Right, only the on-device model runs with no network. It loses on raw quality, but it's the only one that clears the hard constraint, so it's the only real option.
- What's a hard constraint vs a soft preference?
- A soft preference (a bit more quality, a bit lower cost) trades off against the others, you weigh them. A hard constraint (must run offline, data can't leave the device, must answer under 200ms) is a gate: fail it and the model is out, no matter how good the rest of its profile is. Always check the gates first, then weigh the rest.
- Frontier here means the GPT / Claude / Gemini flagship tier; Small is the mini / flash tier.
- You can match a model to a task by trading six axes, not by chasing one leaderboard number. The biggest model is just one column of the table.
- And the only score that's really about your task is your own eval set (lesson 7.10 / 7.3). Shortlist by profile here; decide with your own evals on your own data.
- A second axis under model choice: closed models (reached only through a vendor's API, like the frontier GPT and Claude models) usually lead on raw quality and need no hosting, but your data leaves your walls and you can't inspect or self-host them. Open-weight models (Llama, Mistral, DeepSeek, gpt-oss and friends) you can download, run on your own hardware, fine-tune, and audit, trading some peak capability for control, privacy, and predictable cost. The pick is rarely which is best, it's: do you need frontier quality, or control over where this runs?
- You're labeling millions of cheap support tickets and someone insists on the top frontier model for best quality. Is that the right call?
- Probably not: match the profile to the task, not the leaderboard. Bulk easy work is dominated by cost and speed, where a small model wins by a mile, so the frontier model's quality buys little while multiplying the bill. Save the big model for the genuinely hard calls.
Key takeaway
You can match a model to a task by trading off six axes, not chasing one leaderboard number.
What you can do after this lesson
You can choose a model for a task by weighing quality, latency, cost, privacy, context length, and tool use.
This is the written summary. The lesson itself is interactive: you predict, drag and operate the mechanism above, and the reveal answers you.