Part 3 opens after the part 2 check

Finish the check at the end of part 2 to open this lesson.

Back to the course

Lesson 9 of 9 · Part 3 · Check yourself · 3 min Draft

Final check

Five questions across the whole course, from the price sheet to the power station.

Five questions across the whole course. Look back at any lesson if you get stuck.

Check yourself

Get every answer right to finish the course.

1. Which part of a request usually costs most per token? The prompt you send The reply the model writes Prompt text read from the cache

Output is written one token at a time. Haiku 5.5 charges five times as much for output as for input.

2. What do you pay less for when your prompt starts the same way every time? The repeated part, read from the cache The reply Nothing, every token costs the same

Providers can keep a repeated opening ready and charge much less to read it again.

3. To write each new token, what must the chip do? Search the internet Retrain the model Read all the model's weights from memory Reread only your last sentence

Every token needs one full read of the weights, which is why memory speed sets the typing speed.

4. At most, about how much power can an H100 need once its server and building are counted? About 70 W About 700 W About 7 kW About 1.5 kW

Its share of the server is about 1,275 W, and cooling and power distribution add 10 to 20 percent.

5. Why does a GPT-4-level answer keep getting cheaper? Smaller new models soon match old big ones Electricity has become free Companies stopped charging for output

New small models soon match old large ones for a fraction of the price. On one hard science test, Epoch AI found GPT-4-level answers getting about 40 times cheaper a year.

← Watch a model type Finish: back to the course →

Books

The Model Race
All books
Errata & corrections

Learn

Facts & News
Courses

Build

Agents
Robotics

Lab

About

Company

Contact