Loading slide

What a Trained Model Actually Is

  1. 01What a Trained Model Actually Is
  2. 02Memory card
  3. 03A model is a file
  4. 04Two files
  5. 05How the two fit together
  6. 06A Piano
  7. 07The valuable file
  8. 08You connect to it
  9. 09Where does training data come from?
  10. 10Cleaning the data
  11. 11The network can't read
  12. 12Guess the next word
  13. 13Right or wrong
  14. 14One nudge, a billion times
  15. 15Take a breath
  16. 16Where your memories live
  17. 17A base model is not an assistant
  18. 18Why it's called pre-training
  19. 19The weights are the model
  20. 20The scale of it
  21. 21Why so many?
  22. 22What a weight actually is
  23. 23Everywhere and nowhere
  24. 24Using the model: inference
  25. 25The whole arrangement
  26. 26Reinforce your understanding
  27. 27Question: What is a model, really?
  28. 28Question: Learning from raw text
  29. 29Question: The model doesn't know what it doesn't know
  30. 30Quiz: answer
  31. 31The machine learned something real.
  32. 32Want to go deeper?
20 / 31
BackNext

The scale of it

The honest answer: more than you can picture. So instead of a number, let's feel the growth.

One quick note on words first. When people count the numbers in a model, they usually call them parameters rather than weights. For now, treat the two as the same thing. More on that word soon.

In the late 1980s and 1990s, when neural networks first started working, a large model might have had a few thousand parameters. Enough to fill a few pages of text. Researchers were proud of that.

By the early 2000s, the number crept into the millions. Still small enough that you could imagine printing them out, given enough paper.

Then something changed. Computing got faster. Data got bigger. And the numbers started climbing in a way that felt almost violent. Within a few years, models crossed a billion parameters. Then ten billion. Then a hundred billion. Then more.

Large language models can contain billions of parameters, or sometimes far more, though the exact numbers are not always made public.

To put a hundred billion in perspective: if you read one number per second, without stopping, it would take you over three thousand years to read them all.

When a company says they have released a new model, this is what they mean: they have produced a new file, with a new configuration of numbers that behaves differently from the ones before it. Different models have different sizes. A smaller model might run on a laptop. A larger one usually requires specialised hardware and a lot of electricity to run quickly. That is part of why you connect to their servers rather than running it yourself.

The file holding those numbers is not a document or a spreadsheet. It is closer in size to a film archive. And yet it fits on a hard drive, travels across the internet, and produces answers in milliseconds.

Which raises an obvious question. Why would anyone want that many numbers in the first place? What do you even get for them?

Citations(2)↓
  1. 1. arxiv.org
  2. 2. arxiv.org
Citations(2)↓
  1. 1. arxiv.org
  2. 2. arxiv.org