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?
22 / 31
BackNext

What a weight actually is

You already met a single weight, back when we watched a dog decide whether to bark. Smell and sound each fed into the decision, and each one had a weight: how much that sense counts toward barking. A strong smell should push hard, a faint background noise barely at all.

That is all a weight is: a dial controlling how much one input counts. Turn it up and the input matters. Leave it near zero and it barely registers. Turn it below zero and it argues against. Drag the dial and watch. Nobody sets these numbers by hand. Training finds them.

So that is one weight. Now the part worth naming clearly.

You will often hear models measured in parameters rather than weights. They overlap, but they are not quite the same word. A weight is the dial on a connection, how much one input counts. Parameters is the umbrella term for every number training adjusts, and that includes one more kind: the bias.

A bias is a neuron's baseline eagerness to fire, before any input arrives. Remember the dog's threshold, the bar a signal had to clear to set off a bark. A bias raises or lowers that bar. A high bias means the neuron fires easily, a low one means it stays quiet unless the inputs really pile up. Like weights, biases are just numbers, and training tunes them too.

So when you hear that a model has 70 billion parameters, that means 70 billion learned numbers, mostly weights, plus a bias for each neuron. Each one a small decimal. Each one found by training.

And here is where the simple picture breaks. With the dog, you could read each dial: this one watches smell, that one watches sound. In a real model, with billions of them, that tidy story falls apart. No single weight watches any one thing.

Citations
Citations