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?
16 / 31
BackNext

Where your memories live

Take a moment with something strange. Think of your grandmother's face. Or the smell of rain. Or how to ride a bike. Now ask: where, exactly, is that stored?

It feels like there should be an answer. A folder in your head labelled "grandmother." A file you open. But there isn't. Your brain has no hard drive, no shelf, no single cell holding the memory. If a surgeon went looking, they would never find the spot where "grandmother" lives, because there is no spot.

Your brain is roughly 86 billion neurons, tiny cells, each one connected to thousands of others. A neuron on its own knows nothing. What matters is the connections between them, and how strong each connection is. Learning something does not write a file anywhere. It changes the strength of those connections. A memory, as best neuroscience can tell, is not a thing you store. It is a pattern: a particular set of neurons that fire together, in a particular way, because the connections between them were tuned by everything you have ever experienced.

This should sound familiar. It is exactly what we just watched happen to the model. The weights are its connections. Training tuned their strengths, one tiny nudge at a time, until knowledge lived in the pattern rather than in any single number. Nobody can point to the weight where "the sky is full of stars" is stored, for the same reason a surgeon cannot point to your grandmother. The knowledge is spread across the whole.

The analogy only goes so far. A real neuron is vastly more complex than a number, and the brain almost certainly does not learn the way a model does. But the deepest idea holds in both: understanding is not kept somewhere. It emerges from connections, tuned by experience, distributed across the whole.

Did you know?

Brains don't do backpropagation. Backprop works by sending one big error signal backward along the exact same connections the signal travelled forward. Real neurons can't do that: there's no known wiring that carries a perfect, reversed error back down the same path, and most of the time the brain has no labelled "right answer" to measure against. Instead it seems to learn with local rules, the famous one being "neurons that fire together, wire together." A connection strengthens based on what the two neurons it links are doing, right there, not on a global correction mailed back from the output. The destination is similar, knowledge living in tuned connections. The route is completely different.
Citations(4)↓
  1. 1. doi.org
  2. 2. news.mit.edu
  3. 3. en.wikipedia.org
  4. 4. doi.org
Citations(4)↓
  1. 1. doi.org
  2. 2. news.mit.edu
  3. 3. en.wikipedia.org
  4. 4. doi.org