How a reply is built · Lesson 10 of 18
The same block, many times over
The same shape of block repeats many times, each with its own learned numbers, and every piece passes through all of them.
How it works, step by step
One block does two jobs
A block runs attention so pieces can share information, then the per-piece part reworks each one. Its output has the same shape as its input.
The same pieces, slightly changed.
Stack the blocks
Blocks are stacked in order. Each has the same shape but its own learned numbers, so it does something different to what it receives.
Four blocks, each with its own weights.
Pass through every block
The pieces go through block one, then two, then three, then four. Only after the last block are scores worked out.
Pieces ready for the scoring step.
In plain language
A model is not one large step. The same shape of block, attention followed by the per-piece part, is stacked and repeated many times with different learned numbers.
A way to picture it
Think of a relay of editors passing one draft along. Each reads the whole thing, makes a small improvement, and hands on a draft of the same length.
A worked example
Say a model has 4 blocks. Your pieces pass through block 1, then 2, then 3, then 4, and each pass uses that block's own learned numbers.
Keep in mind
Every block stores its own weights, so this is not one set of numbers applied repeatedly. More blocks usually mean more stored numbers, more memory and more work for every token.
What these words mean
- Block or layer
- One repeat of attention plus the per-piece part.
- Depth
- How many blocks are stacked. More blocks usually mean more stored numbers.
- Hidden state
- The lists of numbers passing between blocks. They are not readable text.
- Shape
- How many pieces and how many numbers each carries. A block keeps this the same so blocks can be stacked.
Do all the blocks share one set of learned numbers?
No, every block stores its own. Each block keeps its own weights. That is why a model with more blocks holds more numbers, needs more memory, and does more work for every single token.
Where this comes from
This explanation is checked against primary documentation. The small arithmetic examples are ours and are not hardware measurements or vendor benchmarks.