Working out
Notes
The entries state what a thing is. These state how I got there — the derivation taken slowly, the wrong turning that made the right one make sense, the question I could not answer and left standing.
17 notes, 1 of them still open — marked at the top rather than quietly left.
Attention from scratch
6 partsbuilding the operation up from the problem it was invented to solve, one step at a time
017 min029 min038 min049 min059 min0610 min
Why attention exists
The fixed-size bottleneck that broke sequence-to-sequence models, and the obvious fix nobody could take.
Building the attention weights
From "which positions matter" to softmax over scaled dot products, by elimination.
Reading an attention matrix
What the n×n matrix contains, what the pretty heatmaps show, and how much of it is an artefact.
The quadratic problem
Where the n² really comes from, and why the escape routes divide into exactly three kinds.
Many heads at once
Why one attention operation is not enough, and what the heads actually turn out to be doing.
What attention cannot do
Four limits that are structural rather than incidental, and the behaviours each one explains.
Alignment from scratch
5 partsfrom "a base model completes text" to a policy trained against a learned reward, one honest step at a time
017 min028 min039 min0410 min0510 min
What alignment is asking for
A base model is not disobedient, it is doing something else entirely — and naming that is the first step.
Where demonstrations run out
Fine-tuning on good answers works, and then stops working for a reason worth understanding.
Turning comparisons into a number
Bradley–Terry, a scalar head, and the moment the proxy starts lying to you.
The leash
The KL term is not a stabiliser. It is what keeps the reward meaningful, and PPO is what it costs.
Removing the reward model
DPO inverts the closed form from part 4, and what it gives up is not obvious until you look for it.
Making it fit
4 partsthe memory arithmetic of training something larger than the machine you have
018 min028 min039 min049 min
Counting the bytes
Where training memory actually goes, and why the answer is rarely the weights.
Throwing work away
Recomputation is a trade with a bad exchange rate applied uniformly, and a good one applied selectively.
Splitting the work
Data parallelism is one collective, and the reason it scales when nothing else does.
When it all runs at once
Four parallelisms, one cluster, and the silent misconfiguration that halves it.