LANDMARK · 8 MIN · UNDERSTAND
Inside a transformer
After this landmark, you can describe what attention does in plain terms, and explain why model quality is not simply a function of parameter count.
↗
This concept is shared, but the Everyday lens is less central here.
You can still explore it. We’re showing the shared explanation and a related practical view without hiding the knowledge.
A transformer processes all the tokens in its input together rather than one after another. Its central mechanism, attention, lets each position look at every other position and weight how much each one matters for interpreting it. That is how “it” gets connected to the noun it refers to four sentences earlier, and how “bank” is read as a riverbank or a financial institution depending on what surrounds it. The model stacks many layers of this, interleaved with ordinary arithmetic layers, and the weights in all of it — the parameters — are the numbers adjusted during training. A model’s parameter count tells you roughly how much capacity it has to store patterns. It does not tell you how good it is: data quality, training length, and post-training work matter enormously, which is why a well-trained smaller model routinely beats a larger, older, or more carelessly trained one.
This is the canonical concept. It stays the same across learner lenses so personalization never changes the underlying facts.
◌ EVERYDAY LENS
What this looks like for you
You don’t need the internals to use these tools, but one consequence is worth carrying: because every token is read in the light of the others, what you put around your question changes how the question is understood. Context isn’t decoration — it changes the answer.
MAKE A DECISION
Two candidate models: one with far more parameters released eighteen months ago, one much smaller released recently. Which should you expect to perform better on your summarisation task?
CARRY THISTake one prompt you use regularly and run it on a large model and a small one. Note whether the difference in output justifies the difference in cost and latency.