Three Kinds of Attention Pro Trailblazer

See the same skeleton as:

Step 1

Score

Step 2

Normalize

Step 3

Allocate

Same machinery, three jobs. The difference that matters most is in step 2 and 3. Search and social sort into a ranking and take a hard cut. Attention runs a softmax that produces weights summing to 1 and blends, it never crowns a single winner. Open the Compare tab for all five axes.

Click a page, then another, to add or remove a link

Top results

PageRank produces a hard ranked list. The top few are page one.

Real PageRank power iteration: PR(p) = (1 - d)/N + d · Σ PR(q)/outdegree(q) over pages q that link to p, with damping d = 0.85, run to convergence. Ranks sum to 1.0. Authority is conferred recursively: you are important if important pages link to you. It is structural and slow to change. PageRank is the classic core of search ranking. Real search engines layer hundreds of additional learned signals on top, so this is the foundation, not the whole system.

Likes weight 0.60

Dwell weight 0.15

Shares weight 0.55

The feed sorts by predicted engagement, not quality. Each round it allocates impressions by rank, draws sampled engagement (every like, dwell, or share is a real Bernoulli draw against the post's underlying probability), updates a running estimate, and re-sorts. That is a live feedback loop, and it produces rich-get-richer dynamics: the posts on top get the most impressions, so they gather the most signal and stay on top. Quality and the engagement probabilities here are illustrative, but the loop is real. Sensational, novel posts win likes and shares. Quality mostly drives dwell. Push the dwell slider up to realign the feed with quality.

Head:

Predicate:

Attention matrix (rows attend to columns)

Query weights

weights sum to 1.00

Real scaled dot-product attention: scores = (q · k) / √d, weights = softmax(scores), output = Σ weight · value. Every token gets some nonzero weight and the weights for a query sum to 1. The pattern is recomputed fresh for each input and it all serves next-token prediction. There is no human feedback loop inside the forward pass. Vectors are hand-built so the example reads clearly. One declared simplification: the pronoun's query folds in the predicate (big or small) from its clause, which is the job deeper layers do in a real model. Swap big and small to watch what "it" refers to flip between the trophy and the suitcase.

Axis	Search	Social feed	LLM attention
What is scored	Pages	Posts	Tokens
Objective	Link authority	Predicted engagement	Next-token prediction
Temporality	Slow, structural. Changes as the link graph changes.	Live feedback. Re-sorts every round from observed engagement.	Recomputed from scratch for every input.
Output shape	Hard ranked cut	Hard ranked cut, plus feedback	Soft blend, weights sum to 1
Failure mode	Gaming the link graph (link farms, spam)	Optimizes a proxy that can drift away from quality and truth	None of that kind, by construction (no objective beyond prediction in the forward pass)
Winner-take-most	Authority concentrates on hubs everyone links to	Impressions concentrate on whatever is climbing	Softmax sharpens onto the most relevant tokens, but never to a single winner

The honest part

The same score, normalize, allocate machinery is benign when the target is prediction or relevance, and corrosive when the target is engagement. Engagement is a proxy. It stands in for "what the user wants," but it can pull away from it, and the feedback loop amplifies the gap. Search can be gamed and attention is just arithmetic, but only the engagement objective has an incentive to show you something worse for you because it predicts you will click.