Exactly - but they can't. I've been using Substack recently - the algorithm for its Notes functionality is an eye-opener, adjusting the stream on the fly as you interact with content. Admittedly, it's matching only for 'apparent relevance', not 'quality'. But that's still half the battle, and as Medium already has a quality metric for each story through constantly updated user read ratio and engagement stats, it would be simple to factor this in:

User X just clicked on a story with topic tags (or, in fact, content analysed word by word as relating to) A, B, C. We have 10,000 articles that could match that reader's interests. We will now include them in User X's feed, ordered according to that user's preference: by date, relevance or quality, or our own blended default recommendation.

I cannot imagine coding such a system would be beyond the wit of their software engineers. If they're paying a senior project manager $300,000 a year, they should be able to find someone to improve the algorithm.

In fact, I reckon a high school computer science student could do better.