The thing is, HN doesn't just contain links for the intellectually curious. It often has stories about a new school of thought, or an approach to design, or a new programming language or something. However, it also often has stories about Internet Drama in the tech startup community, or American political events, or practical advice for people entering the business world for the first time. They're popular here because they're relevant to a large proportion of the HN user-base, but it's certainly possible for someone to be both intellectually curious and uninterested in American politics.

I would argue that American politics as they affect the internet and copyright impacts pretty much everyone, unfortunately.

Which doesn't necessarily mean everyone would be interested in reading about the newest developments weekly.

> It often has stories about [...] a new programming language [...]

That's something I think could be automated (at least for me.)

I can (and do) guess whether I am likely to be interested in a new programming language, based on what other languages the commenters compare it to.

In my case it would probably be something like:

* Compared to Haskell, Io or SETL => very interesting

* Compared to Java, JS or Scheme => neutral

* Compared to Groovy, PHP and VB => uninteresting

I like the approach of going outside of your "comfort zone" and start discovering new things. Limiting my exposure to what I enjoy most is not an intellectually curious strategy.

> By filtering out stuff, you'll never expose yourself to things outside your "pattern".

What makes you think that a Naive Bayes classifier will automatically put all unseen/unexpected/surprising data in just one of two categories?

Indeed, to the classifier it's just two categories, it doesn't "know" what's "in" or "out".

In fact it is much more likely that the classifier will sort of distribute the data that is unexpected (read: doesn't contain many features that are trained on) rather evenly among the two categories based on other features. Which is exactly what you would want it to do.

The author also said he plotted the precision-recall curves? (too bad he just showed a screenshot with the numbers instead of the graphs) That sort of analysis is bound to bring out such behaviour.

I wonder if a filter that excluded things that I've trained it that I don't like would be more effective?