Hacker News Reader

Comment by rmk | Hacker News Reader

This is common. Most publication houses, including magazines and newspapers, have a 'house style' that is enforced by an editorial staff.

View on HN

This is very similar to how code bases are 'edited' to fit the coding style of the team/project/company.

I guess the main difference is that code editing is done in collaboration with the author (code review) as opposed to an editor.

If refactoring code did not have the potential side-effect of changing its run-time behaviour (or performance) I'd argue that software engineering would also benefit from a 'code editor' role :)

Most companies institute code reviews and invest in tooling that stops nonstandard code (formatting and convention-wise) getting into the codebase, once they grow large enough. I think these two processes serve the same purpose, albeit with numerous exceptions depending on the seniority of the committer and reviewer(s) :)

They can probably be replaced by a neural network in time, leaving humans to just doing content reviews.

Doesn't 'the voice' evolve overtime though? How can a neural network evolve on its own.

Train it every year on the senior editors' publication.

So we can't get rid of the senior editors after all?

Why get rid of them? Automatisation should not mean getting rid of people, it should mean letting them do other things. In this scenario the editors could spend more time on their own writing instead of having to look at junior writer's texts

Editors are not necessarily investigators.

(metaphor/similar example follows)

In the same sense that someone who reviews a (IT) technical procedure is not necessarily a systems administrator, but someone who sees things from an angle to ensure RACI model is followed throughout without any generalizations such as "backup is taken" (by who? using which tools? which systems/data are backed up? how often? how long do we keep the tapes? how often do we trash the tapes? where do we store the tapes?). I haven't taken a backup for more than a decade, but I would shred a backup procedure to pieces. If you remove me from the review cycle, I won't go back to writing procedures.

The publications that were replaced by a neural network?

You could do small random variations (evolving the network) and assign some score to them (either number of sales, or scores by the staff)

Hopefully their editors are operating on principles other than increasing sales.

>How can a neural network evolve on its own.

Same way as humans? Via training on new inputs?

> Via training on new inputs?

There are no new inputs in this case because we got replaced editors with NN.

Nothing stops us from feeding them the work of human editors in 2000000 other publications...

Or their own work, as ranked by readers (e.g. their past posts that had the more views and better engagement).

No. This is drastically missing the point of what The Economist (or almost any media org) is. It isn't trying to be an average from 20000000 other publications, or what an average of readers think, or what had good engagement stats (though obviously the latter is a good target for the application of analytical software). It's not a neutral voice, or some down-the-middle average of what readers liked. It's an entity with its own political goals.

It's not even vaguely similar to a person, so how would that work? ln this example, if you had a hypothetical AI that understood the UK class system + the political system + the press system + the way these interlocking relationships work from a particular elite PoV then sure, but that's magical thinking. The actual written style is a small part of that; the other things are the inputs.

>ln this example, if you had a hypothetical AI that understood the UK class system + the political system + the press system + the way these interlocking relationships work from a particular elite PoV then sure, but that's magical thinking.

It doesn't have to understand any of those in any "human" sense, where it would be able to articulate what it understood on a meta level, etc.

It would just have to reinforce the right patterns and fuzzy connections that lead to success in writing credible looking articles....

> It would just have to reinforce the right patterns and fuzzy connections that lead to success in writing _credible looking articles_

We're probably at a point where we can manufacture shallow fakes of the current writing style, and I'm sure techniques like this will become prevalent (because convenience), but I feel you're missing the point of what I'm saying. It does need to understand these things. If it doesn't, then what it is is a mechanical reproduction of a current style of writing: that's a completely different category of thing, it's not even close to similar.

Reinforcement learning.

Isn't that more or less what Grammarly Premium can do? I'm not a subscriber, but their site says it offers suggestions on tone and clarity.

I’d bet against this. Neural networks still don’t actually understand the text at all. I would imagine true understanding is required for this task. You can’t get to the moon climbing progressively higher trees.

Reply on news.ycombinator.com