Hacker News Reader

Comment by ggggtez | Hacker News Reader

> explains why the bot became racist in less than 24 hours

Eh, the connection is tenuous. Sure ai becomes biased by the users who use it the most, but then you really still have to stop and all yourself: why is it we are taking about conspiracy videos instead of how people spend millions of hours watching cat videos, or memes?

There is a lot to say about the topic, but the answer really boils down to the fact the recommender systems are just good at working with variables it has access to. I don't know of any AI that has access to: "this person murdered someone" as an input. So literally, how should an AI be expected to optimize against that?

Manually tweaking the AI based on things in the real world makes sense, I'd just wish the author didn't have have it and mention Tay which is not really the same issue at all. Tay is an instance of deliberate manipulation, but YouTube is a case of wrong incentives.

View on HN

> Tay is an instance of deliberate manipulation, but YouTube is a case of wrong incentives.

I think it near-certain that YouTube is both. Considering the obvious value of getting your video into as many people's recommendation streams as possible, there's surely a whole "SEO, but for YouTube" industry out there, working furiously to game the system.

Anecdotally (albeit not "an industry" per se), on a number of occasions I've observed... let's call them "radical activist groups", organising off-YouTube to flood a favoured video or channel with activity, with the explicit goal of causing YouTube to make it more visible to "the masses".

> Sure ai becomes biased by the users who use it the most

There's an easy fix. Exclude the hardcore users (the top 10% say) from your recommendation engine.

Hmm, regardless of whether that works, that makes me wonder if it would work for the vicious cycle that is "popular things getting more popular because they are already popular".

It would probably lead to some interesting delayed feedback effects where things that are nearly popular get recommended, push out the popular, then the displaced stuff gets pushed into the nearly popular zone, etc. Perhaps some popularity + cooldown system would work.

Of course, all of this goes against the short-term incentives of maximizing engagement, so there's no way YouTube will ever implement this (even though it would probably greatly improve diversification of content).

I'm excluding the most active users, not the most popular things, so I think it would behave quite naturally. It just wouldn't have input from the obsessives. I agree it's unlikely to happen though.

Ah, yes. I misread. Yours is fundamentally simpler (that is a compliment) and at face value it indeed seems like a good solution. Outliers distort data, after all.

The system I thought you suggested in my misreading seems complimentary to your solution though.

But why must it be an AI? There was a time when systems were designed to let let me decide what I want.

Edit: For example. YouTube could have a feature that would let me filter out phrases like “top 10” or “amazed” from results.

> why is it we are taking about conspiracy videos instead of how people spend millions of hours watching cat videos, or memes?

Because we're begging for oversight. It is clear reading comments and articles day after day. We also highlighted alcoholics, communists, druggies, terrorists, etc in the past. The vast majority of people that, say, flew airplanes without consequence didn't matter anymore. We take the ones that validate our assumptions, appeal to emotion, and get oversight. We, the meager, need protection from the bogeymen.

Reply on news.ycombinator.com