Hacker News Reader

Comment by M2Ys4U | Hacker News Reader

> Numerous issues arose during longer conversations. Blender would sometimes respond with offensive language, and at other times it would make up facts altogether.

I mean, to be fair, I've had many conversations like that...

View on HN

That's what happened with Microsoft's Tay.

Is it possible to learn sentiment analysis from Reddit? If they had access to modmail they could determine what's offensive to individual subreddits or groups, but I'm not sure if there's a way to gauge fiery reactions without that.

Maybe you could bootstrap it with an existing sentiment analysis tool, but that could easily lead to Garbage In Garbage Out.

Tay went bad because of a different mechanism. In the case of Tay, trolls figured out a way to make it repeat back arbitrary strings, and used that to create seemingly offensive dialogue. In the case of this chatbot, the offensiveness is coming from the underlying training data.

It was a bit of both.

There was a post after Tay came out that argued that Tay's answer to "Is Ted Cruz the Zodiac Killer?" came from the training the data, because that was already a meme, and it came back with the quip within minutes of launch.

I'm still not convinced he isn't!

There is something similar to Tay on Facebook/Instagram but it's not as "offensive" so the problem with Tay must have been in the dataset or something. It's basically a bot attempting to be a social media influencer.

See: https://www.facebook.com/brit.o.tara/

So it's just your average Facebook exchange?

Facebook is worse. On Reddit you get insulted by an anonymous account with a novelty name. On Facebook you discover how crazy your former friends and relatives can be.

> > Blender would sometimes respond with offensive language, and at other times it would make up facts altogether.

So EXACTLY like reddit, then.

What did they expect training it on Reddit posts?

Seems to me the bot is working fine.

Risky chat.

It’s one of the reasons I refuse to open Reddit anymore, 95% of the comments are insults, made up facts, misinformation or people living in their own bubble and refusing to accept any opinion other than their own.

The remaining 5% are the same jokes and memes repeated ad nauseam.

There are some nice communities out there, but they’re usually well hidden and relatively small. Once a community grows, it turns into the same shit show as the rest of Reddit.

Reddit is fantastic for learning a niche skill.

Want to learn how to make Kombucha, lock picking, or 3D printing? Then you probably want Reddit.

But anything approaching the popularity of a moderately successful video game turns into a shit show. One of the free to play games I use to play actually had two subreddits, after a community schism. Bonkers.

What percent is made up statistics?

Which subreddits specifically? I have a largely different experience on Reddit.

>and at other times it would make up facts altogether

Isn't that the entire point of an AI?

Reply on news.ycombinator.com