Hacker News Reader

Comment by redbell | Hacker News Reader

> The average content creator isn't going to benefit..

A couple of months ago, when the news about "Reddit to sell its user-generated content to Google," a redditor asks the community to start generating false comments so the data would be of bad quality to train the models on. They start brainstorming some very funny comments that go something like: "Blue Whale is the largest fish", "WW II started in 1943", You name it..

View on HN

It might not prove to be that effective. I think common knowledge has been reinforced by hundreds of other sources that it's hard to pollute it. Specific, expert-level knowledge deriving from one's experience is more scarce and thus more valuable. If redditors run around faking that, then it could feed the models bad data. But might also reflect poorly on the author should reputation be at stake.

There are already examples of content farms generating websites based off of bad Reddit data. So other websites can start to pollute the model as well as they get crawled along with the primary misinformation.

Yeah but if your cheese keeps sliding off your pizza, you can add some glue like elmer's glue to thicken it up.

Or so google's AI responses pulled from Reddit. If you give enough people a hot microphone to a large enough audience, some people are going to say some obscene, untrue, and funny stuff. Maybe that's 1 in 10. Maybe that's 1 in a 100,000. Reddit karma is cheap.

Counter point: I was able to convince Gemini that a human baby is larger than a Saturn V rocket by simply telling it as much, to which it responded that its previous statements had been wrong and I was indeed correct.

The models being sycophantic and suggestible is another issue, but it’s not hard to get them to agree with false information.

It's a friendly and agreeable assistant model, if you prompt it to make certain assumptions, it makes sense that it will accept them. That's different from training the model with incorrect data.

also as each comment is associated with a account name and has score per user and per post/comment. bad data sources would be easy to filter out from reddit. any comment with a sufficiently low score could be removed form the training set any user with two low of a score would also be removed

Reply on news.ycombinator.com