Hacker News Reader

Comment by rckoepke | Hacker News Reader

Indeed, that seems to be from at least as early as 2013: https://www.facebook.com/dan.ariely/posts/904383595868

It is said that GPT-3's parameter space is enough to encode/memorize nearly 1/3rd of it's training corpus[0] as pointed out by 'GIFtheory in another GPT-3 thread here on HN. It seems you're finding the effects of that.

Additionally I'm curious about the recursive effect of this over-fitting after more and more output from GPT-n is published on the internet and inevitably gets included in the training corpus for GPT-(n+m) as pointed out by 'jobigoud[1]. Especially as people start using GPT-like models to spam the internet. We may lack a ground-truth corpus in the future to label "human".

It would be a bit like how carbon dating or production of low background steel changed after 1945 due to nuclear testing.

0: https://lambdalabs.com/blog/demystifying-gpt-3/

1: https://news.ycombinator.com/item?id=23887405

View on HN

It's funny, ever since the most recent thread about low background steel the other day it seems to be popping up with some frequency. I was aware of it before, so I'm not sure this is just a case of Baader-Meinhof (when you learn of something and then "start" to hear about it all the time, but really you're just now paying attention to it)

Edit: actually, I think I saw this tweet: https://twitter.com/rantlab/status/1284849214653034497, and remember thinking "I bet this person just read the hn thread about low backround steel". It doesn't seem to have come up on hn other than that since the low background steel post.

It's a really fun feeling when you notice that someone posted something because they read the same thing you read and it sparked a similar association.

There are probably already manipulation techniques in play where an actor 'plants' (or 'incept' if you will) a future post by posting a lot of things that will lead people to organically 'find' the subject they want to promote.

> It's a really fun feeling when you notice that someone posted something because they read the same thing you read and it sparked a similar association.

Someone please tell us the German word for that. There must be one. (Or maybe a French phrase.) ;-)

Probably something like: assoziationmanipulation

In my case I definitely only included it as an example due to the recent submission on HN, so not Baader-Meinhof in this case. I'm still surprised that @rantlab and myself came up with the exact same analogy independently though.

Clearly it wasn't as creative of an analogy as I originally thought. For myself, just a very small leap of logic after reading 'jobigoud's comment. Very surprised to see other people making it as well, and a good re-calibration for me!

I also came up with the same analogy after reading both posts. Wasn't just lurking and when I later had a moment to post my brilliant analogy it was already posted. So if many of us thought of this independently, does that make it a better or worse analogy?

Probably better, but I guess it depends. I suppose one of two cases would apply to analogies that people come up with independently (although 'polytely explores some of the non-independence of this particular case):

1) The analogy has more symmetries between the left and right side of the analogy than most analogies, allowing people to arrive at it from many different perspectives.

2) A lack of options for the analogy - maybe there just aren't that many good examples with any real symmetries at all, so everyone who is inclined to create analogies is forced to the same one, which may or may not be particularly quality.

I'd hypothesize overall (1) will generally dominate (2) just because if an analogy is poor, I'd expect few people to come up with it even if it's the only possible analogy...it wouldn't pass the "impact"/"relevance" threshold needed to spend effort converting to a tangible form for communication.

I also think (2) may be a false mechanism given the incredible breadth of experiences everyone has. There might not be a lack for analogy targets for almost every conceivable topic.

Whereas (1) feels quite a bit more defensible overall.

That said, we're in a bubble here as 'tlarkworthy points out, so I wouldn't use this as an example.

There are so many discourse fads, especially on HN, like the time there was a lot of logical fallacy taxonomy. The echo chamber effect is very real. Even the use of echo chamber is and echo chamber. I guess its unavoidable that memes are a very real method we use for group cognition. I also noticed the background steel thing came up a lot recently.

It'd be interesting to compare plagiarism detector scores of average outputs of various generations of GPT.

Memorizing a web corpus and answering questions about it as well as Google would be an interesting result.

Well, it's not close to that, but it's close enough to be amusing. Here are some questions it answered in Q&A:

https://tildes.net/~games/qmc/ai_dungeon_dragon_model_upgrad...

But I think for factual knowledge, it will need to be better about explaining where it got the information.

Reply on news.ycombinator.com