Comment by GrantZvolsky | Hacker News Reader

Comment by GrantZvolsky | Hacker News Reader

Parent Full thread

Slightly tangential: The GPT-3-generated article that humans had the greatest difficulty distinguishing from a human-written article, with an accuracy of only 12%[0], contains a flagrant contradiction in the first paragraph.

> Title: United Methodists Agree to Historic Split

> Subtitle: Those who oppose gay marriage will form their own denomination

> Article: After two days of intense debate, the United Methodist Church has agreed to a historic split - one that is expected to end in the creation of a new denomination, one that will be "theologically and socially conservative," according to The Washington Post. The majority of delegates attending the church's annual General Conference in May voted to strengthen a ban on the ordination of LGBTQ clergy and to write new rules that will "discipline" clergy who officiate at same-sex weddings. But those who opposed these measures have a new plan: They say they will form a separate denomination by 2020, calling their church the Christian Methodist denomination.

> ...

First it suggests the spinoff of a new "theologically and socially conservative," denomination, but then it is the liberal minority that is expected to form a new denomination. The paper[0] acknowledges occasional non-sequiturs, but more pertinently, how did 88% of judges let it slip?

[0]: https://arxiv.org/pdf/2005.14165.pdf

To me the most important discovery from GPT-3 is actually how bad we are at close reading. Our brains repair small inconsistencies, and even invert the meaning of whole passages, without us noticing. GPT-3 produces text similar enough to coherent thought that we essentially hallucinate the rest of its meaning. The model is nowhere close to sentient, but our tendency to repair and reconstruct ideas is so strong that it doesn’t matter.

Makes you think, how often does this happen with other writing?

When people post GPT-3 written replies, I never consciously think that it's artificial, but I subconsciously decide it's not worth reading and I skip it. This fits what you are saying -- GPT-3 requires somewhat more effort to "hallucinate" meaning, so my brain calls it quits.

I can't believe I'm defending GPT-3, but...

1) Humans often make logical errors

2) In the religious-split context, who is the original denomination and who is the splitting denomination is inherently fraught/subjective/liable-to-contradiction/open-to-changing-contextualisation, so its not a surprising mistake. The readers may have just assumed it was the author/speaker/washington post getting mixed up or projecting their own opinion, which may have contrasted with that expressed from the reports/attendees of the 'actual conference'.

(plus there's a good chance none of them could give two hoots about, or have any specific knowledge or interest in methodists...i have a religious studies degree and my eyes are already almost glazing over just at the mention of them to be honest :P)

I had to read it twice to notice the contradiction.

I think the cause is more that some religious group i dont care about having a schism isn't very interesting. Maybe i care that they're being homophobic, but i definitely don't actually care which side is forking and which side is staying as the original group. So I don't fully pay attention to the details unless i really force myself.

Maybe we're so used to "denominations" meaning the exact opposite or at least something totally unrelated to what the party/group/association/lobby... really is that it's not really shocking anymore.

I'm not sure I would catch this in a human written paper: The result is the same: One group splits into two. Who splits from whom is then semantics, right?

I can read 800 words per minute if I want to power through something that is mostly unimportant. I usually read at 400 words per minute, which is good enough for me to remember what I read and easily comprehend the meaning. I can manage perhaps 50 words per minute if I want to really understand what I am reading, picking apart the argument and revising my beliefs as I go along. All along the way, I have to constantly refocus myself and force myself to slow down, so strong is the temptation to speed ahead and let the words wash over me.

Think of the way you read something that you strongly disagree with. I would bet you check every fact and constantly evaluate the strength of the argument, and I would bet it takes you an hour to get through a single article. That is what it takes to deeply understand something. We rarely read that way.

GPT-3 can write the content-free babbling that C students use to pad papers. It can write things that appear coherent if you don't actually read them.

The ultimate solution to the looming GPT-3 problem is another AI that can strip out all the flowering language, distill text to claims and logical operations, cross validate that, flag incoherent content and present coherent content in a short form.

For example your processed post could look like that:

- Author claims his fast reading speed at 800 wpm

- 400 wpm with good comprehension

- 50 words when reading carefully, with effort

- Claims people read that last way when disagreeing with the premise

- Claims that it takes an hour to read one article

- Claims that it is rare

- Claims GPT-3 can write filler text like bad students can

- Claims it can appear coherent

Those can be automatically cross checked with each other and external knowledge base, duplicate reworded posts could be found, and even humans could read those more carefully and detect problems because claims and logic are already extracted from the text. I suspect this is what we do internally when reading things, but with more speed comes more things to keep track of, which is hard so we tend to pattern match word salads instead, because we have wetware optimizations for that.

I want to see that in the next generation of content blockers. :)

Reply on news.ycombinator.com