Hacker News new | past | comments | ask | show | jobs | submit login

I have been doing a lot of writing lately, short and longer essays. I have also written on-and-off for the last ten years.

Writing well is not easy.

I used to think I wrote well until I started reading better writers. The whole idea that we will just slap together a bunch of text using some “AI” and it will resonate with people is crazy to me. As an engineer whose dabbled with some AI and ML, and has even seen what GPT3 can do, we are as far away from being able to use AI to write a best seller as we are from landing on the Sun. Not saying it will never happen just that nothing we are building right now will be able to do that.

Good writing factors in human culture, sparks emotions in people, tells stories that resonate. I will keep going to sleep at night safely assuming that no AI will out-write humans anytime soon and you should too.




You are right. AI will not replace good writing. AI will replace typical writing – seo spam, news reports, ads, coding tutorials, etc

Good writing has insight. AI can’t do that yet. Most writing is filler lettuce. AI is great.


Most news articles don’t seem to be even proofread anymore! At least AI spells perfectly. It will at least likely soon replace the uncreative who/what/where/when news writing.


Only for papers that don't care about who/what/where/when, because GPT-3 itself doesn't really care about those questions.


But if you write such an article, you need to convery specific information - about the "who/what/where/when" of the news story. This is information that was not available to GPT-3 at the time of training (e.g. it doesn't know that Biden is the PotUS now) so someone will have to sit down and craft a prompt that contains this information and such that it causes GPT-3 to generate a piece that repeats this information. To do all that sounds like a lot more work than just writing the article by hand.


we are as far away from being able to use AI to write a best seller as we are from landing on the Sun.

Harlequin Romances, though, those little soft-core porn pamphlets for women found at supermarket checkouts, could probably be generated. There are about 4,000 of them, and they're written to a set formula - “Boy meets girl, boy loses girl on page 56, and, by page 180, the book would end with a marriage proposal.” Load up the training set and profit.


But that’s not good writing, it’s just junk remixed, what would be the point? And that could be procedurally be generated.

In general I don’t see utility of having AI write stuff for us to read, it’s not like we are running out of things to read. What I’m more interested in is getting an computer to memorize facts, understand the context and then subleties of a convetsarion to the extent that you could ask back questions and infer answers from that ingested knowledge we fed it. And I think we’re far from it though we’re making fast strides towards a different direction.


AI is pretty far away from those, too. The "window" mechanism that GPT-3 and similar methods use basically prevents it from staying on-topic over the course of a few paragraphs.


Well, it's not just women, I have read and enjoyed such writing. But you are correct, writing that lacks substance intended for people who expect nothing more. Sex is easy to write about. It doesn't take much to turn people on.


I used to read a lot of junk paperbacks - mostly horrors, but also stories about truck drivers, detectives and western stories. They were schematic to the extreme. Looking back, it was the same story over and over and over. I liked them back then. I used to read also a lot of repetitive sci-fi: those small start trek books.

I never read Harlequin, but it being repetitive is not something special. Back when people read a lot of books, we did not read James Joyce and Charles Dickens exclusively. I mean, spiderman comics, ninja turtles comics, they were all the same story repeated again and again too.

We read easy for fun books.


>I used to read a lot of junk paperbacks - mostly horrors, but also stories about truck drivers, detectives and western stories. They were schematic to the extreme.

He, ditto in Spain. In most cases the (same) authors would rehash a detective story as some sci-fi based short novel by just changing some devices and lore and call it done.

Well, in the end cyberpunk it's just futuristic noir.


Dan brown seems to do this. Da Vinci code, angels and demons, and the one about the NSA all have the same plot structure.


Altough my case was about really cheap short novels from the 70's, they almost were Pulp Fiction.

An infamous crime in Chicago with mafia and detectives -> some crime in Orion with spaceships, Martian troops and so on.

Altough some of them were fairly good, like one about a time travel paradox.


Just copycat Broken Sword so nobody notices anything.


Junk paperbacks about truck drivers sound interesting... Remember any titles?


Yes, when signing up for the beta I almost included 'eroticc but I thought better of it.


Very true.

GPT-3 (and everything like it) is the perfect tool for SEO spam and filling out college papers with blather if you're confident no person is actually reading it.

It's impressive. No, really, it's an incredible simulacrum. But it's also just a slightly more practical version of the infinite monkeys with typewriters thought experiment.


Computer generated text have tens of millions of people they can put out of a job, manipulate into emotions, phish for scams, etc. without needing to produce text like James Joyce.


I almost feel like the obscure to most effects of joyce may be easier for gpt3 to generate than other things an only joycian scholars of which there are few would be able to tell


> we are as far away from being able to use AI to write a best seller as we are from landing on the Sun

We might be as far away from both. But our velocity approaching those goals could not be more different. In the last 10 years we've made tremendous progress in terms of writing. GPT-3 is leaps and bounds better than GPT-2. I'm not convinced a GPT-7 couldn't be mostly responsible for a best seller.


This comment is correct.

The most interesting question that it raises is whether the comment was written by GPT-3. How would we ever know for certain?


It's simple : GPT-3 doesn't have anything it wants to communicate. If you've taken some message from the comment, it's either written directly by a human, or some human has looked through god knows how many GPT-3 outputs until they found one that actually sounds like their message.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: