Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

The computers are getting better at writing

newyorker.com

Post reply on HN
Print view View original

31–40 of 165 posts

Re: The computers are getting better at writing

#31

Post by minimaxir »

See also my experiments with GPT-3 on sane prompts, which have wildly varying quality even after generating them in bulk: https://github.com/minimaxir/gpt-3-experiments

Creative writing hasn't been one of the super-hyped use cases by OpenAI for the OpenAI API outside of AI Dungeon, surprisingly. For just random generation, the necessary curation can detract from the time-savings advantages. (as an aside, the API is also extremely expensive for long-form content to the point I'm not sure how the economics work for these startups even with charging monthly fees).

I'm more bullish on small bespoke models for a given use case, which is what I spend my time researching. (example: a ~1M parameter GPT-2 Magic: the Gathering card generator: https://colab.research.google.com/drive/1VOt090UzvltoBgMdUZm... )

Re: The computers are getting better at writing

#32

Post by groby_b »

amilios wrote: post #30

It's clear that the system occasionally struggles in making coherent sentences that communicate actual meaning -- others have pointed this out well enough. I think what's occasionally lost in these types of criticisms of these systems is the point the article is making about the transformation of the role of the writer from a producer to a curator. Already this transformation has occured in other fields; the job of a…

> the job of a professional translator (for text) at this point is mostly starting with a machine translation and slightly tweaking it in order to produce something more natural.

Not really, no. That might be true for instruction booklet translation (they sure read like it), but no, for a decent translation, you'll need to let a human have a full go at it.

Writing is as much about filtering & tweaking GPT-3 output as sculpting is about "removing everything that doesn't look like the final sculpture", as the old adage goes.

It is useful for one purpose: To do an absolutely minimal job and to externalize some more work onto end users. It'll also be popular for spam^W SEO, and for writing filler content for content graveyards. It will not, however, be useful for writing.

I have no doubt that we'll see a lot more tooling supporting writing, and machine-assisted writing. I have equally no doubt that GPT-3 and friends aren't going to be it.

Re: The computers are getting better at writing

#33

Post by Animats »

GPT-3 is already better than most blog comments, many bloggers, and most online text porn. Maybe most ad copy. Many pundits. Some college professors.

It has no clue what it's talking about, of course. It's terrible at fact-based material, because what comes out looks plausible, but contains made-up facts. Some facts will be correct, because they were mentioned a few times in the training database. Others are just plausible words strung together.

What it really demonstrates is that much of literature has less behind it than was previously supposed.

Re: The computers are getting better at writing

#34

Post by FeepingCreature »

Animats wrote: post #33

GPT-3 is already better than most blog comments, many bloggers, and most online text porn. Maybe most ad copy. Many pundits. Some college professors. It has no clue what it's talking about, of course. It's terrible at fact-based material, because what comes out looks plausible, but contains made-up facts. Some facts will be correct, because they were mentioned a few times in the training database. Others are just pla…

I expect GPT-3 to be amazing for people who want to write stories or articles but are really terrible at one specific aspect of the writing process.

Re: The computers are getting better at writing

#35

Post by Animats »

lbacaj wrote: post #3

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 G…

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.

Re: The computers are getting better at writing

#36

Post by JamesBarney »

lbacaj wrote: post #3

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 G…

> 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.

Re: The computers are getting better at writing

#37

Post by barry-cotter »

amilios wrote: post #30

It's clear that the system occasionally struggles in making coherent sentences that communicate actual meaning -- others have pointed this out well enough. I think what's occasionally lost in these types of criticisms of these systems is the point the article is making about the transformation of the role of the writer from a producer to a curator. Already this transformation has occured in other fields; the job of a…

> Already this transformation has occured in other fields; the job of a professional translator (for text) at this point is mostly starting with a machine translation and slightly tweaking it in order to produce something more natural.

That certainly isn’t the case for Chinese-English translation. It’s faster to translate from scratch than to edit machine translated text if you know what your doing and have quality standards. One of my friends did German-English translation and machine translation just isn’t for purpose there either if you have any real quality standards. You can’t skip the part where you know what’s going on.

Re: The computers are getting better at writing

#38

Post by tsimionescu »

Teever wrote: post #21
lbacaj wrote: post #3

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 G…

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.

Re: The computers are getting better at writing

#39

Post by Clewza313 »

groby_b wrote: post #32
amilios wrote: post #30

It's clear that the system occasionally struggles in making coherent sentences that communicate actual meaning -- others have pointed this out well enough. I think what's occasionally lost in these types of criticisms of these systems is the point the article is making about the transformation of the role of the writer from a producer to a curator. Already this transformation has occured in other fields; the job of a…

> the job of a professional translator (for text) at this point is mostly starting with a machine translation and slightly tweaking it in order to produce something more natural. Not really, no. That might be true for instruction booklet translation (they sure read like it), but no, for a decent translation, you'll need to let a human have a full go at it. Writing is as much about filtering & tweaking GPT-3 output as…

Sure, "decent" translation as in novels etc will continue to be hand-crafted, but the vast majority of translation is low-end bulk stuff like business reports and technical documentation, paid pennies per word. I know one (former) professional translator, and dumping text into Google Translate was always her starting point.

Interestingly, her job got harder when Google rolled out the new deep learning models, because instead of outputting visibly broken garbage that kind of flagged itself for checking, it now tends to produce beautifully formed sentences with subtle but often critical fuckups (wrong sense of a word, missing negation, parsing a pronoun incorrectly, etc) that can affect meaning.

Re: The computers are getting better at writing

#40

Post by Clewza313 »

TheDudeMan wrote: post #18

"teraflops per second"

Teraflops per second is the speed at which computing speeds up. Eyeballing Wikipedia's list of fastest supercomputers, between 2018 and 2020 we clocked around 100 petaflops per year, which comes to around 0.00317 teraflops per second.

For homework, you can calculate teraflops per second per second, which would measure the acceleration of teraflops.

Post reply on HN