Show HN: I co-wrote a novel with GPT-3 in 118 hours | Hacker News Reader

Show HN: I co-wrote a novel with GPT-3 in 118 hours

docs.google.com

55 pointsggillas38 comments

Have asked this before, but GPT-3 users seem tight lipped about it. Let’s try you:

How did you get accesss to GPT-3?

I’m afraid that the way OpenAI is flicking aside applications for access from mere humans, and instead granting access to large corporate entities, tells us something about how AIs will deal with humans in the future.

I would also like to know this. This is exactly the kind of AI-assisted stuff I want to do, but no matter what I try, I just can't seem to get any information about how to actually get access.

Use AI Dungeon. It's the same thing just wrapped up into a webpage. AI Dungeon tries to steer you into using GPT-3 for dungeon crawling but you can edit the prompts, text, responses, anything. I used it to simulate a paper on curing Covid.

Greg brockman has said before on HN that if you have a cool use case you can email him and he’ll see what he can do.

I think it's more that they're trying to avoid association with the most obvious use case for GPT3. Generating spam, fake content, fake reviews, fake news.

It's whole purpose is generating unlimited amounts of believable fiction. Without tight control the internet would be overrun with GPT generated trash. Bans would be called for, just like what's happening with deepfakes. That's bad for business

They are trying to keep a genie in the bottle then!

And for these wonderful few months, they have succeeded. Eventually someone will train a similarly sized model and release it, but the costs involved may well be high enough for long enough that the common spammer won't be able to for a little bit.

> How did you get access

Well judging by recent posts you get it by cold emailing the CEO repeatedly, writing blogspam articles of zero worth and making YouTube videos with flip charts. At this stage GPT-3 founders seem more concerned with marketing themselves as an exclusive club than actually producing anything of note.

https://www.flowrite.com/blog/how-we-got-access-to-gpt-3-in-...

> Described as "The best access request" by Sam Altman

Cringeworthy. Yet to see anyone with access actually build a product with value, and that's before pricing kicks in.

Expect more hype, more clicks and then an exit.

The name “OpenAI” is downright Orwellian in its hypocrisy. What are they even hoping to achieve at this point? They should just change the name.

Do initial research as a non-profit. Much cheaper, all kinds of tax breaks, funders can write off "donations". Build something to make money, switch to regular corporation, profit.

You could easily accomplish what the OP did by just using AI Dungeon which allows for collaborative writing with GPT-3. I've used it to write erotica.

AI Dungeon is still free to use, just with a cost to use their better model.

I cold emailed the CEO with the description of the project I wanted to do. Had no prior connection, got access within two days.

I didn't get direct access. I've requested it. But everything I did here was done in AI Dungeon and mainly, Shortly Read.

Reminds of an episode of the programm (audio series). [1] They used GTP-3 to generate most of the answers the AI gives in that episode.

It is generally a fascinating series, set in a fictional future, where the program took control of the world. But it is less about that future, but more of a look back on today's world from fictional future.

[1] https://programaudioseries.com/14-more-parrot-than-predator/

When reading novels, I usually find myself skipping the portions that setup the atmosphere or looks of characters. I feel like a lot of it is forced. The latest book I read was Gideon the Ninth, and I found myself doing that quite a bit. I am more interested in the story and interaction between characters. Of course a bit of setting is indisposible such as where, when, rules of the universe, tech, etc. but a lot of the minute details I find don't matter.

If I were to write a novel, I would probably find it really mundane to flesh things out this way and perhaps GPT would be an ideal companion to fill in the details of the sketch while I focused on the main plot and interactions. Or maybe I would aim for readers who wouldn't mind reading a novel without all the fluff.

As a reader, I want to immerse myself in the world that the author creates for me. It's a communication between the author, and me, the reader.

If an author were to outsource this communication to an algorithm, I'd never read anything from this author again.

It's betrayal, just like outsourcing a phone call to your mother to GPT-3.

It's a waste of everyone's time: The authors time to set up and train the rubbish generation, the CPU cycles and energy wasted, and the reader's time.

Writing is a fine art, and captivating the reader is hard work. The fact that you get bored when reading work by authors that set up the scenery thoroughly might be a sign that those authors stuck to some kind of template for story drafting. There are tutorials and bootcamps for novelists, just like we have coding schools and Create-React-App.

If you get bored by texts created based on templated story layouts, imagine how readers would feel being fed GPT-3 nonsense.

I mostly agree with you except that the experience of every reader is different. What we find boring because of our over-exposure to similar themes might be exciting for someone who is encountering them for the first time. So if those readers are your targets, then I don't see anything wrong with GPT-3 aiding us.

In any case, do you mind recommending me some novels that you find exciting and do a good job of describing character and other details without being boring or generic?

You can feel the gpt fluff in between the plot. It’s unbearable. Beginnings of good prose that ramble off into word associations and nonsense content that reads like real sentenses. The computer clearly has no idea of the plot.

Superfun. Two Dutch podcasters had an episode[1] made using GPT-3, based on previous episodes, as an experiment. After the usual introduction they read up the generated text. It soon becomes illogical and incoherent but it took me 10 mins to figure out something was wrong. In the final few minutes there's some outtakes of them cracking up as they have to read the nonsense. Dutch only.

[1] https://art19.com/shows/een-podcast-over-media/episodes/fe6d...

> Peter has adopted the code and molded it for years, Anakin to a digital C3PO. He'd grown attached to his creation. He loved Art more than anyone else alive.

True gold.

Thanks for reading! Edits, updates, and full length version will be posted soon. Email ggillas@protonmail.com to subscribe.

Artwork generated from Artbreeder.com GPT-3 through Shortlyread.com

How long is the book? Is 118 hours short? What percentage of the content is from GPT-3? How extensively was it edited? Did you use a single prompt for the whole book, one per chapter, per section? How did it keep characters straight for an entire book length?

Without some of this information, it’s both impossible to tell what significance this has and less interesting, and just feels like you want me to sign up to a mailing list just because of the GPT-3 keyword.

Great questions. The book came in at 55,000 words.

I've competed in Nanowrimo several years and could never hit the 50k word (200 pages) benchmark during the 30 days. This took me only ~3hrs a day, with a few longer sprints on weekends.

I updated the prompt every break and made edits. The book is ~90% GPT written, with edits, that will drop to ~70% by word count.

The interesting part for me was that the prompt needed to be updated frequently. I also would rewrite opening texts to cram in the scenes characters and context, so that it could be carried through.

That's so cool! Are there any services that allow you to train the GPT on own your knowledge repository.

For example, if you wanted to write a research report but want the GPT work off an information corpus that you created. For example, to write a current month's summary on Tesla, you want the GPT to use all articles you archived over the previous month.

I still think the output is inferior to automatic paper generators from 2005.

This is literally just stringing words and putting random events together.

Can you please try to sell it on Amazon and tell us how that goes?

I wonder if you wrote a few seed pages per chapter on a contentious topic like ‘Why Bill Gates is funding the Coronavirus vaccine - and the answer is not good’ and released that book to a certain demo what your sales would be.

Ha, if I do I'll post an update.

Yes, I agree, there's a lot of potential here to generate controversial and fairly fake content.

I've only spent a month playing with gpt2 and gpt3, and technical literature, historical, and political content all seemed very viable for mass production with AI.

How do you get access to those apis? Also, how much would a freelance editor charge to edit it? I’m just fascinated with going all the way with this unethical idea lol.

Next time do the same, like he does, don't give him what he is looking for, like comments.

I don't have API access. I've applied and am still waiting.

I used the tools I posted above to get GPT-3 output. Editors can range from Fivver prices to $50k for a real novel.

Do you think if you wrote it slower it would be a little less weird?

Eel is the weirdest thing to eat with strangers. Her biology was like a movie. These things are brilliant in a Delillo like way but the flow is too fast. Also making a butterfly thing?

Absolutely.

As mentioned below, this is the first draft release. I loved some of the absurd things and re-rinsed when it got too bizarre.

Could not bring myself to read it. Thank for the excerpt :)

https://gimmeserendipity.com/mbtimodel/text/

Personality Probability

INTJ 0.239249

INFP 0.237849

Let me know when they can predict The Big Five

IBM already has this capability, and anecdotally, it seems quite accurate given 1-2k words of text.

Sorry for the lame question. But as a Marketer, how can I learn to build such products using GPT-3. Is there any way how to make it?

How many words did you have to input to obtain this novel?

Reply on news.ycombinator.com