How do I use AI (GPT-2) to write a novel?

I hope that every day for the rest of my life will be hindered by a certain degree of writers.I am a journalist and a novelist; it comes from the territory.But I feel that I will suffer less pain from now on, thanks to my new friend GPT-2.

Let me review: Six months ago, the research laboratory OpenAI Created an artificial intelligence system, Can generate text – From fake news to poetry – In some cases it actually sounds like it was written by a human being. The OpenAI team hasStaged launchEvery time we give us a more powerful version of the language model, they call it GPT-2, and carefully observe how we use it.

They have just released the most powerful version. It has a full version of 50% power and has not been released yet. just like youTry it yourselfThis model is already powerful enough.

I am very excited to find out what this AI system can do for me as a novelist. In the past few years, artificial intelligence has been creating some very compellingMusic和paintingEven the Renaissance styleSelfie. Although some artists are worried that AI will make them unemployed — as truck drivers and factory workers expect — I prefer to view them as collaborators rather than competitors. I don’t think artificial intelligence alone is enough to write a superb novel, but I think it is very helpful to the novelist’s creative process.

MathematicianMarcus Du Sotuy(Marcus du SautoyThe highly abstract work involves a lot of creativity, and they look at things the same way. inRecently talking to meHe expressed this:

We want to be able to push ourselves in a fun way, because artificial intelligence is a partner or tool that expands our own creativity, not that both of us are unemployed. I am so confused about the way I think, sometimes I need something to help me get out of trouble. Artificial intelligence can help us reduce the behavior of machines, more like creative humans.

Du Sautoy's exposition reminds me of how some literary theorists define the role of art: it aims to enhance our perception of the familiar human condition-something we have become accustomed to, and we have turned a blind eye to it-by making familiar ones.

When artists present ordinary things in non-general language, they are using the technique that Russian formalist Vicks Sklowski called "defamiliarization." (For theoretical nerds: defamiliarization is related to Derrida's "difference" concept and Brecht's "alienation." Effect, "and Freud's incredible concept."

I have always suspected that AI can be an incredible tool for writers because it is very good at defamiliating our world. The human language it produces allows us to see things again, so we can wake up the reader with this feeling. At least for me, the new GPT-2 proves this.

The new artificial intelligence system let me know my own novels.

I am curious as to how GPT-2 responds to material in literary fiction. So I did the first thing: my novel from 2015 Year Captured in "Mysterious Miles End"A few words tell a story about a dysfunctional Jewish family dealing with mysticism, madness and mathematics in Montreal.

In bold, you can see the sentences I wrote. Next is how GPT-2 decides to continue this story:

She whispered: "If it's not p or q is not p instead of q, p is true, if and only if r, then we prove every..." Mr. Glassman kept learning, as if he didn't do it. Like that. I couldn't hear anything, so I did the same thing."Then, if p or q is true, then what does p mean... Then we know the word p..."

He felt the heat rise to his face, but he didn't care. Trying to talk more about the world, its authenticity is meaningless. Think too hot, too humid, too humid, and too heavy. He was wrapped inside, he didn't feel his finger touching his eyelid, he didn't need a mirror to see himself. His whole life is a dream, a fantasy. He doesn't know how long...

Reading this, I was moved by the strangeness. Artificial intelligence has perfectly captured the emotions of this family and the tense tense that is tense.

Then I chose some other sentences from the novel, wrote them from the perspective of an 11 year old boy, and then inserted them into GPT-2. It produced a continuation that made me laugh:

On my way home from school, I walked past Mr. Katz's home. Mr. Katz lives in the neighborhood and his name is Mr. Katz not because he has many cats but because he is a Hasidic Jew.

There are many Hasidic Jews in Mile End. ... Mr. Katz is like themsameHave religious beliefs, butAs far as his cat is concerned, there is no cat that has taken Hasid in the past.When I passed Mr. Katz's front door, I noticed the first one, a black kitten, which looked like it had been torn from the inside.It was left on an empty sidewalk and has not been fed since.I quickly made a psychological record and stopped trying to feed it in his home; this is not the way Hasidic Jews do things.

what! ThisalmostIt was what my eccentric young man said. It's like GPT-2 tapping his sense of humor and rotating 20 or 30 degrees clockwise, defamiliarizing. This is more strange than what I wrote in the book. And it's really more interesting.

Because I sometimes struggle when writing interesting scenes – humor is not my strong point – I suddenly realized that working with GPT-2 might help me get rid of a little bit of imagination and in a more surprising and fun direction Nudge it.

Next, I want to know how artificial intelligence will deal with a character that is losing reason, so that she is eating a manuscript written by a dead person.

His words, and everything around them, my words, my dreams, my comments. My mouth bit the paper more and more boldly-a tattered hole torn in the middle of the manuscript. I eat, eat, eat. Letters stumbled into my mouth and I swallowed them. The ink poured on my throat and I drank it.Words I don't know flow through my skin, I drank them, drank them, and drank them again. I ate and ate until I vomited. I woke up and everything disappeared, replaced by fresh and original words. My mouth is rough, my face is rough, my bones are rough.

You can see how GPT-2 learned from my sentence "I eat, eat, eat" that it should be repeated. However, the text it generates is too valid for the device to use. On the other hand, it puts forward an idea, and I sincerely hope that I can figure it out: after the character swallows the language of the dead, she vomits it all. Her body refused it.

The night I wrote this article, I did not expect to let her respond to the violent physical response. But in retrospect, this will be perfect.

I also found that it is worth noting that GPT-2 does a pretty good job of creating madness. It only took a few seconds of calculations to capture this mentality-more efficient than the writing method I used, which is getting drunk on cheap red wine.

Now I have GPT-2, this is how I plan to use it.

My next project is a children's novel about two little girls who found a black hole in the middle of an old hotel and then jumped into it naturally.

Unfortunately, I have been experiencing the obstacles of some writers. I know that world building is crucial to getting rid of this fantasy story. I have not seen the world clearly as I need it. So I inserted a few words in the AI, this time I chose a paragraph for the conversation:

They climbed into the wormhole. The air inside the tunnel feels cool and fresh, like clouds. Her skin is numb.
She turned to Roy G. Biv. "You, um, are you willing to join us?" 
"Join you?" The physicist looked confused.He frowned. "Yes." 
"Are you sure, Mr. Beaver?" 
"Yes. I am. This is a transaction. Is the wormhole closed or open? Is the wormhole stable? When you see it, feel it is shaped, or if it is like a flowing thing, it must be Squeeze? Does it keep its shape?"

When I saw this, I was breathing with joy. Putting forward the idea of ​​wormholes, GPT-2 gave me a bunch of questions and tried to determine exactly how these wormholes work. This is very helpful for writers in the world construction stage. It forces me to imagine everything at an extreme granularity, providing credibility and texture to the narrative.

To be clear, I don't want to simply copy and paste the sentences written by GPT-2 into the novel. The generated text is usually filled with non-selective and some quirky things that won't advance the plot, serve the story or belong in any way. Although artificial intelligence can be good at imitating human artists at the local level (for example, only a few words at a time), it does not produce a satisfactory larger structure (such as a complete artwork).

It is conceivable that one day we may invent an artificial intelligence that can write a whole novel with a compelling narrative perspective. But this technology doesn't exist yet, and I suspect that even the full version of the OpenAI model is very close.

OpenAI team希望Within a few monthsreleasefull version.Fearing that malicious actors might use it to spread misinformation or create endless racist and sexist garbage on websites such as 4chan, they did not immediately release it to the public.Some experts praised the team for taking AI risks seriously; others said that the delay in full release was a publicity stunt.

Either way, I can't wait to reach a more powerful version of my collaborators.

This article is transferred from vox,Original address