The current prompt project is too much like divination, and communicating with art AI is like a word game

Enter "Pac-Man game interface, Pac-Man, ghost, ink, blink, Clyde, Pac-Maze, Pac-Man, Mondrian style, modern art" to the AI ​​painting tool Midjourney , the picture obtained after "modernism bloomed".

Isn’t the input phrase of “prompt project” interesting?

When you enter a text prompt into an AI drawing tool such as DALL-E or Midtravel to have it generate a picture, or ask Copilot, an AI tool that automatically generates code, to write some software, they get The result can be called a work of art.

We can call this process "engineering", which sounds precise and logical. But if you go to the Discord platform and look at the prompts people type into the Midjourney app, you'll see something like this:

galaxy arising from a brain, 8k, octane render, micro detailed — upbeta — test — creative

my teeth are yellow, hello world :: would you like me a little better if they were white like yours — s 5000 — q 2 — upbeta — v 3

hg giger lovecraft nightmarish realm where monsters eternally reign terror​

chaos corrupted the once valor knight, transforming them into a powerful villian. Horns bursted from their heads, wing and tails grew from their sides, fingers and toes grew into claws. this is what does the void does. this is how life loses….

reason There must be a correct way to write prompts. The reality is that writing prompts often feels traceless. It's like when using a magic spell, you accidentally put the words in the spell in the wrong place. , it’s easy to mess things up.

To put it funny, writing prompts seems like humans trying to coax "an eager and confused pack animal" into doing work. We think it understands what we're saying, but the way it communicates is by yelling and running around.

What causes this phenomenon?

It can be said that this is a very strange moment in the history of artificial intelligence. For decades, artificial intelligence has advanced in the "shadow" of the Turing test (not always, but often), which holds that "intelligent" AI behaves and communicates in exactly the same way as intelligent humans.

According to Turing's ideas, for example, if an artificial life form can discuss current events, then it can be considered intelligent. In recent years, we've expanded this expectation of clear, precise, natural language into everyday devices: talking to Apple Siri and Amazon Alexa, asking about the weather or setting a timer.

But it is completely different from the artificial intelligence "dialogue" that produces works of art. We try to get them to create something. This means that if the AI ​​makes a mistake, the consequences are much more severe. No one cares if an online chatbot suddenly goes offline while chatting. It wouldn’t be a big deal if the chatbot wasn’t streaming the NBA live.

But what if we have a specific creative need that AI can satisfy? What if we want it to write a blog post with a specific content and style? We certainly need to make sure we can communicate with it correctly.

This means we have to start thinking about what AI is thinking, or rather, how it thinks. We must further develop what psychologists call machines’ “theory of mind.” “Sounds like fantasy, right?” As OpenAI co-founder Andrej Karpathy told me when talking about Copilot. "It's not something you're used to seeing. It's not like human theory of mind. It's like an alien artifact that emerged from a massive optimization process."

Andrej Karpathy

The author is not saying that these artificial intelligences are actually conscious, intelligent or anything else. They are just very subtle pattern recognizers and sequence completers, internally more like a chaotic ocean of mathematics.

But, because we give them commands with words, this puts us in a strange psychological relationship - trying to figure out what's going on inside.

The author is reminded of how the ancient Greeks interacted with the Delphic oracle. The Oracle of Delphi was believed to have knowledge of the past, present and future. The answers to questions can be weird because essentially it's like talking to a foreigner and who knows what results you'll get?

Communicating with artistic AI is like a word game

Scientists studying the inner workings of artistic robots have documented some of the strange internal states of these machines. Recently, two researchers at the University of Texas at Austin discovered that DALL-E 2 generated an apparent garbled phrase that appeared to have some consistent meaning within the model itself.

They noticed that the model generated the phrase "Apoploe vesrreitais," and when they fed it back to DALL-E 2 as a prompt, it drew birds. Similarly, receiving "Contarra ccetnxniams luryca tanniounons" will draw an insect or pest. Use "Wa ch zod ahakes rea" to create pictures of seafood.

Why is this? How did the model generate this strange new internal language? Scientists know nothing about this, although it appears to be an adversarial artifact of DALL-E 2's text encoder.

Similarly, prompt writing experts say that repeating phrases is a skill, as Michael Taylor writes in Prompt Engineering: From Words to Art.

Link: https://www.saxifrage.xyz/post/prompt-engineering

DALL-E 2. Midtravel or other AI art tools need to really capture important features when generating images, and simple repetition works surprisingly well here. Take this set of prompts as an example: "homer simpson, from the simpsons, eating a donut, homer simpson, homer simpson, homer simpson"

It feels like we need to hypnotize artificial intelligence to use It focuses on topics we care about. You can also see this in the large number of descriptive words that prompt writers typically use. Take a look at the image generated by Xe Iaso combined with stable diffusion:

I have to say that the picture is still a bit poetic. Communicating with the artistic AI feels like a word game - like playing Charades or Taboo, you have to trigger the AI ​​to generate the right results by having a conversation around a topic. Beyond that, the goal is to find the right incantation to awaken the spirits that inhabit that altar of intermediaries and summon them to do your bidding. As Xe said, "I'm not sure why people call prompt 'project'. I personally prefer to call it 'divination'."

Perhaps, we need to make some rigorous clarifications on the prompt generation model. Because it requires us to communicate in a completely insane way, it's unlikely to meet the requirements of the Turing Test and is not intellectually "like" us. The author firmly believes that one day artistic AI will be like us! But now, they're really, really weird.

The above is the detailed content of The current prompt project is too much like divination, and communicating with art AI is like a word game. For more information, please follow other related articles on the PHP Chinese website!