Shulou(Shulou.com)11/24 Report--

AI is more and more indispensable in our life. Obviously, whoever can better master the prompt project will be able to master a powerful weapon more than others.

The popularity of ChatGPT has also led to a rise in the salaries of prompt engineers.

Now, some people are earning an annual salary of $250000 to $330000.

An important skill of the 21st century: reminding the project at a breakneck pace, the artificial intelligence arms race has begun.

Google released Bard, a search chat robot based on a large-scale language model LaMDA, and Microsoft unveiled a next-generation Edge browser driven by ChatGPT and a new search engine, Bing Bing.

Although big technology companies have been quietly developing their own generative artificial intelligence tools for years, since OpenAI's ChatGPT became popular, these giants have been scrambling to show off their abilities to the public.

The technology of generative AI is also thought to be destined to be a substitute for creative or repetitive work, from search engines to storywriting, scientific research, homework, schooling and so on. In this context, it will reshape the world.

If this paradigm shift in artificial intelligence does come, then an important skill in the 21st century is likely to be how to effectively talk to machines, or "cue engineering".

Image generation models like DALL-E 2 and Midjourney, as well as text generation tools like ChatGPT, will be a means of creation. But the prerequisite for successful use of them is that we first understand how to guide AI to achieve the desired results.

For example, asking ChatGPT to write a five-paragraph book report on "animal farms" may produce impressive but inaccurate results.

But if you give AI a clear hint: "write a college-level five-paragraph book report in elegant prose, drawing on the history of the satirical fable novel Animal Farm." Referring to Orwell's Why I write and explaining the author's choice of style in the novel, we can produce a more mature and persuasive output.

It can be said that good tips are not only specific, but also reflect a deeper understanding of what you want AI to do.

New career: prompt engineer Charlie Warzel, a well-known American writer, shared the story of how he and his 8-year-old nephew experienced Midjourney together last summer.

They originally wanted to create a pleasant scene: "A flea surfing in the tsunami wave, fighting a giant wasp."

But as it turned out, the flea didn't look cartoon enough at first, but adjusted by the author, it was too surreal, completely beyond the reach of his nephew's age.

The writer says that although he writes for a living, it is clear that the description he asked for in mentoring AI is too vague, at least for AI.

However, it doesn't matter if you don't know how to talk to AI, because now there's a whole new industry that can help you play with AIGC-- 's Tip Store.

For example, on the website PromptBase, you can buy a prompt that allows AI to steadily output a topic at a very low price.

Website address: https://promptbase.com/ is here, from DALL-E to Midjourney to Stable Diffusion, there are all kinds of tips.

One of the most popular prompt on PromptBase is the cute 3D emoji generated in clay style with DALL-E 2.

Or for $2.99, Stable Diffusion can infinitely generate movie heroines with similar styles.

PromptBase's sixth-ranked seller, a Spaniard nicknamed Imagineer, said Tip Engineering is only a sideline, after all, he has made just over 800 euros since September. "to me, it's more like a game. These prompt are my little babies. "

Imagineer's cue writing process is based on knowledge of design, illustration and photography.

When asked why he thought he was good at writing prompts, Imagineer thought it was a combination of natural skills and strategies. "I realized that I was better at talking to Midjourney than anyone else," he said. And in Midjourney, experience can be taken, and less effort can produce great results than DALL-E 2 and Stable Diffusion. "

The most critical factor is iteration, says Imagineer. When you produce a large number of images and change some words or parameters, a good hint "will produce consistent and predictable results."

In the prompt project, the mastery of professional knowledge in related fields is also essential.

Dan Shipper, an entrepreneur and writer, has been using ChatGPT to help write blog posts about the future of artificial intelligence tools since ChatGPT was released in November.

When he needs to describe a concept (for example, to write an article about the unpopular cryptocurrency CEO Sam Bankman-Fried, he needs to use the philosophy of effective altruism), he asks ChatGPT to sum up the key points of effective altruism for the average reader in a few words.

After ChatGPT outputs the text, Shipper checks the accuracy of the text and uses his own rhetorical devices to make the sentence more beautiful.

ChatGPT allows me to skip one step, but only if I know what I'm talking about, so I can write a good tip and then check the facts of the output, says Shipper.

Shipper likens the job of reminding ChatGPT to managing a smart and enthusiastic junior employee.

ChatGPT is enthusiastic, skilled, but also inexperienced, so it is more likely to make minor but critical mistakes. When it doesn't know the answer, it is very good at talking nonsense.

Shipper concluded that taste and experience are necessary conditions for creating a successful prompt.

For example, he will ask ChatGPT to write an impressive and thorough outline for a long article he is writing.

"I wrote a bunch of key points and told ChatGPT: these are all kinds of quotes, ideas and phrases that I have accumulated. Then I wrote: can you write it as an outline of a paper? Obviously, the more work Shipper does on fine-tuning tips, the better the output will be.

Sometimes, reminding that writing itself gives people a special joy.

Meg Conley, a writer who uses AI tools in her spare time, believes that prompt engineering is a challenge, similar to writing a persuasive article.

It is suggested that the project is very difficult and often fails. "when words look a bit like the world in your head, there is pure happiness," she wrote on Twitter in November.

But Conley's experience is special. She suffers from heart blindness, so it is difficult for her to imagine images. After the release of Midjourney, she often stayed up late to describe what she imagined, polishing the prompt until the image became correct.

In Conley's view, the most important thing is to understand the model in front of you. Each tool is built and trained in a different way, so they have unique aesthetics and dialects, just as people who share a language have dialects and cultural eccentricities.

"just as prose writing is different from academic writing, different listeners will have different ways of expressing themselves. "

"I've seen people who are really good at DALL-E 2, and it seems that the stronger the ability to mix and match references and vulgar culture, the better they are. But the way I think about the world is more like working along the Midjourney model. "

Over time, Conley has become very familiar with the sequence of Midjourney operations.

In Midjourney, if you type the word girl before the adjective red, it will focus more on girls than on red.

When there is a longer hint, it is like a puzzle, you need to solve the puzzle and know which terms should be given more attention.

Smart teachers are already training students to use ChatGPT. Although many teachers regard ChatGPT as a scourge, some enlightened professors have realized that learning to use prompt is a necessary skill for students in their academic life.

Ethan Mollick, a professor at the University of Pennsylvania who teaches a course on Innovation and Entrepreneurship, revised the syllabus after ChatGPT was released.

In one of his new courses, Mollick asked his students to think of ChatGPT as a student and teach the robot through prompt to write an article about specific course concepts.

In the process, students must, like professors, help ChatGPT improve the article until it seems to have fully mastered the topic.

Mollick hopes that this exercise will help students learn through explanation and help them learn to write clever prompts.

In Mollick's view, cue engineering is a zone between linguistics and problem solving.

"hints are like programming in prose, producing strange and random results. And good tips reward people with divergent thinking who can find experimental methods quickly. It rewards people with deep curiosity. "

In addition, ChatGPT will reward deeper technical knowledge.

One striking image on Midjourney's Discord server is that it was generated with the painfully detailed prompt below.

Metaphysical style, using colored pencils to draw smooth shadows and mixtures, sunrise shining in the deep mountains and forests of the pond, willow branches hanging on the edge of the pond, moody, strong emotion, deep perspective, natural lighting, ultra-fine, super high contrast, complex details, photography, ray tracing, octane render Illusory engine-ar 3:2-s 999-chaos 50-v 4-v 4-

In fact, this process not only requires concreteness, good tips often show awareness of the user's ability to replicate media.

Octane Render and Unreal engine are digital animation tools for making 3D graphics.

Inputs like "--v4" instruct Midjourney to use a newer, more powerful, experimental version of the model.

Why can some tips produce realistic images? Because they will require models to imitate specific cameras or lenses, or use some knowledge of art history or specific art styles.

This is reminiscent of the early days of search, when experts who could navigate Boolean operators, authors, keywords, sources, and date ranges could unlock better results.

Of course, today's search engines have long been less demanding.

To cater to the masses, Google's tools have become more powerful, allowing users to get high-quality results through simple or even clumsy queries.

In Mollick's view, cue engineering is just a basic way for us to interact with artificial intelligence, and our goal is to get them to synthesize what we want through naked cues or other unknown means.

He is probably right.

Some people who have seen the early testing of Open AI's undisclosed GPT-4 model say it is like science fiction, and when it comes out, all the old tools will be out of date.

The Microsoft press conference has shown us the possibility that chatbots will talk to us like personal assistants.

You don't need to search for the size of the Honda Odyssey and the size of the Ikea Klippan sofa. You can ask it directly: can the Ikea Klippan sofa fit into my Honda Odyssey in 2019?

Real-time engineering may just be a bridge to enter a brave new world, no matter what kind of generated media landscape comes next, but it's hard to know to what extent this is hype.

Now, Warzel says, before the paradigm shift, his family was still attracted by artificial intelligence cues.

In his view, these tips are usually more intriguing than their output.

When people share art or text generated by artificial intelligence, they often share prompt together.

Shuttling back and forth between instructions and the final result makes sense, and even makes people feel intimate.

It's kind of like allowing you to get into ta's brain to see how ta puts different knowledge points together, how to deduce a problem, and how to use ta's creativity to produce something unexpected.

Like writing and coding, cue engineering is a new form of thinking.

It is between dialogue and query, between programming and prose.

It is part of this fast-changing and uncertain future that feels very "human".

Reference:

Https://www.theatlantic.com/technology/archive/2023/02/openai-text-models-google-search-engine-bard-chatbot-chatgpt-prompt-writing/672991/

This article comes from the official account of Wechat: Xin Zhiyuan (ID:AI_era)

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.