The hottest and strangest new job in the technology industry: Just talking to AI can earn nearly 8 billion VND / year

Riley Goodside, 36 years old, now works at the startup Scale AI with the strangest job in the field of AI. He spends many hours every day talking to the GPT-3 model. Goodside’s job is to create and tailor the passages that people are likely to type into the AI ​​for the best results.

Unlike traditional programmers, “suggestion engineers” will feed into the AI ​​system commands written in plain text, which the system will then execute.

Goodside often tells this model that it is a good tool but not perfect, and to follow what he says. “ You are GPT‑3 and you can’t do math. Your memory is impressive but you tend to make up false information ,” Goodside entered the system during an AI training session.

According to the Washington Post, when Google, Microsoft or OpenAI released AI search and chat tools, they reversed the traditional way of interaction between humans and machines that had existed for decades. Users do not currently need to write code in languages ​​like Python or SQL to control the computer. The hottest new programming language right now is English,” Andrej Karpathy, former head of AI at Tesla, shared on Twitter.

Engineers like Goodside can take full advantage of what these AI tools can do: understand their strengths and weaknesses, and turn plain text input into the best results.

“AI has no background in practice, but it understands that: Any task can be done. All questions can be answered,” Goodside shared. The trick is to “build it on a premise, a story that can only be completed in one way”

Simon Willison, a British programmer who is studying suggestion engineering, said: “ I’ve been a software engineer for 20 years and everything is fixed. You write code and the computer does exactly what it says. But in reality, the language modelers don’t even know what the model is going to do .”

Willison added: “ There are many people who despise suggestion engineers and think they are paid just to type into the computer. But AI systems can deceive engineers and lead them into time-consuming things with no results .”

Karpathy compares “suggestion engineers” to psychologists for AI and the companies that hire them in hopes of uncovering new AI potentials.

Some experts argue that the suggestion engineer doesn’t really control the AI ​​either, because no one knows how the system will react. One question can generate dozens of conflicting answers, as ChatGPT is showing.

” Whatever is driving the behavior of the models in response to cues is not a deep understanding of the patterns ,” said Shane Steinert-Threlkeld, assistant professor of linguistics at the University of Washington. language. This is an indication that the computer’s response is not based on understanding, but rather crudely simulating speech to solve problems they don’t understand.”

He worries that the development of suggestion techniques will cause people to overestimate not only the technical rigor but also the reliability of the results.

New AI tools, also known as big language models, are trained by billions of words taken on the Internet such as Wikipedia, news, social networks or from the user himself. They are taught how to analyze words and sentences, then simulate when asked.

“ AI tends to make up small details to fill in the gaps in the story, so it can confidently give a wrong answer,” says Goodside. Sometimes the AI ​​suffers from ‘hallucinations’ and makes meaningless statements. They do not represent real human knowledge and thinking .”

Engineers suggest not disclosing what commands they need to put into the computer, and see that as the key to unlocking the precious rewards of an AI model. In 2022, an artist won a painting competition in the state of Colorado (USA) thanks to AI Midjourney. He declined to publicly disclose the description he entered into the system, saying only that it took 80 hours to perfect with more than 900 different versions.

Some people are also selling description text snippets on platforms like PromptBase. Ben Stokes, the founder of PromptBase, says about 25,000 accounts have bought and sold description text on the platform since 2021. A 50-word piece of text for Midjourney to create an image can sell for $1, 99 USD. About 700 suggested engineers are using PromptBase to sell user-requested documents.

But the work is becoming more and more professional. AI startup Anthropic, founded by former OpenAI employees, recently posted a “suggestion engineer” in San Francisco with a salary of up to 335,000 USD / year (equivalent to about 7.9 billion USD). copper).

Suggestion engineers are also sought after in companies outside the tech industry. Boston Children’s Hospital (USA) in February also posted an “AI suggestion engineer” to help build documents that analyze healthcare data from clinical research and treatment. The London-based law firm Mishcon de Reya also posted a search for “legal advice engineers”, in which applicants were asked to submit screenshots of conversations with ChatGPT.

Reference: Washington Post