CHARBONNEAU – AI learns about the world from army of low-wage workers

IMAGINE A CHILD who is able to write entire paragraphs and essays but has no knowledge of the world. They write about trees without being able to identify a tree from a picture.

In order to function in the real world of objects, to be aware of human nature and social mores, this precocious child needs to be told every detail of the way the world works.

Regurgitating contents does not make the child smart; it only gives the illusion of intelligence. The gifted child will need to learn mundane things like naming fingers and toes, eyes and nose. Some things learned are more critical, like how to safely cross the street.

That child is where artificial intelligence is now. It regurgitates the information and misinformation of the world without knowing the difference. To make the precocious AI smart, an army of nannies around the world is spoon-feeding AI.

Much progress has been made in filling the endless void but the appetite of the child AI is voracious.

It’s fun watching the gifted AI grow up. I was amazed when I first used ChatGPT to write a coherent paragraph; impressed when DALL·E 2 “painted” a image of Kamloops in an expressionistic style. I have it hanging on my wall, like a painting that your kid would bring home from school.

The army of nannies, called “raters,” annotate web content with the goal of making AI truly intelligent. Most are low-paid gig-workers but some actually make a decent wage.

In an investigative report by Joe Castaldo, he interviews dozens of raters and finds out what they do. Some of the work is incredibly dull and some requires expert knowledge.
Raters rank responses from chatbots and converse with them in real time. They edit AI-written copy and catch mistakes; decide which AI-generated image is best and teach AI about biology, chemistry, computer science, history, law, marketing, physics, poetry, creative writing. They transfer a massive amount of knowledge representing centuries of discovery and humanity into the bright but nonhuman child.

Aesop Khaemba, who lives in Nairobi, makes $50 a week if he’s lucky. Sometimes he spends a whole day labeling objects sent to him. Sometimes he gets pictures and videos of street scenes, along with images generated by an autonomous vehicle’s various sensors. Then he painstakingly draws boundary boxes around cars, pedestrians, traffic lights and other objects. One scene could take a few hours or as long as two days to complete.

Like teaching a child how to safely cross the road, teaching self-driving cars the real world is critical. The spoon-fed models have taken technology so far –but not far enough. Self-driving vehicles still respond incorrectly in unfamiliar situations.

Other raters can put their expertise to work. One outsourcing company recently posted job openings for experts to train large language models, or LLMs, in subjects such as law, philosophy and marketing, with offers up to US$45 an hour. LLMs power ChatGPT and other chatbots.

Artificial intelligence is actually based on human intelligence—centuries of human intelligence.

Will the capricious AI child bite the hand that feeds it once it grows up? We’ll see.

David Charbonneau is a retired TRU electronics instructor who hosts a blog at http://www.eyeviewkamloops.wordpress.com.