If your AI seems smarter, it’s because of the smarter human trainers at Reuters

By Supantha Mukherjee and Anna Tong

STOCKHOLM/SAN FRANCISCO (Reuters) – In the early years, getting artificial intelligence models like ChatGPT or its rival Cohere to spit out human-like answers required vast teams of low-cost workers to help the models distinguish basic facts such as whether a picture was of a car or a carrot.

But more sophisticated updates to AI models in the highly competitive arena now require a rapidly expanding network of human trainers who have specialist knowledge — from historians to scientists, some with doctorates.

“A year ago, we could get away with hiring students to generally teach AI how to improve,” Cohere co-founder Ivan Zhang said of its in-house human trainers.

“Now we have licensed doctors teaching models how to behave in medical settings, or financial analysts or accountants.”

For more training, Cohere, which was last valued at more than $5 billion, is working with a startup called Invisible Tech. Cohere is one of OpenAI’s main rivals and specializes in business AI.

Startup Invisible Tech employs thousands of trainers who work remotely and has become one of the main partners of AI companies from AI21 to Microsoft (NASDAQ: ) to train its AI models to reduce errors, known in the AI ​​world under the name of hallucinations.

“We have 5,000 people in more than 100 countries around the world who are doctors, masters degree holders and knowledge specialists,” said Invisible founder Francis Pedraza.

Invisible pays up to $40 an hour, depending on the worker’s location and the complexity of the work. Some companies like Outlier pay up to $50 an hour, while another company called Labelbox said it pays up to $200 an hour for “high expertise” topics like quantum physics, but starts at $15 for topics basic.

Invisible was founded in 2015 as a workflow automation company that turns to food delivery companies DoorDash (NASDAQ: ) to digitize their delivery menu. But things changed when a relatively unknown research firm called OpenAI contacted them in the spring of 2022, before ChatGPT’s public launch.

“OpenAI came to us with a problem, which was that when you asked an early version of ChatGPT a question, it would hallucinate. You couldn’t trust the answer,” Pedraza told Reuters.

“They needed an advanced AI training partner to provide reinforcement learning with human feedback.”

OpenAI did not respond to a request for comment.

Generative AI produces new content based on previous data used to train it. However, it sometimes cannot distinguish between true and false information and generates false results known as hallucinations. In one notable example, in 2023 a Google (NASDAQ: ) chatbot shared inaccurate information about which satellite first photographed a planet outside Earth’s solar system in a promotional video.

AI companies are aware that hallucinations can reduce GenAI’s business appeal and are trying various ways to reduce it, including using human trainers to teach the concept fact and fiction.

Since getting on board with OpenAI, Invisible says it has become AI training partners for most GenAI companies, including Cohere, AI21 and Microsoft. Cohere and AI21 have confirmed they are customers. Microsoft has not confirmed that it is a customer of Invisible.

“These are all companies that had training challenges, where the number one cost was computing power, and then the number two cost was quality training,” Pedraza said.

HOW DOES IT WORK?

OpenAI, which started the frenzy around GenAI, has a team of researchers called the “Human Data Team” that works with AI trainers to collect specialized data to train its models like ChatGPT.

OpenAI researchers are coming up with various experiments, such as reducing hallucinations or improving handwriting, and are working with AI trainers from Invisible and other vendors, a source familiar with the company’s processes said.

At any given time, dozens of experiments are underway, some with tools developed by OpenAI and others with vendor tools, the person said.

Based on what AI companies want – from improving Swedish history or making financial models – Invisible hires workers with relevant degrees for those projects, reducing the burden of managing hundreds of trainers for AI companies.

“OpenAI has some of the most incredible computer scientists in the world, but they’re not necessarily experts in Swedish history or chemistry questions or biology questions or anything you can ask,” Pedraza said, adding that more than 1,000 contract workers OpenAI. single.

Cohere’s Zhang said he personally used Invisible’s trainers to find a way to teach his GenAI model to find relevant information from a large data set.

COMPETITION

Competitors in this space include Scale AI, a privately held startup last valued at $14 billion that provides AI companies with training datasets. It has also ventured into providing AI trainers and counts OpenAI as a client. Scale AI did not respond to interview requests for this story.

Invisible, which has been profitable since 2021, has only raised $8 million in seed capital,

“We’re 70 percent owned by the team and only 30 percent by investors,” Pedraza said. “We facilitate secondary rounds, and the most recent transaction price was at a half-billion-dollar valuation.” Reuters could not confirm this assessment.

Human trainers first entered AI training through data labeling work that required less skill and also paid less, sometimes as little as $2, mostly done by people in African and Asian countries .

As AI companies roll out more advanced models, the demand for specialized trainers in dozens of languages ​​is growing, creating a high-paying niche where workers in a variety of disciplines could become AI trainers without even knowing how to code.

Demand from AI companies is leading to the creation of more companies offering similar services.

“My inbox is practically flooded with new companies popping up here and there. I see this as a new space where companies hire people just to create data for AI labs like us,” Zhang said.