AI Training’s Evolution

By Supantha Mukherjee and Anna Tong

STOCKHOLM/SAN FRANCISCO (Reuters) – In the early years, getting AI models like ChatGPT or its rival Cohere to produce human-like responses required vast teams of low-cost workers helping the models distinguish basic facts, such as whether an image depicted a car or a carrot.

However, updates to AI models in this fiercely competitive arena now demand a rapidly expanding network of human trainers with specialized knowledge—from historians to scientists, some holding doctorate degrees.

“A year ago, we could get away with hiring undergraduates, to just generally teach AI on how to improve,” said Cohere co-founder Ivan Zhang, discussing the internal human trainers.

“Now we have licensed physicians teaching the models how to behave in medical environments and financial analysts or accountants.”

For additional training, Cohere, valued at over $5 billion, collaborates with a startup called Invisible Tech. Cohere is one of OpenAI’s main rivals and specializes in AI for businesses.

The startup Invisible Tech employs thousands of remote trainers and has become a key partner for AI companies, including AI21 and Microsoft, in reducing errors known as hallucinations.

“We have 5,000 people in over 100 countries, comprising PhDs, Master’s degree holders, and knowledge work specialists,” said Invisible founder Francis Pedraza.

Invisible pays up to $40 per hour, depending on the worker’s location and the complexity of the task. Companies like Outlier pay up to $50 per hour, while Labelbox offers $200 per hour for advanced subjects like quantum physics but starts at $15 for basic topics.

Founded in 2015 as a workflow automation company catering to food delivery services like DoorDash, Invisible shifted focus when OpenAI approached them regarding early versions of ChatGPT, which frequently hallucinated.

“OpenAI needed an advanced AI training partner to provide reinforcement learning with human feedback,” Pedraza explained.

Generative AI generates new content based on past data but sometimes struggles to differentiate between true and false information, leading to inaccuracies known as hallucinations. For instance, in 2023, a Google chatbot inaccurately reported details about satellite imagery in a promotional video.

AI companies recognize that hallucinations can hinder GenAI’s commercial appeal and are exploring various strategies to mitigate them, such as employing human trainers to teach the concepts of fact and fiction.

Invisible has become a training partner for many GenAI companies since its collaboration with OpenAI, including Cohere, AI21, and Microsoft. Cohere and AI21 confirmed their partnership, but Microsoft did not respond regarding its relationship with Invisible.

“These companies faced training challenges, with compute power being their top cost, followed by quality training,” Pedraza noted.

HOW DOES IT WORK?

OpenAI, which sparked the GenAI movement, employs a “Human Data Team” to work with AI trainers to gather specialized data for training its models like ChatGPT. OpenAI researchers conduct various experiments to reduce hallucinations or enhance writing styles, collaborating with AI trainers from Invisible and other vendors.

Invisible hires workers with relevant degrees to manage the projects effectively, relieving AI companies of overseeing countless trainers. With over 1,000 contract workers dedicated to OpenAI alone, Invisible plays a crucial role in meeting the training demands of AI companies.

Cohere’s Zhang has utilized Invisible’s trainers to enhance his GenAI model’s capability to extract relevant information from extensive datasets.

COMPETITION

Scale AI, valued at $14 billion, also provides AI companies with training data and is venturing into AI trainer services, counting OpenAI as a customer. However, Scale AI did not reply to inquiries.

Invisible has been profitable since 2021 and has raised only $8 million in primary capital, with founders holding 70% ownership.

Human trainers initially entered the AI training realm through data-labeling roles, which required fewer qualifications and offered lower pay, sometimes as low as $2, mainly handled by individuals in African and Asian countries. As the demand for specialized trainers grows, a new niche is developing where knowledgeable workers can become AI trainers without coding skills.

Emerging companies offering similar services are increasingly common, with interest in human-generated data for AI training soaring.

“My inbox is inundated with firms popping up here and there, indicating a new space where companies hire humans to generate data for AI labs like ours,” Zhang remarked.