If Your AI Appears Smarter, It’s Due to More Skilled Human Trainers – Reuters

By Supantha Mukherjee and Anna Tong

STOCKHOLM/SAN FRANCISCO – In the early stages of developing AI models like ChatGPT and its competitor Cohere, substantial teams of low-cost workers were engaged to assist in differentiating basic elements, such as identifying whether an image depicted a car or a carrot.

However, advancements in AI have led to a greater need for a diverse network of human trainers with specialized expertise, ranging from historians to scientists, some of whom hold doctoral degrees.

"A year ago, we could rely on hiring undergraduates to provide general input for AI improvement," said Ivan Zhang, co-founder of Cohere, regarding their internal team of human trainers. "Now, we have licensed physicians training models to operate effectively in medical settings or financial experts guiding financial models."

For additional training, Cohere collaborates with a startup called Invisible Tech, which is valued at over $5 billion and focuses on AI for business applications. Invisible Tech employs thousands of remote trainers and partners with various AI companies, including AI21 and Microsoft, to minimize errors classified as hallucinations in AI models.

"We have 5,000 specialists across more than 100 countries, including those with PhDs and master’s degrees," stated Francis Pedraza, the founder of Invisible Tech.

Invisible offers compensation that can reach up to $40 per hour, depending on the worker’s location and the complexity of the task. Some companies pay up to $50 per hour for specialized knowledge, while others, like Labelbox, may offer up to $200 per hour for highly specialized subjects but start at $15 for more basic topics.

Founded in 2015 as a workflow automation service for various companies, including food delivery businesses, Invisible’s trajectory changed dramatically when it was approached by OpenAI in 2022, prior to the launch of ChatGPT.

"OpenAI identified an issue where an earlier version of ChatGPT would frequently hallucinate, making its answers unreliable," Pedraza explained. "They required an advanced AI training partner to enhance the model through reinforcement learning with human feedback."

Generative AI, like that used in ChatGPT, creates new content based on existing training data. However, sometimes it struggles to differentiate between accurate and inaccurate content, leading to misleading outputs. AI firms understand that these inaccuracies can undermine the appeal of generative AI to businesses and are actively seeking methods to address these issues, including employing human trainers to clarify the distinction between fact and fiction.

Since partnering with OpenAI, Invisible has managed to secure partnerships with several prominent generative AI companies, including Cohere and AI21. Though Microsoft has not confirmed its client status with Invisible, both Cohere and AI21 have acknowledged their collaboration.

"These companies face significant training challenges where their highest expenses stem from compute power, closely followed by the costs associated with quality training," Pedraza noted.

HOW DOES IT WORK?

OpenAI, a leader in the generative AI field, utilizes a dedicated team known as the "Human Data Team," which collaborates with AI trainers to gather essential data for training models like ChatGPT. The team conducts various experiments aimed at minimizing hallucinations and improving writing styles, utilizing resources from Invisible and other vendors.

At any given time, numerous experiments are underway, utilizing both OpenAI’s in-house tools and those from external vendors. Based on the specific needs of AI companies, Invisible recruits trainers with relevant academic backgrounds, alleviating the challenge of managing large numbers of trainers directly.

"OpenAI has some of the best computer scientists globally, but they need support in specialized areas such as Swedish history or biology," Pedraza remarked, noting that over 1,000 contract workers are involved in supporting OpenAI’s efforts.

Cohere’s Zhang added that he has utilized Invisible’s trainers to enhance his generative AI model’s ability to extract pertinent information from large datasets.

COMPETITION

Scale AI is a noteworthy competitor in this space, reportedly valued at $14 billion. This startup provides AI companies with training datasets and has begun offering human training services, counting OpenAI among its clientele.

Invisible has been profitable since 2021, despite having raised only $8 million in primary capital. "Our team owns 70% of the company, with investors holding the remaining 30%," Pedraza stated, adding that recent trades valued the company at around half a billion dollars, although this could not be independently verified.

Originally, human trainers entered the AI space primarily through data-labeling jobs that required less expertise and typically offered lower pay. As AI companies develop more advanced models, the demand for specialized trainers across various languages is increasing, creating lucrative opportunities for individuals with expertise in diverse subjects, even if they lack coding skills.

The rising demand from AI companies is spurring the creation of additional firms providing similar services. "My inbox is constantly filled with new companies emerging in this field. I see this as a burgeoning market where businesses hire humans specifically to generate data for AI applications," Zhang concluded.