In the evolving landscape of artificial intelligence, the enhancement of models like ChatGPT and Cohere has shifted from basic training techniques to the involvement of highly specialized human trainers. This evolution is a response to the increasing complexity and sophistication required in AI systems, reflecting the competitive nature of the industry.
In its early stages, training AI models involved large teams of low-cost workers helping the systems identify simple concepts, such as differentiating between a car and a carrot. However, as the demand for advanced capabilities has surged, companies are now enlisting professionals with specialized knowledge, including historians and scientists, some holding doctoral degrees.
According to Ivan Zhang, co-founder of Cohere, “A year ago, we could get away with hiring undergraduates to teach AI on how to improve. Now, we have licensed physicians instructing models on behavior in medical environments, along with financial analysts and accountants.”
Cohere, valued at over $5 billion, collaborates with Invisible Tech, a startup employing thousands of remote trainers. Invisible Tech partners with various AI firms, including AI21 and Microsoft, to train models in an effort to minimize errors known as “hallucinations.”
Invisible Tech founder Francis Pedraza stated, “We have 5,000 people in over 100 countries, including PhDs and specialists in various fields.” The compensation for trainers varies based on location and expertise, with rates reaching as high as $200 per hour for topics like quantum physics. Other companies, like Outlier, pay up to $50 per hour, while basic topics may start at $15.
Founded in 2015, Invisible Tech initially focused on workflow automation for companies like DoorDash. However, its trajectory changed dramatically after a partnership with OpenAI in 2022, aimed at addressing the hallucinations prevalent in early versions of ChatGPT. Pedraza explained, “OpenAI came to us with a problem: early iterations of ChatGPT produced unreliable answers. They required an advanced AI training partner to enhance performance through human feedback.”
Generative AI, which creates new content based on historical data, sometimes struggles to differentiate between fact and fiction, leading to erroneous outputs. A notable incident in 2023 involved a Google chatbot sharing incorrect information about a satellite’s photographic capabilities. AI companies recognize that these inaccuracies can undermine the appeal of generative AI, prompting efforts to utilize human trainers to clarify distinctions between truth and falsehood.
Invisible Tech has positioned itself as a primary training partner for many generative AI firms, including Cohere and AI21, while Microsoft has yet to confirm its relationship. Pedraza noted, “These companies face training challenges, with compute power as their primary expense, followed closely by the cost of quality training.”
How the Process Works
OpenAI, the pioneer of generative AI, employs a team known as the “Human Data Team” to collaborate with AI trainers in gathering specialized training data. Researchers at OpenAI design various experiments to reduce hallucinations and enhance writing styles, working alongside trainers from Invisible and other vendors.
With numerous ongoing experiments—some utilizing OpenAI’s proprietary tools and others from external vendors—Invisible effectively manages the recruitment of workers with relevant degrees for specific projects. This alleviates the burden on AI companies that may not have expertise in niche areas.
Pedraza emphasized, “OpenAI boasts some of the finest computer scientists, but they may not be experts in specialized subjects like Swedish history or biology.” He noted that more than 1,000 contract workers are engaged solely for OpenAI’s needs.
Cohere’s Zhang has leveraged Invisible’s trainers to enhance his company’s generative AI model, focusing on extracting pertinent information from extensive data sets.
As the AI landscape continues to mature, the integration of specialized human trainers is set to play a crucial role in shaping the future of generative AI, ensuring that models produce reliable and accurate responses while meeting the demands of various industries.
Related topics:
Swiggy Files for $1.25 Billion IPO Amid India’s Stock Market Surge
U.S. Justice Department Launches Investigation into Super Micro Computer
Dell Mandates Full-Time In-Office Work for Global Sales Team