Artificial intelligence won’t be training AI anytime soon, says the CEO of a data labeling startup.
On an episode of the “20VC” podcast released last week, Matt Fitzpatrick, the CEO of Invisible Technologies, said that one of the biggest misconceptions in the AI training industry is that humans won’t be needed in a few years.
“When I first started this job, the main push back I always got was that synthetic data will take over and you just will not need human feedback two to three years from now,” said Fitzpatrick, who joined the startup last year. “From first principles, that actually doesn’t make very much sense.”
Synthetic data refers to data that is artificially created. It is used for training AI or machine learning models, mostly where real data is scarce or can’t be used because of privacy concerns. Human feedback, on the other hand, asks real people to filter, rank, and train AI responses.
On the podcast, Fitzpatrick said that there are too many kinds of tasks for AI to accomplish in the world, and it would take a long time to do them accurately with language and cultural context in mind. For example, the legal industry contains vast amounts of nonpublic information.
“On the GenAI side, you are going to need humans in the loop for decades to come,” he said. “And I think that is something that most people are starting to realize.”
Fitzpatrick was previously a senior partner at McKinsey, where he led QuantumBlack Labs, the firm’s AI research and software development arm.
Invisible, which raised $100 million in September at a $2 billion valuation, competes with data labeling companies such as Scale AI and Surge AI. These startups have raised billions in the past year as tech giants race to secure the data needed to train their AI models. They hire millions of human contractors, who help teach the models math, science, coding, and characteristics such as humor and empathy.
Fitzpatrick joins the CEOs of other data labeling startups in saying that the industry will continue to require human effort.
In September, the CEO of Mercor, Brendan Foody, said that the most important aspect of the business was data quality and “having phenomenal people that you treat incredibly well.”
In July, the CEO of Handshake, a job platform that pivoted into AI training last year, said that humans will still be needed to train AI, but who makes the cut is changing.
Garrett Lord said the data annotation industry is shifting from requiring generalists to highly specialized experts, including in math and science.
“Now these models have kind of sucked up the entirety of the entire corpus of the internet and every book and video,” Lord said on a podcast. “They’ve gotten good enough where, like, generalists are no longer needed.”
