The Overvaluation of AI Hype: A Wake-Up Call from Within

Andrew Ho, a former OpenAI researcher, left the company after just eight months. But it’s not his departure that’s generating headlines – it’s the $700,000 in equity he’s stuck with, unable to cash out until the IPO and lockup period are over.

In an interview with Fortune, Ho expressed concerns about the unsustainable growth of AI labs. He has started a new company catering to frontier labs’ need for high-end reinforcement learning datasets. Ho’s warning is not just about individual investors; it’s also a cautionary tale about the broader implications of unchecked hype in the tech industry.

The idea that AI will soon surpass human intelligence, rendering current investments obsolete, has become a rallying cry for many. However, Ho is skeptical. He believes that the models being developed are not as revolutionary as their proponents claim. The concept of Recursive Self-Improvement (RSI), which suggests an AI system can improve itself through research, is based on a fundamental misunderstanding of how research works.

“Research isn’t limited by models’ intelligence,” Ho says. “It’s hindered by a lack of ‘research taste’ – the ability to propose experiments, recognize important results, and reason about complex phenomena.” This critique highlights the limitations of current AI capabilities. While models excel in narrow domains like mathematical proof verification or data compilation, they struggle with more nuanced tasks that require human judgment and creativity.

The billions of dollars being poured into AI research have indeed produced impressive gains in verifiable areas. However, Ho believes that the next round of breakthroughs will come from tackling messier, less tractable problems – exactly what his new company aims to address. Some argue that compute costs could increase by 10x, leading to greater revenue for labs. But Ho remains unconvinced, pointing out that this scenario relies on the assumption of RSI-driven growth, which he believes is unlikely.

Ho’s stance aligns with Wall Street’s growing skepticism about the AI buildout. The recent sell-off in Meta and Google stocks reflects fears that companies funding the AI revolution may not see a sufficient return on their investment. Ho argues that labs are stuck in a Red Queen’s race – constantly paying more for each cycle of training, only to see cheaper competitors close the gap.

The winners in this scenario are unlikely to be the labs themselves but rather the companies providing them with infrastructure: Nvidia and Micron, which sell the chips driving AI development. Ho notes that labs have two possible escape routes from this trap – building their own chips or moving up into the application layer – but both options are daunting.

The story of Andrew Ho’s departure from OpenAI serves as a reminder that even within the tech elite, there are those who question the narrative of unbridled AI progress. As investors continue to pour money into AI research and development, it’s worth considering Ho’s warnings about overvaluation and the limitations of current models. Will the next big breakthrough come from RSI-driven growth or will it emerge from addressing more complex, human-centric problems? Only time – and the market – will tell.

For now, Andrew Ho is stuck with his locked-up equity, but he’s not dwelling on it. “You’re going to work,” he says; “don’t think too hard about these questions.” But for those willing to confront the uncertainty head-on, Ho’s story offers a timely and important wake-up call – one that should give pause even to the most ardent proponents of AI hype.