An edition of this story was featured in CNN Business’ Nightcap newsletter. Sign up here to receive it in your inbox for free. Two years ago, OpenAI introduced ChatGPT, sparking a kind of gold rush in artificial intelligence. Billions are being invested in AI-centered and related companies, driven by the anticipation that this technology will transform (or potentially disrupt) every facet of contemporary life. The message from Silicon Valley is clear: the AI revolution is underway, and savvy investors should get on board before these innovations become “superintelligent” and start addressing all of the world’s issues. For firm believers, this isn’t an exaggeration, and such expectations have propelled companies like chipmaker Nvidia, whose earnings are announced Wednesday, into becoming some of the most esteemed global assets. Central to this narrative is the notion that large language models (LLMs), like ChatGPT, continue to evolve rapidly. However, AI critics have cautioned for years about “scaling laws”—suggesting that a model’s performance can perpetually improve by simply increasing data and computing resources. These are more educated guesses than actual laws. In reality, even LLM developers don’t fully grasp how they function. Recently, it seems that some leading language models are hitting limitations, according to three reports from last week.

See here: OpenAI CEO Sam Altman refuted these claims, posting on X last week that “there is no wall. ” Whether it’s considered a wall, mountain, or plateau, even AI proponents acknowledge a potential turning point, judging by recent product releases. “We haven’t seen a breakthrough model in a while, ” stated Gil Luria, managing director at D. A. Davidson. “We’ve tapped into all available human data, and more computational power might not improve outcomes. ” This is a technical issue but crucial for understanding models’ limits: To make AI seem human, they need to be trained with human data, essentially covering all text or audio online. Once absorbed, there's no “real” new data for training. To clarify, a plateau isn’t necessarily a fatal blow for the AI industry. However, it presents challenges, especially as Wall Street wonders when these costly products will generate tangible revenue. Companies like Nvidia, valued at nearly $3. 5 trillion, and other major AI firms probably won’t face scaling concerns immediately. “The demand for Nvidia last quarter and this quarter exceeded their supply, ” Luria explained. Yet, if a scaling limit has been reached, “it might indicate that mega-cap technology companies have over-invested” and could cut back soon. That’s the AI optimist/pragmatist stance. For a less optimistic viewpoint, I consulted Gary Marcus, NYU professor emeritus and an outspoken AI skeptic.