Bubble or Breakthrough? Marcus, Mollick and Narayanan Read AI's Big Week Very Differently

In a single day, Anthropic opened its books for a Nasdaq listing and OpenAI said its models had settled problems that stumped researchers for decades. Three of the most-followed voices in AI looked at the same news and told very different stories, and the gap between them says a lot about where this technology really stands.

Photo: Pavel Danilyuk / Pexels

What Gary Marcus sees

Gary Marcus, the cognitive scientist and longtime critic of large language models, spent Monday, October 26, picking apart Anthropic's S-1. His main complaint is timing: the filing covers the six months to June 30, and he argues that leaving out July through September hides a slowdown after what he calls a peak in heavy token use. He noted that Anthropic's token prices were hitting a record low in the same week as the listing and summed it up in one line: "Fantasy math and reality are about to collide." He called this possibly generative AI's make-or-break week and flagged mid-November, when he expects updated third-quarter numbers, as the dates to watch. His worry goes beyond one company, too, since he agreed with Elon Musk that most of the smaller cloud firms renting out AI chips will not survive tighter debt markets.

Marcus was more cautious than usual about OpenAI's science claims. He first called the claimed refutation of the Exponential Time Hypothesis a big deal if correct, then wrote that his reaction had been too generous and that the result needs context. For him, the pattern is familiar: bold announcements arrive first, and careful checking arrives much later, if at all.

What Mollick and Narayanan answer

Ethan Mollick, the Wharton professor who studies how people actually use AI, is watching a different scoreboard. On Monday, October 26, he described asking GPT-6 Pro to revisit an earlier finding that violent video games may reduce crime because people stay home to play them. The model downloaded fresh crime data and extended the old work, linking the 2013 launch of GTA V with a dip in violence internationally, while other game launches gave murkier results. He also shared OpenAI's announcement and argued that a small group of people predicted today's AI with remarkable accuracy a decade ago. His standing advice is to keep pushing your use of AI to be more ambitious until you either hit the frontier or discover the model can do more than you expected.

Arvind Narayanan of Princeton, co-author of the book "AI Snake Oil," takes a third position that cuts against both hype and doom. In recent posts he argued that full automation never depends only on a model crossing some capability line, pointing to robotaxi companies that still rely on first responders, remote staff and gig workers when riders fall ill or leave doors open. He also shared an argument that people who are good at predicting benchmark scores often carry crude models of how society works, which produces forecasts of explosive change that rarely arrive. On that view, a proof in a lab and a profitable business are separated by years of slow, messy adoption.

Our take

All three are partly right, and readers should hold two ideas at once. The capability gains look real: a model that can redo an economics study in an afternoon, or produce a proof that leading complexity theorists agree to examine, is not a parlor trick, even though OpenAI's claims still need independent review by the group led by Russell Impagliazzo, Ramamohan Paturi and Ryan Williams. Yet Marcus is right that the money side is fragile. Anthropic's first-half revenue grew about fourteenfold to $16.29 billion and its operating loss shrank to $693 million, but its compute and infrastructure costs alone reached $8.95 billion in those six months, and the filing stops at June 30. Investors asked to buy in November deserve the latest quarter before they pay, and a company that sells itself on trust should want to publish it.

Narayanan's point is the one most people miss. Even brilliant models change the economy at the speed of institutions, contracts and habits, not at the speed of benchmarks, which is why robotaxis still need people standing by. So our verdict is that this is neither a bubble about to erase AI nor a breakthrough about to remake everything by spring. It is an expensive, genuinely useful technology whose financing has run ahead of its adoption, and the next real test comes when Anthropic's third-quarter numbers do.

Published by the Serey team with AI assistance.

Would you buy shares in an AI lab before seeing its latest quarter?