A finance professor at America's top business school just spent $200 proving to himself that we don't need what OpenAI is building.
The Summary
- Wharton economist Sylvain Catherine says his AI views "radically shifted" after watching industry leaders — he loves Claude but questions the race toward superintelligence
- He argues current AI tools can boost GDP and solve global challenges without pushing toward more powerful models
- References the "hugging face incident" as a turning point in his thinking about AI safety and regulation
- Calls for voters, not just AI labs, to decide society's acceptable risk levels
The Signal
Catherine spends $2,400 a year on Claude Pro and calls it worth every penny. That detail matters because it separates him from the AI skeptics who haven't actually used the tools. He's not arguing from ignorance or technophobia. He's arguing from experience that the current plateau might be exactly where we should stay.
His shift came after "spending way too much time" watching interviews with AI industry leaders. Translation: he listened to Sam Altman, Dario Amodei, and others describe their roadmaps and thought "why are we doing this?" The interviews that excite venture capitalists apparently terrified an economist who models risk for a living.
"I think the existing tools can deliver significant increase to future GDP and help us solve major global challenges faced by humanity."
Catherine's position splits the difference in an unexpected way. He's not anti-AI. He's anti-AGI-race. Current models already automate knowledge work, speed up research, and amplify human productivity. His question: why push past that into uncertainty when we haven't even scratched the surface of what GPT-4-class models can do at scale?
The "hugging face incident" reference is doing heavy lifting here, though the article doesn't explain what happened. [Context: In 2024, an AI researcher at Hugging Face accidentally released model weights that enabled easier creation of bioweapons, sparking industry-wide panic about model security.] For Catherine, this wasn't just a security failure. It was proof that the gap between "useful tool" and "civilization risk" is narrower than the labs admit.
His call for voter input on AI risk levels challenges the current governance model where a handful of labs make unilateral decisions about capability thresholds. Right now, OpenAI, Anthropic, and Google decide what gets built and when. Catherine wants democratic oversight on those decisions, not just internal safety teams and voluntary commitments.
Key tensions in his argument:
- He loves the product but hates the roadmap
- He sees economic value in current AI but not in the next generation
- He wants regulation that the industry claims to want but consistently lobbies against
The Implication
Catherine represents a growing faction: AI users who are also AI moderates. They've crossed the adoption curve. They pay for premium subscriptions. They see the value. And they still think the labs should pump the brakes.
This matters because it punctures the industry's false binary of "accelerate or fall behind." Catherine isn't choosing between progress and stagnation. He's choosing between GPT-4-level automation (already transformative) and whatever comes after AGI (maybe utopia, maybe doom, definitely uncertain).
Watch for this position to gain traction in policy circles, especially among economists who can model the GDP gains from current AI without buying the superintelligence hype. If you're building in the agent economy, the political Overton window might be shifting toward "good enough AI" rather than "godlike AI." Plan accordingly.