AI Is Dumber Than Investors Think — ft. Gary Marcus
Summary
Prof G interviews AI skeptic Gary Marcus, who argues that current large language models are inherently unreliable and driven by hype rather than genuine intelligence. Marcus discusses the technical limitations of next-token prediction, the lack of profitability in the AI sector, and the societal risks of hallucinations and sycophancy. He also analyzes recent regulatory shifts, including government subpoenas and executive orders, criticizing weak policies while advocating for stricter oversight to address cybersecurity and public safety concerns.
Topics discussed
Sponsors: Odoo, Plod, and Mount Sinai Women's Health
Gary Marcus introduced: AI critic and former Uber AI lead
LLMs are unreliable: Next-token prediction vs. true cognition
The Eliza effect: Humans over-attribute intelligence to AI
AI commoditization: Price wars and lack of differentiation
Sponsors: Odoo, VCX, and LinkedIn Ads
Regulation and liability: OpenAI subpoena and societal costs
Political shift: From anti-regulation to scrutiny of AI risks
Claude Mythos: Cybersecurity risks and vibe coding vulnerabilities
Sponsors: Prof G Markets, Indeed, and Toyota
Executive orders and the need for nuanced AI policy
The business model crisis: OpenAI's massive losses
Productivity myths: Why AI hasn't delivered ROI yet
Valuation reality check: Anthropic vs. OpenAI prospects
Conclusion: The danger of betting everything on LLMs
Credits and final sponsor: Shopify
Listen ad-free on Castria