Chatbots Are a Waste of A.I.'S Real Potential (www.nytimes.com)

In a New York Times opinion, Gary Marcus argues that Big Tech’s fixation on general-purpose generative models and chatbots is diverting resources from AI’s most productive paths. He points to persistent hallucinations, safety failures and weak business returns—citing an MIT NANDA study where 95% of AI pilots showed little or no ROI and a financial projection of an $800 billion revenue shortfall for AI firms by 2030—to argue that one-size-fits-all LLMs are overhyped and often unreliable in practice. Marcus contrasts that trend with the demonstrated power of narrow, domain-engineered systems. Examples include AlphaFold, which combines biochemical priors and tailored ML/classical AI to predict protein structures at scale; Waymo’s modular stack for sensing, perception and decision-making in autonomous driving; and classical chess engines that enforce rules and search game trees reliably, unlike LLMs that can make illegal moves. He notes real-world failures of generative approaches (e.g., driverless startup Ghost Autonomy) and safety risks from unconstrained models. Marcus doesn’t call to abandon AGI research or generative tools entirely—he urges a strategic pivot toward specialized models that embed domain knowledge, constraints and safety guardrails to deliver more reliable, useful and provably safe AI in science, medicine and industry.