Is AI About to “Eat Everything”? | AI Reality Check

In this AI Reality Check episode of Deep Questions, Cal Newport dissects the widely circulated METR Time Horizon chart that shows a steep rise in AI capabilities, particularly in programming tasks, from 2025 onward. He debunks the widespread panic that this indicates an imminent 'intelligence explosion' or that AI is 'about to eat everything.' Instead, Newport explains that the chart measures only how long a specific software task took human programmers to complete, and how well AI models with coding harnesses can solve those same tasks at 50% success rates. The dramatic jumps are not due to general AI intelligence but to two key developments: the shift from pre-training to post-training (fine-tuning models for specific tasks like coding), and the creation of highly sophisticated, hand-coded 'coding harnesses'—essentially expert systems that guide and verify AI-generated code. These tools, built over years of focused effort, are what enable the leap in performance, not a sudden emergence of superintelligence. Newport contrasts the flawed 'water level' mental model—where rising AI capability means universal transformation—with a more accurate 'river and tributaries' model: AI progress is not a single unstoppable force, but a series of narrow, hard-won breakthroughs in specific domains. He critiques the influence of transhumanist and existential risk communities, whose apocalyptic or utopian narratives distort public discourse and pressure AI leaders to adopt extreme rhetoric. Ultimately, he calls for AI companies to distance themselves from these cult-like narratives and communicate their tools in grounded, practical terms—celebrating real progress without fear-mongering or over-extrapolation.