An untidy history of AI across four books (hedgehogreview.com)

This piece traces AI’s arc from symbolic, logic-driven aspirations to today’s data-and-compute‑driven machine learning, highlighting the contingencies that shaped the field: vast web-scale datasets, commodity GPUs, and the neural‑network breakthrough (AlexNet’s ImageNet success) that redirected research and industry resources. It notes OpenAI’s 2015 founding and the blockbuster, understaffed public debut of ChatGPT in 2022 as inflection points that made generative models ubiquitous and reallocated enormous capital to AI. Technical takeaways: symbolic systems hit representational limits; modern progress rests on optimization over massive data and compute; generative models are probabilistic and often brittle, while claims about predictive AI (accurately forecasting complex real-world outcomes) remain overstated. The article’s main intervention is a critique of AI boosterism found in recent popular books. Narayanan and Kapoor’s AI Snake Oil is praised for urging skepticism and clarifying key distinctions—especially between “generative” and “predictive” AI—and for warning about societal harms from hype, misuse, and opaque systems (e.g., COMPAS’s contested ~64% recidivism accuracy). By contrast, works by Harari and by Kissinger/Mundie/Schmidt are accused of technical misunderstandings and fear‑mongering that mystify tools, legitimize fatalism, and skew public debate. For the AI/ML community the significance is clear: technical literacy, precise claims about capabilities and limits, and careful public communication are essential to avoid policy missteps, ethical harms, and market distortion driven by sensationalism.