Boston Today

In a wide-ranging Hoover History Lab conversation, historian Stephen Kotkin argues that good history starts with strong questions—and the discipline to hunt for disconfirming evidence, not just proof that flatters your thesis. The archive is essential, but also dangerous: state records reflect the state’s obsessions and blind spots, so historians have to triangulate across genres—periodicals, memoirs, travel accounts, local knowledge—to see what official paperwork misses. He also describes the historian’s craft as a practice of empathy: not sentimental identification, but the hard work of understanding motives and mental worlds, including those of perpetrators, without excusing them. Kotkin rejects “history is one damn thing after another” by insisting historians must pair depth with pattern recognition across cases. The payoff is not just better scholarship, but better public reasoning—especially against “junk history,” where lazy analogies (his example: reflexive Munich comparisons) substitute for analysis and shut down debate. He champions “analytical narrative”—showing rather than sermonizing—and writing that respects both specialists and the public by keeping the audience in view. Turning to today’s information ecosystem, he treats AI “slop” as a demand problem as much as a supply problem, and warns that attempts to suppress speech often produce perverse, unintended consequences—history’s recurring lesson.

A Google Doc roundtable throws cold water and rocket fuel on the same question: are we watching a genuine platform shift—or a historic misallocation of capital? The optimism case is simple: the 2017 “tabula rasa agents” era gave way to Transformers + scaling laws, producing general-purpose models—and now agents are back, riding on pretrained foundations (“this is the worst it will ever be”). The skepticism case is even simpler: economies don’t grow at AI-speed. If GDP expands ~2–4% annually, the spending base is arithmetically constrained—and the application layer still looks small next to the infrastructure binge. One sharp mismatch: roughly $400B in Nvidia chip sales versus < $100B in end-user AI product revenue, plus the reminder that total SaaS spend is < $1T. Capex can become the “escalator” problem—everyone must spend just to avoid falling behind, with no durable margins to show for it. Meanwhile, the productivity evidence is muddy: a widely-cited METR result suggests coding tools can slow experienced developers on familiar codebases, even as internal surveys report big self-perceived gains. The most actionable consensus lands somewhere else: energy. If AI demand does not stop climbing, infrastructure—especially power—becomes the real bottleneck, and the real policy story.