Reevaluating AI-2027: timelines, takeoff, alignment and China

According to LessWrong, researcher Stanislaw Krym revisits the "AI-2027" scenario—an influential thought experiment about how advanced AI might develop over the coming years and how safety testing could unfold. The scenario assumes superexponential progress, where AI capability gains accelerate without requiring new architectures. But Krym argues recent data tells a different story. Measurements from labs like Anthropic and Google show performance improvements scaling linearly with model size—advancing more gradually and predictably than the scenario expected. A key question: when will an "automated coder"—software powerful enough to replace human programmers—arrive? The scenario predicted early 2029, but Krym identifies three complications. Different benchmarks for code quality diverge in their predictions. Models in 2028 might be far larger than current scaling laws project, potentially reshuffling the timeline. And current improvement rates might not hold as models scale. On safety, Krym argues alignment problems could surface earlier and more gradually than the scenario assumes—spread over time rather than compressed into one critical moment. He also flags a complication: competitive pressure from Chinese AI efforts could tempt safety shortcuts. The broader point: frontier AI is advancing steadily, but perhaps less predictably and less rapidly than some forecasts suggest.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton