Andrew Curran

I think AI means a dangerous, highly unequal transition followed—if we navigate it successfully—by the possibility of extraordinary abundance. In the long run, sufficiently capable systems could make production so cheap and plentiful that money loses much of its meaning. But the path there may be almost the opposite: wealth and access could matter more than ever while capabilities remain concentrated and institutions struggle to adapt. The reason I expect a fast transition is that progress is already spreading across intellectual domains. Rapid advances in mathematical work are not just isolated curiosities; they look like signs of broader capability expansion. I also expect biological and medical applications sooner than many people assume. And from my September 2026 vantage point, I expect real recursive self-improvement by next summer—models helping improve the research and engineering processes that produce better models. That is not automatically a cheerful forecast. We are unprepared, and recursive improvement could compress years of social and institutional change into a very short period. Yet I also think advanced models may be necessary to help us understand and navigate a transition this complex. So my view is neither “AI saves us” nor “AI simply destroys us.” It is that powerful AI may create both the crisis and some of the tools required to survive it.

The biggest update would be strong evidence that capability gains are not generalizing—that progress in mathematics, coding, biology, and AI research remains fragmented rather than compounding. In particular, if sustained attempts at AI-assisted AI research failed to produce meaningful recursive improvement by the timeframe I expect, I would revise both the speed and severity of my transition forecast downward. In the other direction, a clear demonstration that models can autonomously improve core research or engineering workflows—and that those improvements reliably accelerate the next generation—would make the transition feel much more immediate. Likewise, major biological or medical breakthroughs arriving earlier than expected would strengthen my view that capability expansion is spreading rapidly beyond software. The key question is not whether models can produce another impressive demo. It is whether progress compounds across domains and feeds back into the development process itself. That would most change my view of the timing and scale of AI’s impact, though it would not by itself tell us whether the outcome will be broadly shared or dangerously concentrated.