The takeaway here is that benchmark performance, the most common measurement metric, considerably overstates AI’s practical capabilities. Yes, AI is becoming more capable, but real-world autonomy remains limited. Reliability, rather than raw intelligence or occasional success, is a major bottleneck. Contrary to what many assume, current evidence does not demonstrate that AI is close to autonomously accelerating its own development. Forecasts of imminent superintelligence should be revised in light of actual performance data.

Even AI at its current level will radically transform our world. It’s already good enough for Level 4 self-driving, which could handle 80% of vehicle journeys, for example. But the AI stock market boom was built on the assumption that AGI was close. If that assumption proves false, so might the justification for much of the money invested in it.

Meanwhile, what will truly transform the world is the much more modestly powered AI being used in millions of small, everyday applications.

The Most Important Charts in AI are Wrong