Progress on benchmarks that measure knowledge and reasoning has been steep enough that several are effectively saturated. If AGI were an exam, the answer would be "close". It is not an exam. The capabilities that separate a strong model from a general agent are the ones current methods find hardest.
The four gaps that matter
- Continual learning: updating from experience without a full retraining run.
- Long-horizon reliability: staying correct across hundreds of steps, not one answer.
- Calibration: knowing what it does not know, and saying so.
- Autonomy: deciding what to do next without a human framing each step.
Why benchmarks hide the gaps
A benchmark is one prompt with a checkable answer. Real work is a sequence of dependent decisions where an early error compounds. Systems that look strong in evaluation often fail in deployment for exactly that reason.
The honest position
We are close to systems that are useful across many domains and far from systems that are dependable across all of them. The distance between those two is the whole difficulty.