Unsolved frontiers
There is no invented percentage "to AGI" here. Instead, we show concrete capabilities still missing from reliable general intelligence.
Three unsolved problems. While even one stays open, the upper part of the line remains dark.
Learning from experience
Today a model does not remember yesterday's work: everything it knows is frozen at training time. We need a system that learns while working without forgetting what it already knew.
Full dossierReliable autonomy
Before trusting AI with an operating theatre, a power plant or a company's books, you must be able to prove it will not fail silently. No such method exists today.
Full dossierActing in the physical world
A model can describe an action, but the real world requires continuous perception, physical reasoning and correction before errors become irreversible.
Full dossierCompute and power bottleneck
AI capability is constrained by more than algorithms. Chips, memory, data centres, electrical grids and cooling determine which models can be trained and how many runs the economy can sustain.
Full dossierAGI
Forecasters cluster on 2027–2033 with a long tail past 2040. The spread is enormous, and much of the argument is about definitions.
Full dossierASI
Extrapolation stops working here. No single branch leads to this point alone: continual learning, verifiability, physical grounding and an accessible compute base would all be required at once.
Full dossierWhat the forecasts say
Estimates run from "already in this window" to mid-century. That spread is the honest picture: nobody knows the date.
Superintelligence in "a few thousand days". The conversation moved from AGI to what comes after.
Timelines are compressing: human-level AI is possible within a few years.
Genuine human-level AGI is a multi-year horizon, definitely not "already here".
The earliest public estimate: "smarter than the smartest human" already in this window.
The bulk of the probability mass, with a visible tail past 2040. Every recent update moved earlier, not later.
50% probability for high-level machine intelligence. Stanford separately insists: no AGI in 2026.