Historical overview
The term was coined by Mark Gubrud in 1997 and popularised by Ben Goertzel and Shane Legg in the 2000s. It was historically marginal but has become central since 2020 as large language models began demonstrating broad capability.
Scientific basis
There is no consensus AGI architecture. Leading research directions combine large language models with tool use, memory, reinforcement learning, and search. Definitions vary from operational benchmarks (Levels of AGI, DeepMind 2024) to economic ones (median worker performance).
Strengths
- Would compress decades of scientific progress into years
- Could be deployed in parallel at marginal cost
Limitations
- No verified instance exists
- Alignment of goals and values remains an unsolved problem
Relationship to other intelligence systems
Artificial Intelligence
AGI is the projected end-state of current AI scaling laws.
Superintelligence
AGI is a stepping stone to superintelligence.
Future implications
Major labs project AGI timelines between 2027 and 2040; independent forecasters span a much wider range. The intervening years are the focus of contemporary alignment research.