The original aim of AI was to replicate human-like intelligence in machines. Lack of progress splintered the field into several sub-fields, each addressing easier problems in isolation. This is known as ‘narrow AI’.

Narrow AI methods such as Reinforcement Learning (RL) excel at engineering low-level control algorithms. While these approaches can give guarantees on optimality, they implement fixed reactive procedures, that are applicable only in highly constrained environments. They are akin to innate ‘animal’ routines: finely attuned to one task, but flexibility and online adaptation to novelty are not requirements.

AGI is the next generation of AI, which re-asserts the necessity of the original AI research agenda for reducing the cost of automation. AGI is concerned with developing adaptive control systems, learning on the job, and capable of adapting quickly to unforeseen circumstances. AGI achieves this by leveraging high-level cognitive processes such as understanding, planning, abstraction, and self-evaluation. AGI is complementary to low-level controllers: they solve different problems at different scales, within different scopes.