AGI race heats up

A key goal for artificial intelligence (AI) is to achieve what researchers call ‘artificial general intelligence’ (AGI). While there is some debate as to what that means, the ARC Foundation which developed a test of AGI defines the concept as “a system capable of efficiently acquiring new skills and solving novel problems for which it was neither explicitly designed nor trained.” Until last year, most AI models scored poorly on ARC’s AGI test, but that changed with the release of OpenAI’s so-called reasoning models. In September, Open AI’s o1 preview model scored 21% on the AGI test, and in December a highly tuned version of their o3 model scored 87.5%. That performance comes at a price, according to ARC, the cost per task of these highly tuned reasoning models approaches USD 20. However, last week, China-based developer Deepseek released a model that while only scoring 20.5%, apparently achieved that result for a mere USD 0.05 per task. As our Digital Innovations strategy team notes in their 2025 Outlook, we are in a new golden age for invention.

Source: ARC Foundation, December 2024