In Today's Issue: Where the terms came from, what the different definitions require, and what self-improving AI has actually demonstrated.

Dear Readers,

In a podcast published on March 23, NVIDIA chief executive Jensen Huang was asked when an AI might be able to build and run a technology company worth more than a billion dollars. His answer: "I think it's now. I think we've achieved AGI." He then described the possibility of an AI creating a viral service that briefly made money before disappearing. He ruled out those agents building NVIDIA. (Lex Fridman Podcast, 03/23/2026)

Since then, labs have put more effort into using AI to develop AI. Reuters reported that Google co-founder Sergey Brin was pushing resources toward recursive self-improvement, citing a person familiar with his efforts. Anthropic's August experiments show agents carrying out research and model training. Whether we call the result AGI affects expectations about which jobs machines can take on and how much responsibility people can hand over. (Reuters, 08/12/2026; Anthropic, 08/28/2026)

Artificial superintelligence, or ASI, raises the bar further. The Singularity adds a prediction about how deeply and quickly superhuman intelligence could change society. Recursive self-improvement offers one possible way AI could become more capable: better systems could help build still better successors.

When someone says AGI is here, what exactly has been achieved, and what would still have to happen for superintelligence and the Singularity to follow?

All the best,

Kim Isenberg

An artificial general intelligence would be able to tackle many kinds of intellectual work, including tasks it had not been built specifically to perform. It would need to learn unfamiliar tasks, use knowledge in new situations and recover from mistakes. The disagreement is over how well it must do those things before it earns the name.

Ben Goertzel traces the phrase to Mark Gubrud's use in 1997. In 2002, Shane Legg independently suggested it to Goertzel as a book title. The book and the conferences that followed helped establish the term. The ambition to build broadly capable AI came earlier.

Ray Kurzweil gave this future a timetable: 2029 for passing the Turing test, a test of whether a machine's conversation can be distinguished from a person's, and 2045 for the Singularity. In a 2024 interview, he connected the latter to humans merging with AI and greatly expanding their intelligence. His vision adds another meaning to the term: the future could involve people becoming more capable through technology, as well as machines surpassing them. (Kurzweil, 03/11/2024)