Import AI 458: Reckoning with the future; and a singularity story

Import AI issue 458 is made up of a lengthy essay drawn from a speech and a fictional story that the newsletter describes as an attempt to think through what a positive singularity might look like.

The talk is the 2026 Cosmos HAI Lab Lecture, delivered at the Human-Centered AI Lab (HAI Lab) in the Institute for Ethics in AI at the University of Oxford, in collaboration with the Cosmos Institute. It argues that rapid AI advances present individuals and societies with a choice: explore the future or retreat from the present. Retreating, in the talk's framing, means ignoring or dismissing the technology's implications and falling into reactivity or passivity.

The essay is organized in three parts. Part 1 outlines recent AI progress and argues that if the technology advances as much as the author expects, it cannot be treated as a normal technology. Part 2 examines the advance of AI through the author's own experience and that of Anthropic. Part 3 discusses what the talk calls humbling, almost unimaginable choices ahead.

The talk cites the Epoch Capabilities Index (ECI), which scores models over time on a basket of more than 40 benchmarks, and lists milestones including AI passing the bar exam in March 2023, silver at the International Math Olympiad in July 2024 and gold in July 2025, AI co-authoring new mathematical proofs in 2025, and systems such as Claude Mythos finding novel software flaws.

It states that AI progress rests on a common technology whose performance grows in relation to invested compute and data, and that companies are investing hundreds of billions of dollars in training facilities, locking in some future progress. It says an AI system smart enough to develop its own successor, starting recursive self-improvement, could arrive within the next two years, possibly sooner. It states that an elegant global slowdown would likely be a good thing, but that no coordinated global slowdown exists and commercial and geopolitical rivalries are drowning out larger aspects of the technology.

The author describes personal AI use from 2023 onward and a repeatable skill built from the Import AI newsletter archives that generated 20 graphs from a few hundred papers. The essay also says that after paternity leave from November 2025 to late February, colleagues had begun delegating more work to the model Opus 4.6, with some no longer writing code themselves, and that code output inside Anthropic rose substantially.