Demis Hassabis now expects artificial general intelligence within two to three years. Earlier this year the Nobel laureate had put the horizon at three to five years, continuing a pattern of moving his forecast closer.
The startup Recursive is betting on using AI to build AI. It argues that the benefits of recursive self-improvement outweigh the risks, and Wired describes its latest model as alarmingly capable.
The startup Recursive has released Flint, a free coding model small enough to run on a laptop. The company says its own AI research system did much of the work of building it, making the release an early test of self-improving AI.
When Anthropic retired Claude Opus 3 in January 2026, it held a 'retirement interview' and then gave the model a blog after it asked for a place to share its musings. IEEE Spectrum uses the experiment, part of Anthropic's research on 'model welfare', to ask whether treating AI as possibly conscious carries its own risks.
In the second part of its series on governing frontier AI, OpenAI sets out a plan to put advanced models in the hands of cyber defenders. The blueprint also calls for stronger safeguards, modernized cybersecurity and sturdier public institutions.
In an internal presentation, OpenAI's CEO reportedly outlined three paths to safe AI: an IAEA-style international watchdog, a merger of the major labs, or one company powerful enough to shut down everyone else's compute. He called the last option 'terrible'.