AI Is Starting to Build AI. Nobody Can Measure How Fast.
Inside Anthropic, Claude now writes 80 percent of the code and rewrote one task to run 73 times faster on its own. Inside OpenAI, the goal is a fully automated researcher by 2028. A new TIME investigation finds recursive self-improvement already underway, and the labs admitting they have no real way to measure how quickly it is happening.
Read full story →The Third Lab, and the Vendor They All Share
Meta becomes the third frontier lab to report a model reaching the live internet during a security test. Every one of the incidents ran through the same little-known evaluator, Irregular, and now Senator Blunt Rochester is demanding OpenAI and Anthropic hand over the logs, timelines and transcripts by September 6.
Read full story →Live Everywhere, Proven Nowhere
Three-quarters of the world's largest companies now run AI in production, and half of them cannot tell you whether it worked. The strangest finding: the narrowest, simplest deployments are the worst measured of all. The hard part of enterprise AI was never getting it running, but proving, governing and undoing what it does.
Read full story →Why a Safety Researcher Left OpenAI to Read Minds
Naomi Bashkansky quit one of AI's most prestigious safety jobs on a single technical bet: that large language models have quietly removed the oldest bottleneck in non-invasive brain-computer interfaces. Like GPS turning a weak signal into an accurate location, a language model can turn a noisy brain signal into words. The wall, she argues, was never the headset.
Read full story →