AI Watch

As agentic AI workflows multiply the cost and latency of long reasoning chains, a team from the University of Maryland, Lawrence Livermore National Labs, Columbia University and TogetherAI has found a way to bake 3x throughput gains directly into a model’s weights.

Unlike speculative decoding, which requires a separate drafting model, this approach requires no additional infrastructure — just a single special token added to the model’s existing architecture.

Over the course of his career, Joseph McMullen has dealt with some of the most powerful agencies in the country: the FBI, Customs and Border Protection, and Immigration and Customs Enforcement. But in early 2024 the San Diego–based civil rights attorney faced a problem of scale. He had three federal trials in three months—two involving deaths in jail, one involving American children detained at the border—and terabytes of documents. He turned to artificial intelligence to help him get through it all.

McMullen’s path to the courtroom has been unconventional. A former analyst at the consulting firm Bain & Company, he received a law degree at the University of Virginia and trained at the Trial Lawyers College (now called the Gerry Spence Method) in Wyoming in a program that specialized in the emotional craft of storytelling. The emphasis he places on both analytical rigor and narrative instinct has led him, unexpectedly, to artificial intelligence.

Elon Musk, the entrepreneur behind SpaceX and the artificial intelligence company xAI, has unveiled one of his most ambitious plans yet. During an all-hands meeting with xAI staff on Tuesday, February 10, Musk announced a proposal to establish a manufacturing facility on the Moon that would build satellites equipped with advanced computing capabilities. According to The New York Times, Musk described the Moon as a necessary step in gaining a competitive advantage for future AI systems, saying simply, “You have to go to the Moon.”The idea involves constructing a lunar factory that could produce satellites outfitted with hardware designed to support artificial intelligence workloads.

Artificial intelligence is already embedded in most modern cameras, from autofocus tracking to in-camera noise reduction, but with the speed of AI advancement ever-increasing, it’s now possible to use AI before you even pick up the camera.

So-called “agentic AI”, where AI positions itself as an assistant in everyday life, is becoming increasingly popular. Naturally, we were curious to see if its usefulness extended to life as a photographer, and what we could therefore learn about how AI positions itself as a companion across every walk of life.

The emerging Age of AI has a tricky PR problem that may turn into a nasty political one. Microsoft chief executive Satya Nadella recently warned that AI could lose its “social permission” to consume massive amounts of electricity unless it delivers clear, real-world gains across society, including healthcare, education, government efficiency, and business performance.

That stark warning should be front-of-mind for techno-optimists. Too much of the public sees AI as extractive rather than beneficial, associating it with job threats, wealth concentration, and rising energy demand. Even worse, the message is getting through to at least some lawmakers.