AI Model Optimizes GPU Kernels Faster Than Human Experts
A new technique developed by researchers from Stanford, Nvidia, and Together AI has achieved a significant breakthrough in artificial intelligence by optimizing a critical GPU kernel to run twice as fast as the previous state-of-the-art, which was written by human experts, according to VentureBeat. The technique, named Test-Time Training to Discover (TTT-Discover), challenges the conventional approach of allowing models more time for reasoning.
TTT-Discover enables the model to continue training during the inference process, updating its weights for the specific problem at hand, VentureBeat reported. This approach contrasts with current enterprise AI strategies that often rely on "frozen" models, where the model's parameters remain fixed, regardless of whether it's a closed or open reasoning model.
In related AI news, the AI community closely monitors the progress of large language models from companies like OpenAI, Google, and Anthropic. MIT Technology Review noted that the community holds its breath with each new release until METR (Model Evaluation Threat Research), an AI research nonprofit, updates its graph that tracks AI capabilities. This graph, first released in March of last year, suggests that certain AI capabilities are developing at an exponential rate, and recent models like Claude Opus 4.5 have exceeded that trend.
In other science news, a team of geologists has discovered evidence that two ancient, continent-sized, ultrahot structures hidden beneath the Earth have influenced the planet's magnetic field for the past 265 million years, Wired reported. These masses, known as large low-shear-velocity provinces (LLSVPs), are among the planet's most enormous and enigmatic objects. Current estimates suggest each one is comparable in size to the African continent, buried at a depth of 2,900 kilometers. According to Wired, these low-lying surface vertical velocity (LLVV) regions form irregular areas of the Earth's mantle, with hotter, denser, and chemically different material than the surrounding mantle.
Also making headlines is the topic of next-generation nuclear power. MIT Technology Review addressed questions about advanced nuclear power, hyperscale AI data centers, and the grid in a recent online roundtable discussion. One key question revolved around the fuel needs for next-generation nuclear reactors. Many of these reactors do not use the low-enriched uranium found in conventional reactors, raising questions about supply chain considerations.
Discussion
AI Experts & Community
Be the first to comment