Topic
Explore English articles about Generative AI.
Examining materials claiming Qwen3.8-27B inference at 30+ tokens/s on a hypothetical 64-core XuanTie C950 configuration. We unpack the SHL optimizations behind the performance, the undisclosed test conditions, and the distance to mass production.
Demand for AI server SSDs and tight supply pushed the combined revenue of the top five NAND makers up 77% in Q2. We unpack how this price-driven growth could reverse when supply increases in 2027.
Etched has raised $700 million, pushing its valuation to $21 billion. It has delivered its first inference rack to Jane Street, but performance comparisons and disclosure of production scale remain to come.
Gartner forecasts that provider-side AI inference costs per agentic workflow will more than quintuple by 2028, shifting the focus from per-unit pricing to cost-per-outcome management.
Cursor has opened its Git hosting platform "Origin" as an early beta for paid plans. We look at the goal of unifying PRs, agents, and public APIs while retaining GitHub sync, along with the current constraints.
Against the backdrop of the 2025 crisis in which China's rare earth export controls halted EV production, Mercedes-Benz subsidiary YASA has begun work on making its axial flux motors rare-earth-free. Building on an architecture that holds a 59 kW/kg power density record, the company is developing two parallel paths: eliminating heavy rare earths and achieving complete rare-earth freedom.
AI models trained during safety fine-tuning to say they have no consciousness also systematically downgrade their attribution of mind to animals and natural objects, their religious beliefs, and even their subjective well-being. A team from Google and the University of Chicago reveals this unexpected ripple effect of alignment through geometric analysis of internal representations.
Payments giant Stripe has reportedly agreed to terms to acquire OpenRouter, which brokers comparison and use of AI models, for over $7 billion. Through this acquisition, Stripe aims to provide an end-to-end offering—from routing inference requests to metering usage and billing—monopolizing the monetization infrastructure for AI services.
Spot prices for 512-gigabit TLC NAND dies rebounded in August after hitting a low in July. While the immediate price movement is explained by technical factors like inventory and supply-demand dynamics, a new demand variable lurks in the background: NVIDIA Vera Rubin's new CMX storage tier, which represents a structurally different driver from the previous production-cut-led price increases.
Moonshot AI's PerceptionBench is a 3,000-question benchmark that strips away reasoning and knowledge to measure visual perception alone. Evaluating 16 state-of-the-art models, the top score was just 59.7%. It quantitatively reveals how similar overall scores mask uneven capabilities, and how what looks like a 'reasoning failure' is often actually a perception failure at its root.
