Topic
Explore the latest artificial intelligence news and research.
Intel's EMIB-T has demonstrated, in separate test vehicles, a package exceeding 120mm square housing more than 9x reticle-equivalent silicon, and a 25µm interconnect. Whether it can deliver HBM4E-class speed while managing warpage and yield at volume production remains the next test.
Facing surging AI-driven demand, Samsung has raised supply prices for new customers by about 15% on its advanced 4nm process and other nodes. With full utilization expected in the second half of 2026, the company is shifting from a discount-driven order strategy to one focused on profitability—though success hinges on stable yields and sustained earnings improvement.
Google has announced LiteRT.js, a new runtime for running machine learning models in the browser. By bringing mobile model formats to the Web, it aims to reduce conversion overhead and make it easier to share models across devices, while also targeting WebGPU and future NPU utilization.
Samsung is reportedly developing a PC AI accelerator called "GAIA," advancing prototype verification with Lenovo and others toward mass production targeted for 2027. The chip will use a 4nm process, and cooperation with in-memory computing technology is also reportedly being considered, though power consumption and software support remain challenges for practical deployment.
SK hynix is set to resume equipment investment at its second NAND fab in Dalian, China. The company aims to bring it online by the first half of 2027, seeking to boost NAND supply capacity ahead of the completion of its new fab in Korea. However, completing the plan hinges on continued approval for equipment imports under U.S. export controls.
OpenAI announced that roughly 30% of the AI coding benchmark "SWE-Bench Pro" contains flaws, withdrawing its recommendation. Because defects such as failing correct solutions prevent accurate model evaluation, continuous auditing of the grading system itself—before comparing rankings—becomes essential going forward. (123 characters)
Meta has released an API for its flagship AI model "Muse Spark 1.1," which significantly strengthens coding and long-running agent performance. By shifting to a strategy of keeping model weights private and setting usage-based pricing far below competitors, Meta aims to capture the developer market.
GridCARE's AI "Energize" has uncovered idle capacity in the US grid, securing over 650MW of actual connection capacity with National Grid and PGE. The 300GW figure the CEO cites, however, remains a forward-looking projection on a different scale entirely from that track record.
Anthropic has released a new feature called "Reflect" that analyzes usage patterns from past conversation history. From a digital wellbeing perspective, it encourages breaks while also suggesting features tailored to individual work styles, aiming simultaneously to help users build an appropriate relationship with AI and to improve convenience.
