News — 2026-09-17

Researchers have introduced a new storage method for Ternary Large Language Models (LLM) that breaks the 1.58-bit barrier, potentially leading to more efficient AI processing. The method, called BITCOS, takes into account the actual symbol distribution of ternary LLM models, which can store weights more compactly than existing methods. Read more

In a significant shift, Anthropic and OpenAI have proposed embedding third-party evaluators to assess AI model safety, sparking both optimism and skepticism among researchers, who welcome the idea but emphasize the need for independence and legislative backing. Read more

Apple is developing an AI server powered by its high-performance M-series Ultra chips, aiming to capitalize on the growing popularity of its hardware among AI developers. The potential product, expected to launch in 2029, would be the company's first server in nearly two decades. Read more

Snap is introducing a new standalone AI assistant, Specs Intelligence, which will work with the company's new AR glasses and have dedicated desktop and mobile apps. The assistant builds an understanding of users' goals, priorities, and routines to surface timely reminders and relevant info. Read more

OpenSpec is a lightweight and configurable framework designed to create and manage software specifications, ensuring teams and coding agents stay aligned throughout the development process. By capturing requirements and validating implementations, OpenSpec helps build the right thing and build it right. Read more

DeepSeek-V4.1 Flash achieves nearly 420 Tokens/s in speed and reduces KVCache storage by 4x, making it a significant improvement over previous models. The model's architecture is designed for aggressive KVCache compression, with a parameter scale of 552B and native support for multimodal input. Read more