AI News — Saturday, August 15, 2026
OpenAI is previewing an 'Ultrafast mode' for its GPT-5.6 Sol model, promising up to 14 times faster inference speeds, which could revolutionize real-time AI applications.
OpenAI has published a comprehensive guide for developers, detailing how to effectively build applications and leverage the capabilities of their latest GPT-5.6 model.
Researchers present Alaya-EVOKE, a novel approach that transitions from linear-scaling supervision to creating an 'endless world' for AI, potentially enabling more expansive and continuous learning environments.
Google announced that users will now have the option to remove visible watermarks from its AI-generated content, offering more flexibility for creators.
A new unified infrastructure called LLMRouter has been introduced to streamline the development, evaluation, and deployment of Large Language Model routers, enhancing efficiency and consistency.
DreamX-Phi 1.0 introduces an action-conditioned video world model designed to improve robotic manipulation by allowing robots to predict and understand the consequences of their actions in a visual environment.
OmniScientist is presented as an ambitious AI model designed to function as an 'omni-modal, omni-discipline' scientist, capable of integrating and processing information across various modalities and scientific fields.
Kog is reportedly intensifying its research and development to optimize GPU utilization, aiming to extract significantly more inference performance from existing hardware.
Intern-S2-Preview is presented as a scientific agentic foundation model, aiming to provide advanced AI capabilities for scientific research and discovery.
Google's AI Mode in Search offers five practical tips on how its advanced features can help users engage more with the real world, from planning activities to discovering new places.
A new report highlights that while Claude Code offers coding assistance at a significant monthly cost, a competitor named Goose provides similar functionalities completely free of charge.
PlayWorld introduces a new benchmark for evaluating AI world models, focusing on their ability to handle long-horizon objectives through agent-based interactions within simulated environments.
This article argues that traditional vector databases fall short in providing durable and comprehensive memory solutions for advanced AI systems, highlighting the need for more robust architectures.
AutoDesign proposes a meta-harness optimization framework to enhance long-horizon agentic design processes, allowing AI agents to more effectively tackle complex, multi-step design challenges.
This research introduces a Spatial Memory Agent equipped with experience-grounded procedural memory, significantly improving its spatial intelligence and navigation capabilities.