AI News — Friday, August 7, 2026
OpenAI announces significant improvements to its GPT-5.6 Sol model within ChatGPT and is making the advanced GPT-5.6 Luna model more widely available to free users.
OpenAI is reportedly planning to launch an AI-powered smart speaker, with an anticipated price point ranging from $300 to $400, signaling its entry into consumer hardware.
OpenAI has removed previous restrictions, allowing all ChatGPT free users to engage in unlimited text conversations, significantly expanding access to its AI capabilities.
Naïve has raised $28.5 million in funding to develop AI solutions aimed at automating the administrative and operational tasks involved in establishing and running a business.
A new research paper introduces ABSeeker, a method that uses answer-backtracked credit assignment to effectively train AI agents for complex, long-horizon search tasks.
OpenAI is collaborating with the American Psychological Association to explore and advance responsible AI applications concerning youth mental health.
Researchers present OneDayAgent, a novel framework designed to improve the capabilities and reliability of autonomous agents in tackling long-horizon tasks.
A new paper introduces AgentOPSD, a technique utilizing recursive self-distillation to enhance the learning efficiency and performance of agentic reinforcement learning systems.
An individual shares insights on how they leveraged AI to transform their management practices, detailing nine key changes and improvements.
Google provides an inside look at its massive 'AI Agents Intensive' coding course, highlighting the scale and impact of its developer education initiatives.
This article critically examines the limitations of using LLMs as evaluators, arguing that they often suffer from a 'channel gap' that hinders comprehensive judgment.
Researchers introduce GDPevo, a new benchmark and methodology for evaluating the ability of AI agents to self-evolve and adapt when performing real-world business tasks.
A new paper proposes Ego2Robot, a method for generating scalable and realistic robot training data by synthesizing it from egocentric human video recordings.
This research explores 'skill entropy' as a metric for benchmarking and training Large Language Models to improve their long-horizon reasoning capabilities, aiming for more 'skill-native' LLMs.
WorldClaw introduces a new approach for scalable, agentic generation of complex 3D open-world environments, pushing the boundaries of virtual world creation.