AI News — Sunday, August 23, 2026
Researchers introduce OmniScientist, an AI capable of performing scientific tasks across multiple modalities and disciplines, marking a significant step towards general AI in scientific discovery.
Startup Inherent, founded by DeepMind alumni, claims its AI 'teammate' has surpassed models from Anthropic and OpenAI in the complex task of replicating scientific research.
OpenAI has publicly advocated for California to enact a stronger AI safety bill, emphasizing the need for robust regulation in the rapidly evolving field.
OpenAI has launched a new version of ChatGPT specifically designed for teenagers, focusing on educational use cases and incorporating enhanced safety protections.
A 12-year-old developer successfully built and launched a full-stack AI SaaS application using only an Android phone, demonstrating remarkable resourcefulness and the accessibility of AI development.
SkillEvo proposes a novel method for AI agents to continuously improve their skills through self-renewing evolution gradients derived from multi-turn interaction feedback.
Harvard's new $699 startup bootcamp is incorporating AI avatars of its instructors to enhance the learning experience and provide personalized guidance.
A new service named Goose is offering code generation capabilities comparable to Anthropic's Claude Code, but at no cost, potentially disrupting the market for AI coding assistants.
A developer created an AI tool that automatically replies to Instagram DMs without requiring users to log in, offering a convenient and privacy-conscious solution.
Repo0 introduces a novel design-driven approach to code generation, aiming to create entire software projects from initial design specifications.
Google has introduced five new AI-powered features within Search designed to enhance learning and study tools for students and users.
This article discusses how simply scaling up an AI model did not resolve persistent planning errors, highlighting the need for qualitative improvements beyond model size.
This paper presents ForgeWM, a new approach for training video world models that can predict future frames based on actions, focusing on few-step causal progression.
This research explores the challenge of 'unlearning' information in Large Language Models, proposing an adaptive popularity mechanism to address the difficulty of forgetting more popular data.
Researchers introduce Centered Residual Signatures as a method to verify the training lineage of language models, providing a way to trace their development and potential influences.