Latest developments in artificial intelligence, including large language models (LLMs), generative AI, multimodal AI, AI agents, robotics, and AI regulation and policy.
Chris Fall, Director of CAISI under NIST, has stepped down after just three months. His predecessor resigned in less than a week. Tensions between the administration and the AI industry appear to be a key factor.
Google is developing a new chip, "Frozen v2," for its LLM "Gemini," aiming for 6–10x efficiency gains over existing AI chips, with a 2028 launch planned. Shares rose 3% amid investor optimism.
At WAIC, four major AI firms—MiniMax, StepFun, Kimi, and Zhipu—partnered with AI infrastructure company Infinigence to bridge the gap between soaring inference demand and limited computational resources with the Agentic Infra strategy.
Y Combinator leader Garry Tan asserts that true competitive advantage in the AI era lies in workflow redesign rather than model performance. He highlights examples like a 400-fold efficiency boost and achieving nine-figure ARR with a 15-person team, presenting new organizational rules that challenge conventional norms.
Moreton Bay City in Queensland will conduct Australia's first trial of an AI-powered traffic light system in late 2026, highlighting not only technical challenges but also issues of fairness in public road priority and cybersecurity.
NVIDIA CEO Jensen Huang visited Japan to finalize three major agreements with Japanese industries and semiconductor suppliers: a national AI factory, robotics collaboration, and chip material supply.
Riemann Dynamics unveils its robot operation model, Riemann-1.0, trained on 200,000 hours of first-person human videos. It achieved a 62.6% success rate on the RoboCasa-365 benchmark, surpassing the previous state-of-the-art (SOTA) by 8.4 points.
At WAIC 2026, Zhongke WenGe revealed an AI decision-making product suite covering data governance to agent execution. Its decision engine, Decitron, achieved a 160.28% annualized return in predictive market simulations.
A developer shares a firsthand comparison of Kimi K3 and Claude, highlighting Kimi K3's lower costs and comparable performance. The article also critiques U.S. AI regulations.
Author Dave Eggers delivered a speech to around 200 OpenAI staff at the invitation of CEO Sam Altman, harshly criticizing ChatGPT for devastating the education sector, stripping students of their creative abilities, and silencing an entire generation.
TeleAI, the Artificial Intelligence Research Institute of China Telecom, has developed AI Flow technology, enabling lightweight drones to directly connect with satellites and transmit high-definition video in real-time using token encoding and generative models. This breakthrough has overcome traditional communication constraints and won the top award at the World Artificial Intelligence Conference (WAIC).
Zoox has issued a software recall for 105 robotaxis due to interference with emergency operations caused by an inability to detect smoke. NHTSA guidance played a role in this decision.
Google's Gboard 'Sign-to-text' feature: converts sign language to text. Found in beta code. Claims on-device AI, but data may go to cloud.
Databricks has solidified its position as an AI-focused enterprise, achieving a valuation of $188 billion. The company raised approximately $3 billion in a new funding round led by Coatue and published research on reducing coding costs with open-weight models.
FireSat, funded by Google, has successfully launched three satellites designed for early wildfire detection, capable of spotting small-scale fires that traditional satellites often miss.
Agility Robotics has opened a 60,000-square-foot training facility in Fremont, California, near Tesla's Optimus factory. Exploring Agility's strategy as a leader in humanoid robot commercialization.
China's high-end appliance brand Casarte, under Haier Group, partners with Youku to launch AIOS-equipped TVs, aiming to diversify beyond refrigerators and washing machines while transitioning to a smart home ecosystem.
Mozilla AI has launched Llamafile 0.10.4, adding the new Transcribefile feature. This allows speech recognition models to run as a single file and supports over 16 models based on the Transcribe.cpp project.
China's Moonshot AI has announced the release of the 2.8 trillion-parameter large language model Kimi K3. Open weights are set to be available on July 27, 2026. Priced at the high end for Chinese AI models, Kimi K3 outperforms Claude Opus 4.8 in benchmarks.
Researchers at Manchester Metropolitan University successfully embedded a backdoor into an open-weight AI model for under $100 in just an hour, exposing vulnerabilities in the AI supply chain.