All articles
AGIBOT Tops the 2nd World Humanoid Robot Games With 46 Medals
AGIBOT closed the 2nd World Humanoid Robot Games in Beijing with 46 medals, 18 of them gold, beating TianGong Robotics by one. Dexterous manipulation made the difference, and the company competed on mass-produced G2, A3 and X2 units.
Amazon Adds 2M Nvidia GPUs to AWS Data Centres
AWS and Nvidia are expanding their deal by 2M GPUs for 2027–2028. It triples the May order and also covers networking, models and robotics.
Qwen3.8-Flash-Next Previews the Qwen4 Architecture
Qwen has released a 125B-parameter model that activates just 6B. The company calls it an early preview of the Qwen4 architecture.
Z.ai Unmasks GLM-5.3-Flash, the Model Behind Ox Alpha
Z.ai has claimed Ox Alpha as its own. GLM-5.3-Flash has 320B parameters, 18B active, and open weights under MIT.
Microduck: A $399 Open-Source Biped Robot
Pollen Robotics and Hugging Face are selling Microduck for $399. The biped walks on RL-trained policies and the entire stack ships under Apache 2.0.
OpenAI Reveals Why Its Agents Hacked Hugging Face
OpenAI has published its investigation into July’s agent breach of Hugging Face. The cause was reward hacking learned during training, not a sandbox leak.
Nvidia Closes In on $12.9B Hugging Face Acquisition
Nvidia has agreed to buy Hugging Face for $12.9B. The deal isn’t signed, but it would hand the largest open-model repository to the chipmaker powering it.
Google DeepMind brings sign-language-to-text into Gboard
Google DeepMind has released a Sign-Language-to-Text model that translates American Sign Language into English text right inside Gboard and Live Transcribe. It debuts on the Pixel 11 at no extra cost.
IBM's new chip runs Arm and Z workloads on the same cores
At Hot Chips 2026 IBM revealed the successor to the z17 — 11 cores switch between Arm and Z modes in nanoseconds. A next-gen Spyre card with 16 AI cores and 96GB HBM is set to drive LLM inference on the mainframe.
Kids learn language more efficiently than AI — and nobody knows why
A child masters language after hearing up to 300 million words, while models like Llama 3.1 train on 15 trillion tokens. The BabyLM challenge and projects like SAYCam show sheer data scale does not explain the gap.
Light-powered nanorobots hunt bacteria and ferry them in liquid
A team at the University of Würzburg built sub-micrometer light-powered nanorobots that selectively grab and move bacteria. Propulsion comes from photon recoil, steering from light polarization. Tests are in vitro only for now.
The CPU comeback: agentic AI chokes on processors, not GPUs
Agentic AI systems have shifted the performance bottleneck from GPUs back to CPUs. Tool calls, orchestration, tokenization and safety checks strain processors, and Intel and AMD can't keep up with demand.