Author: uet200
-
Google Launches Gemini Embedding 2: Its First Multimodal Embedding Model
On March 10, 2026, Google DeepMind released Gemini Embedding 2 — the company’s first natively multimodal embedding model. Unlike traditional embedding models that handle only text, Gemini Embedding 2 maps text, images, video, audio, and documents into a single unified vector space, enabling cross-modal search, classification, and clustering across over 100 languages. Intermediate Key points […]
-
NVIDIA Nemotron 3 Super: 120B Hybrid Model Activates Only 12B Parameters for Agentic AI
On March 11, 2026, NVIDIA released Nemotron 3 Super — a 120-billion-parameter open-weight model that activates only 12 billion parameters per token, delivering frontier-class reasoning at a fraction of the compute cost. Designed for multi-agent AI systems, the model combines three distinct architectures into a single hybrid backbone and ships with a native 1-million-token context […]
-
Karpathy Open-Sources Autoresearch: 100 AI Experiments Overnight on One GPU
Andrej Karpathy, former director of AI at Tesla and co-founder of OpenAI, has open-sourced autoresearch — a compact 630-line Python framework that lets AI agents autonomously design, run, and evaluate machine learning experiments on a single GPU. Released on March 7, 2026, the tool has already garnered over 8,000 GitHub stars and sparked a broader […]
-
Shenzhen Longgang Backs OpenClaw with Millions in Subsidies for One-Person AI Companies
Key points On March 7, 2026, Shenzhen’s Longgang District Artificial Intelligence (Robotics) Bureau released a draft policy to support OpenClaw and one-person company development. The ten-point plan targets individual developers and micro-enterprises building on OpenClaw, with subsidies totaling millions of yuan, and was open for public comment through April 6. OpenClaw is a free, open-source […]
-
FlashAttention-4: Algorithm and Kernel Co-Design for Blackwell GPUs
Key points On March 5, 2026, Tri Dao and collaborators from Princeton University, Together AI, Meta, NVIDIA, and Colfax Research released FlashAttention-4, a ground-up redesign of the attention kernel optimized for NVIDIA’s Blackwell GPUs. The kernel reaches up to 1,605 TFLOPs/s on the B200 at 71 percent hardware utilization, running 1.3x faster than cuDNN 9.13 […]
-
LTX-2.3: Sharper Video, Native Portrait, and Cleaner Audio in Lightricks’ Latest Open-Source Model
Key points Lightricks released LTX-2.3 on March 5, 2026 under an Apache 2.0 license as a refinement of the 22-billion-parameter LTX-2 Diffusion Transformer that generates synchronized video and audio from a single architecture. The update focuses on four areas: a rebuilt VAE for sharper detail, an upgraded gated-attention text connector for prompt adherence, native portrait […]
-
Microsoft Releases Phi-4-Reasoning-Vision-15B: Small Model, Big Vision
Key points Microsoft released Phi-4-reasoning-vision-15B on March 4, 2026, a compact open-weight multimodal model that combines high-resolution visual perception with selective reasoning. It uses a mid-fusion architecture pairing a SigLIP-2 Naflex vision encoder, which processes up to 3,600 visual tokens at dynamic resolution, with the Phi-4-Reasoning language backbone. About 80% of training data uses <nothink> […]
-
NVIDIA Open-Sources SONIC: A Foundation Model for Humanoid Whole-Body Control
Key points NVIDIA has open-sourced SONIC, a 42-million-parameter humanoid whole-body control foundation model trained on over 100 million frames of motion-capture data, released on February 20, 2026. SONIC scales model capacity from 1.2 million to 42 million parameters, using 100M+ frames of motion capture and 9,000 GPU hours across 128 GPUs over 3 days. Its […]
-
Junyang Lin Steps Down as Qwen Tech Lead in Abrupt Departure
Key points On March 3, 2026, Junyang Lin, the tech lead and public face of Alibaba’s Qwen open-source AI project, announced his departure in a brief post on X reading “me stepping down. bye my beloved qwen.” The exit came just 24 hours after the Qwen team shipped the Qwen 3.5 small model series, and […]
-
Reverse Engineering Apple’s Neural Engine to Train Transformers on M4
Key points A developer reverse-engineered Apple’s Neural Engine (ANE) on M4 silicon and used it to train a transformer model, which Apple has never officially supported. The open-source project bypasses CoreML and talks directly to the hardware through private APIs, reaching 9.3ms per training step at 1.78 TFLOPS sustained on an M4 Mac Mini running […]