Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
Workers in Germany average 1,330 hours per year, the fewest among OECD countries, and generate about $98 per hour, while Mexican workers average 2,200 hours and produce only $25 per hour. Across 40 OECD economies, longer…
Data science and data analytics are distinct career paths with different focuses, skills, and educational requirements. Data analytics interprets existing data to identify trends and inform business decisions, while data…
Microsoft Edge introduced a developer preview of the Aion-1.0-Instruct small language model in Edge Canary and Dev channels, designed to run on more devices including those without GPUs. The browser also launched Languag…
ARC and AIcrowd launched the ARC White-Box Estimation Challenge, a contest to improve estimation algorithms for randomly-initialized neural networks. Participants must design algorithms that estimate the expected output …
Kapa, a company building AI assistants for technical documentation, developed a method to incorporate images into its RAG pipeline without incurring per-query vision costs. The firm indexes images by describing each one …
A Swiss family doctor has developed shii • haa, a breathing app that uses a phone's microphone to provide live biofeedback on rhythm, depth, and regularity without uploading any speech or raw audio. The app processes all…
OpenAI will wind down its self-serve fine-tuning platform in May 2026, formalizing a split in how AI teams iterate on products. A small fraction of companies continue intensive model training, while most teams now focus …
Baz improved its AI Agent Code Review accuracy by building a Spec Review agent using Amazon Bedrock AgentCore, automating the validation of code against product and design requirements. The system orchestrates multi-stag…
The field of LLM explainability is advancing with new frameworks like SMILE and gSMILE that use statistical methods to create visual heatmaps showing which parts of a user prompt most influenced a model's output. Researc…
Snowflake is launching Adaptive Compute, a new workload-aware compute engine that automatically adjusts to changing demand without manual configuration or capacity planning. The technology, available soon in Adaptive War…
InfiniBand is a purpose-built network interconnect designed to link compute, storage, and accelerator nodes with high bandwidth and low, predictable latency. The technology integrates Remote Direct Memory Access (RDMA) d…
MLectio, a new company founded by a veteran eHealth researcher, is launching a platform to license academic curricula and exams for AI training, aiming to replace the current "shadow extraction" of educational material b…
Speculative decoding accelerates large language model inference by using a small draft model to propose multiple tokens, which a large target model then verifies in a single forward pass, reducing the number of expensive…
Google engineers deployed a Gemma 3 large language model across two GKE clusters in separate regions, using four TPU v6e chips per cluster, to test multi-cluster failover with the Inference Gateway. The setup leveraged m…
Financial institutions are shifting from siloed, task-specific AI models to unified transaction foundation models trained on billions of proprietary financial events, according to NVIDIA's 2026 State of AI in Financial S…
NVIDIA Apex's FusedAdam optimizer and FusedLayerNorm normalization layers can accelerate Transformer training by up to 30% compared to standard PyTorch implementations, according to benchmark tests. The performance gains…
Open-weight models now account for 69.1% of named token volume on the OpenRouter API platform, compared to 30.9% for closed models, according to the latest platform data. The shift, driven by rapid competition among deve…
NVIDIA released Cosmos 3, a multimodal world foundation model that processes text, images, video, audio, and actions within a single architecture for robotics and physical AI development. The model generates physically a…
NVIDIA released Cosmos 3, a family of world foundation models that generate photorealistic synthetic video data for training robot arms and physical AI systems. The models address the data scarcity problem in robotics by…
Muhammad Khalifa, a research scientist at NVIDIA, is advancing AI agents through his work on test-time techniques, reasoning agents, and LLM post-training. He completed his PhD at the University of Michigan in 2026 and p…