Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
A company is hiring a Computer Vision / ML Engineer to work on video understanding, tracking, pose estimation, and spatial reasoning, according to a job posting. The role requires experience with Python, PyTorch, and Ope…
The Transformer neural network architecture, introduced in the 2017 paper "Attention is All You Need," has become the go-to deep learning architecture for text-generative models including OpenAI's GPT, Meta's Llama, and …
Amazon and Stanford University launched the Stanford and Amazon Research Initiative, a formal framework for joint research in AI, energy, and healthcare, announced at an event on Stanford's campus on September 16. More t…
James Hurley gave a talk on real-time video upscaling using convolutional neural networks and WebGPU. The presentation focused on performance considerations for running CNN-based upscaling in the browser.
Researchers Weifan Jiang and colleagues at Harvard University proposed H-Spec, a hybrid Mamba-attention parallel speculative decoding drafter that eliminates the drafter-side KV cache by reusing target KVs in place and i…
Northeastern University electrical and computer engineering professor Francesco Restuccia received a Presidential Early Career Award for Scientists and Engineers and $1 million to fund a five-year project developing math…
A new reinforcement learning approach targets long-horizon robot manipulation by training on individual failed subtasks rather than collecting full-task demonstrations, according to the paper "From Pretraining to Profici…
Amazon Bio Discovery published three papers on AI-driven antibody design, including a new sequence-based binding affinity predictor called MochiBind that ranks candidate antibodies by relative binding strength rather tha…
AI researcher Alexander Mattick argued on Machine Learning Street Talk that scaling prediction alone cannot produce genuine machine intelligence, citing the mathematical realities of scaling laws as continuous function s…
A paper submitted to arXiv on 11 September 2026 by Rodrigo Nicolau Almeida proves that Medvedev's logic of finite problems is undecidable, settling a longstanding open problem. The proof reduces the periodic tiling probl…
Modulate's ML team released the Entity Transcription Benchmark on Hugging Face, a free public dataset measuring per-entity transcription accuracy rather than word error rate. The benchmark contains 2,151 clips, six hours…
TypeSafe AI released Jev in September 2026, a classification model that returns a choice, score, or calibrated yes/no probability instead of generating chat text, and open-source projects quickly replicated its interface…
Verne Robotics, founded by Neil Nie and Aditya Jha, is building an hourly robotics-as-a-service model around its dual-arm Nemo robot, which learns packaging and fulfillment workflows from demonstrations instead of requir…
Microsoft filed a patent application in April 2025 for a machine learning system that would time in-game advertisements to moments when players are least likely to be annoyed, published by the United States Patent & Trad…
Corporate AI deployments largely fail to learn from operational experience, according to a Fast Company analysis that distinguishes memory accumulation from genuine learning and cites Microsoft guidance that gains come f…
Running a large AI model on modest hardware requires reducing the memory it occupies, cutting the calculations it performs, or offloading work to slower hardware, according to a ByteByteGo article that outlines technique…
Cisco submitted the industry's first multi-vendor accelerator entry to MLPerf Inference v6.1, running a distributed inference service across NVIDIA H200 and AMD Instinct MI350X systems over a Cisco Silicon One-based Ethe…
NVIDIA published a tutorial on using the AI data assimilation tools in NVIDIA Earth-2 to incorporate real-time observations into weather forecasting pipelines, covering Score-Based Data Assimilation (SDA) for regional mo…
A project called Mini-AGI is training a dynamic continual learning model on 8GB of VRAM by combining dynamic Mixture of Experts, which adds and prunes experts during training, with batch-1 training on a single continuous…
Multiverse's new paper, "LLM Compression by Block Removal with Constrained Binary Optimization" (arXiv 2602.00161), reformulates block-removal pruning of large language models as a constrained binary optimization problem…