Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
A developer built Phocinae-Largha-150M-v1, a 144.3M-parameter typed decision model that answers yes/no, pick-one and 1-10 questions in a single forward pass at 18.6 ms p50 on GPU, scoring 0.797 on English typed-decisions…
A Nature paper introduces 'byteification', a two-stage conversion procedure that retrofits existing subword-tokenized large language models into byte-level models with minimal extra training. The resulting byte-level mod…
Alexium International has made its microencapsulated phase change material (mPCM) solutions available to US data center developers, contractors and engineering firms as thermal buffers integrated into walls, ceilings and…
A developer built a minimal Lewis referential game in which two tiny neural networks with no shared vocabulary invent a communication code, achieving 0.970 accuracy on seen objects and 0.937 on held-out objects across 8 …
Liquid AI's D1-Omni-600M, a roughly 587-million-parameter multimodal decision model that returns probability scores instead of generated text, ran fully loaded under 2GB of VRAM on a single consumer GPU in hands-on testi…
Google's open-weight EmbeddingGemma 2 model embeds text, code, images, video, and audio into a single shared vector space using a 270 million parameter Gemma-based text backbone plus optional 170M vision and 300M audio e…
Apple researchers Jiatao Gu, Tianrong Chen, Ying Shen, David Berthelot, Shuangfei Zhai and Josh Susskind published Normalizing Trajectory Models (NTM) in October 2026, a generative modeling method that treats each revers…
AI agents in production can accept in-the-moment user corrections but typically fail to convert them into reusable behavioral guidance, so the same mistakes recur across users and sessions, according to Reflexio's analys…
Astronomer Brice Ménard and a team of AI assistants built a full-sky ultraviolet map by combining observations from several telescopes, with about one third of the finished map being prediction rather than observation. T…
A team of researchers submitted an arXiv paper on 24 April 2026 proposing Action-Conditioned World Models, adapting the LeJEPA framework to physiological time-series to model cardiac disease progression as a transition v…
A paper submitted to arXiv on 26 Sep 2026 proposes describing deep-learning training as the ordered acquisition of "quanta" — reusable computations acquired suddenly and binary-activated across examples to reduce loss — …
A Stanford team led by Jiacheng Miao, Joe R. Davis, Yaohui Zhang, Jonathan K. Pritchard, and James Zou published Paper2Agent in Nature on September 16, 2026, an automated pipeline that converts research papers into MCP a…
Diogo Almeida, a former OpenAI researcher who helped build ChatGPT and co-invent RLHF, left the company and co-founded TypeSafe AI with Erik Gafni and Sasha Sheng, raising $40M. The startup emerged from stealth this mont…
Researchers introduced DeCoPrune, a KV-cache pruning method for autoregressive video diffusion that uses denoising consistency to compress the continuously growing cache. The work targets the memory cost of streaming vid…
OPSWAT will fully integrate MetaDefender Aether into MetaDefender Cloud starting October 1st, turning the platform's sandboxing step into a five-layer detection pipeline that combines threat reputation, AI-driven static …
Aleph Alpha released Kolibri 1, a 78.1B-parameter mixture-of-experts model with 3.46B active parameters, under Apache 2.0 on Hugging Face on 3 October 2026, making it the only one of three open-weight models announced th…
A developer built BirdBuddy, an offline bird call identifier that runs entirely on-device, combining Cornell's BirdNET ONNX model for species classification with Piper neural text-to-speech for spoken results. The projec…
Liquid AI released Open d1, two open-weight multimodal models in its d1 decision model family: d1-3B, which reads text and images, and d1-omni-600M, which reads text with an image or text with audio. Neither model writes…
NVIDIA's Seattle Robotics Lab, working with the NVIDIA Isaac engineering team and contract manufacturer Foxconn, has demonstrated that robots can assemble GB300 tester trays, tackling busbar assembly and multi-connector …
A cluster of 2025-2026 studies led by the InnoEval evaluation framework found that frontier AI models recover the central idea of a paper from its pre-publication reference list only 3-15% of the time across 643 papers, …