Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
NVIDIA released Nemotron 3 Ultra, a 550B-parameter hybrid Mamba-Transformer model with 55B active parameters, achieving up to 6x higher inference throughput than state-of-the-art LLMs while maintaining accuracy. The mode…
Researchers introduced IRTS-ToolBench, a benchmark of 1,700 questions across 10 task types and 13 domains, to evaluate large language models and AI agents on irregular time series question answering. The benchmark addres…
A developer building a Data Engineering Copilot argues that enterprise AI products should not start as chatbots. Instead, the MVP is a controlled workflow application that takes structured metadata as input and generates…
Andrej Karpathy's open-source 'autoresearch' repo introduces a pattern for structuring AI-agent work that generalizes beyond machine learning. The pattern divides responsibility into a fixed evaluator, a modifiable imple…
A developer built an AI pipeline using FastAPI, Kafka (Redpanda), and Python workers to decouple services and handle bursty workloads. The architecture splits the API from background processing, improving scalability and…
Goodfire researchers introduced predictive data debugging, a method that uses interpretability to predict which behaviors reinforcement learning will amplify or suppress from a preference dataset before training, achievi…
Hermes Agent introduces a 5-layer hybrid routing plugin that enables unlimited skills with near-zero latency. The system uses BM25 and trigger phrase matching for O(1) lookup, supports bilingual Chinese and English queri…
A developer argues that agent memory systems should store events (proof) rather than derived lessons (morals) to avoid superstition and enable revisability. Storing only conclusions leads to unfalsifiable rules that pers…
Developer dmatth1 released quicktok, a fast exact BPE tokenizer written in C++ that runs 2–3.5× faster than bpe-openai and 4–11× faster than tiktoken on CPU. The tokenizer is byte-identical to tiktoken and available as a…
Reinforcement Learning from AI Feedback (RLAIF) is increasingly replacing RLHF in enterprise LLM deployments due to lower cost and higher consistency, but AI feedback fails in domains requiring ground truth, multi-step a…
A developer argues that balanced ternary (-1, 0, +1) could replace binary for AI hardware, citing 20× model compression, 3× inference speedup, and 8× power reduction. Microsoft's BitNet b1.58 demonstrated ternary weights…
A developer compares four dominant alignment methods—RLHF, DPO, IPO, and KTO—for fine-tuning large language models, detailing their mathematical formulations, data requirements, and practical tradeoffs. The analysis high…
A Reuters investigation found that AI-enabled medical devices have caused numerous patient injuries, including strokes, due to malfunctions and misleading guidance. The FDA's adverse event reports for the TruDi Navigatio…
A researcher argues that success in machine learning research depends more on discipline and consistent effort than innate talent, comparing the process to meditation. The author advises beginners to focus on fundamental…
A developer built an offline skin lesion segmenter for mobile devices by optimizing the Med-SAM model with CoreML and ONNX Runtime. The approach enables real-time, privacy-preserving segmentation on iOS and cross-platfor…
WPP Media raised its 2026 global advertising growth forecast to 8.9%, excluding US political ads, nearly two percentage points above its December projection of 7.1%. The revision, driven by stronger Q1 results and a US m…
NVIDIA researchers and collaborators from Duke University introduced a communication-efficient vertical federated learning method that handles limited overlapping samples, presented at CVPR 2023. The approach reduces com…
Whissle Research Team introduced a meta-aware automatic speech recognition (ASR) model, META-1, that outputs transcription and metadata tokens (emotion, intent, demographics) in a single forward pass at ~200ms latency. B…
NVIDIA researchers and collaborators published a review in the Annual Review of Biomedical Engineering detailing the transition of federated learning in healthcare from research to real-world deployment. The paper, autho…
NVIDIA researchers published a study in IEEE Transactions of Medical Imaging investigating whether gradient inversion attacks compromise federated learning safety, finding that such attacks can reconstruct training data …