Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
Developer Jon Ready released vllm-ios, a native Swift implementation of vLLM-style continuous batching for MLX on iOS, achieving 88% faster multi-agent inference than llama.cpp at 8 concurrent streams with identical weig…
A technical essay argues that mastering Physics-Informed Neural Networks (PINNs) requires a hybrid background rather than a single major, recommending Computer Science or Mathematics paired with electives in Computationa…
A developer fine-tuning open vision-language models with supervised fine-tuning and GRPO-style reinforcement learning reports three failure modes where training runs looked healthy but accomplished nothing. The developer…
At Hot Chips 2026, Anurag Agarwal and Radhakrishna Giduthuri presented simulations and projections for High Bandwidth Flash (HBF), a new memory technology that combines SSD-like flash storage with HBM-like packaging, aim…
Frink, a pure-Rust GGUF inference engine written by Antonello Frink, reached parity with llama.cpp on Apple Metal decode and passed it on prefill at version v0.12.0, with OLMoE-1B-7B prefill rising from 587 tok/s to 1402…
Ferrox v0.9.1, a pure-Rust GGUF inference engine, shipped today with MoE prefill on Apple Metal 2.4x faster, closing the gap to llama.cpp from 2.62x behind to 1.11x on OLMoE-1B-7B. The update also improved decode perform…
Ferrox, a pure-Rust GGUF inference engine, now runs mixture-of-experts (MoE) prefill on Apple Metal 2.4x faster, reaching 1402 tok/s on OLMoE-1B-7B and closing the gap to llama.cpp from 2.62x behind to 1.11x, with decode…
Engineers at an open-source project have developed Online-SDFT, a prototype that enables a 230M-parameter language model to continually learn on Android devices from delayed user interactions. The system uses hindsight d…
GLM-5.3, a coding model from Zhipu AI, scored 100% in Reinvently's 28-task Ed-o-meter benchmark on August 23, 2026, with a total test cost of $0.28, and completed a root of an Amazon Fire HD 10 in one day after earlier m…
A developer introduced Model Cascade, a technique that uses a cheap model's confidence scores to decide when a large, expensive LLM is needed for classification tasks, potentially reducing costs significantly. The approa…
Inherent Labs, a 12-employee London startup founded by Google DeepMind alumni, announced that its 27-billion-parameter AI agent Faraday outperformed Anthropic's Claude Opus 4.8 and OpenAI's GPT-5.5 at replicating scienti…
Surya and Cameron Franz trained a language model to generate editable p5.brush JavaScript sketches that render into watercolour-style images, using reinforcement learning with a judge model comparing outputs against 581 …
Samsung Electronics expects its Exynos 2700 chip, destined for the Galaxy S27 series, to outperform Qualcomm's Snapdragon 8 Elite Gen 6 and Snapdragon 8 Elite Extreme Gen 6 in CPU, GPU, and AI benchmarks, according to a …
Harvey released Harvey Tenet, its first post-trained model, as a research preview on August 20, 2026, reporting that it completes almost twice as many held-out tasks on its Legal Agent Benchmark (LAB) and 20% more on LAB…
A linear probe on frozen embeddings from Meta's Prompt Guard 2 achieves an AUC of approximately 0.999 on out-of-distribution data, revealing that the model's poor attack detection stems from a miscalibrated classificatio…
MonkeyCode, an open-source AI coding tool, has developed a feedback loop that turns free AI tokens into verified C++ tests, achieving 91% line coverage on a small string utility. The pipeline iteratively generates tests,…
Apple's M4 Pro with 64GB RAM can run vision models like Llama 3.2 Vision, Moondream2, and Qwen2-VL locally with low latency, according to a developer guide. The guide recommends using Ollama to manage models and suggests…
Coding agents fail most often because they guess too much, according to an analysis of the 'Superpowers' methodology, which advocates a multi-gate workflow that treats code generation as the final stage after clarificati…
Researchers at an undisclosed institution developed an AI-driven intrusion prevention system that uses a honeypot feedback loop to reduce false positives from 1.13% to 0.0068%. The architecture, detailed in an IEEE paper…
A blog post explains machine learning concepts, stating that feature scaling works in supervised learning because models learn patterns of X to Y, so normalizing input data (e.g., X/2) maps to normalized output (Y/2). It…