Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
A developer argues that LLM-based judges cannot serve as build gates in RAG evaluation due to cost and non-determinism, advocating for deterministic metrics like precision@k, recall@k, and lexical overlap proxies. The de…
Google announced AlphaGenome Atlas on Tuesday, a resource that predicts the consequences of every possible single-base variant in the human genome, evaluating 9 billion DNA bases. The tool aims to identify functional non…
Neuroscientists are applying machine learning and AI to decode brain activity, aiming to reconstruct visual experiences in unresponsive patients, a technique they say could bring 'mind reading' to psychedelic neuroscienc…
AWS has launched Ray Serve Deep Learning Containers (DLCs), pre-built Docker images for model inference that bundle PyTorch, Ray Serve, and the GPU stack, addressing the maintenance gap left by TorchServe's discontinuati…
IBM released Granite Time Series PatchTST-FM-r2, a ~385M-parameter time-series foundation model that as of September 8, 2026, ranks as the top-performing zero-shot model under a permissive, commercial-friendly open-sourc…
A developer built a hybrid search system using Node.js's built-in node:sqlite module and SQLite's FTS5, eliminating the need for native dependencies like better-sqlite3. The system combines lexical search with optional v…
A developer released hinterland, a zero-dependency agent memory library for Node.js that commits writes to SQLite before attempting to embed them, ensuring durability on unreliable connections. The library offers lexical…
Elbit Systems, Israel's largest defence contractor, launched Fuse, a new division developing AI-based autonomous military technologies including drones and unmanned systems, operating from Rosh HaAyin, Israel. The unit w…
Google DeepMind released the AlphaGenome Atlas on September 8, a 1-petabyte database predicting the molecular effects of roughly 9 billion single-nucleotide changes in the human genome, making it freely available for non…
Microsoft Research researchers Minseon Kim, Zhengyan Shi and colleagues present FrogNano, a 4B coding agent post-trained exclusively with reinforcement learning on about 1,500 software engineering environments using synt…
A developer's guide to training ML anomaly detection systems argues that evaluation must start with the base rate of attacks, citing Los Alamos National Laboratory data where malicious authentication events occur at abou…
Instacart launched Clementine, an AI shopping assistant integrated into its Marketplace app across the U.S. and Canada on September 9, 2026, that builds ready-to-buy carts from text, recipes, or photos, with orders alrea…
A technical blog post by an author writing a paper on the Chebyshev-Jacobi method argues that, contrary to common belief, Gauss-Seidel is not always faster than Jacobi, and presents a modern convex optimization framework…
Databricks unveiled Adaptive Instructed-Retriever, a retrieval model that takes extra search steps only for complex queries to balance answer quality, latency, and cost. The model, trained with synthetic enterprise data …
SKF, a Swedish bearing manufacturer, has implemented a focused AI solution for smart product defect detection, improving product quality in its supply chain. The company's approach demonstrates that targeted AI applicati…
A developer found that a single LLM evaluation case in their regression suite fluctuated from 5/5 to 2/5 and back without any code or prompt changes, revealing that the eval's measurement was noisy. To address this, they…
Machine learning models are mathematical functions that map inputs to outputs through a forward pass, using features, weights, and bias to make predictions, as explained in an educational article on traditional AI. The a…
Casepoint launched Casepoint IQ, an AI framework spanning its e-discovery, legal holds, FOIA, and investigations platform, integrating new semi- and fully autonomous agents, generative AI assistants, legacy CaseAssist ma…
Harvey and Baseten report that a recursive language model (RLM) harness for M&A diligence improved rubric criteria pass rates by 39.1 percentage points on average across seven models on LAB Diligence tasks, and post-trai…
A technical guide outlines seven engineering techniques for training large language models on limited hardware, including QLoRA, DoRA, and GaLore, which reduce memory usage by quantizing weights or projecting gradients i…