Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
A developer deployed Google's Gemma 4 E2B quantization-aware-trained (QAT) checkpoint on a 4 GB GTX 1650 Ti laptop GPU, shrinking the model from 9.5 GiB in bfloat16 to a 3.35 GB Q4_0 GGUF file with only 1.31 GiB resident…
A developer extended Sebastian Raschka's GPT-2-style code from the book "Build a Large Language Model (from Scratch)" to add mixture-of-experts support and trained a 446M-parameter model with 220M active parameters from …
Google Research, in collaboration with HHMI Janelia Research Campus and other partners, has published a complete wiring diagram of the adult male fruit fly brain and central nervous system, containing roughly 166,000 neu…
IBM and NASA released Prithvi-Lunar, an open-weight foundation model trained on two million lunar surface tiles, published on HuggingFace under Apache-2.0 alongside SomBench, the first unified ML-ready lunar dataset unde…
A developer argues that SHAP and other post-hoc explainability tools cannot account for the autonomous decisions and tool calls made by agentic AI fraud-detection systems, leaving an explainability gap for regulators and…
A Hacker News submission asks why AI-based decompilation remains unsolved despite the availability of thousands of source-executable pairs for training and the compiler being a parameter AI can identify and weight. The p…
Sage has embedded domain-specific AI models directly into its accounting platforms—Sage Accounting, Sage Intacct, and Sage X3—to automate ledger reconciliations, close-cycle operations, and transactional accuracy. The ar…
Feyn released MultiMatte, a promptable image background-removal model built on Meta's SAM 3 that uses low-rank fine-tuning to modify 19.49M of SAM 3's 860M parameters, or 2.27% of the model weights. MultiMatte scores 0.9…
Nokia and Orange France completed a live optical network tomography trial using algorithms from Nokia Bell Labs, which Nokia said can pinpoint physical fiber wear, failing amplifiers, and bad configurations with more acc…
A technical breakdown of how tokenizer training data skewed toward English — with English and code making up 95% of Llama 3's training data — creates measurable cost, context-window, and accuracy penalties for other lang…
A new paper, "What fits (into few tokens) doesn't overfit: Compression and generalization in ML research agents," argues that LLM-based research agents avoid overfitting on heavily reused benchmarks because the agents' l…
ByteDance's Seed team, working with University of Hong Kong researchers, adapted DeepSeek's Group Relative Policy Optimization (GRPO) technique into a visual-generation framework called DanceGRPO, reporting benchmark gai…
Google released WeatherNext 3, an AI weather model that ingests 12 recent hourly frames from an 11-channel geostationary satellite mosaic alongside ECMWF analysis fields to refresh forecasts hourly at up to 5-kilometer r…
A developer built a supervised machine learning pipeline that classifies Apple-related tweets as negative, neutral, or positive using the Apple Twitter Sentiment dataset. The system cleans raw tweet text, converts it int…
A developer outlined a six-layer AIoT architecture for turning industrial sensor data into operational decisions, spanning physical assets, sensors, connectivity, a data platform, AI models, and applications. The pipelin…
A developer built a system that connects a biologically accurate fly brain simulation to a Pong game, finding that the brain did not learn to play despite producing behavior that appeared alive. The experiment was design…
Department of Defense Chief Information Officer Kirsten Davies said at the Billington Cybersecurity Summit in Washington, D.C., on Wednesday that the Pentagon must shift from adding personnel to deploying technology, inc…
A paper by Federico Barbero and co-authors, revised 13 May 2025 as v3 on arXiv, argues that Rotary Positional Encodings (RoPE) in Transformer-based large language models are not primarily useful for decaying token depend…
IBM and NASA have developed an open source AI foundation model of the Moon, trained on a lunar dataset aggregating more than 30 spatially-aligned layers from nine instruments across four missions, IBM announced. A techni…
Traditional machine learning models are limited to static input-output mappings and cannot handle multi-step, dynamic, or action-requiring problems, according to an article explaining how combining trained models with ag…