Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
A fraud detection model that predicts "not fraud" for every transaction scores 98.3% accuracy but is useless, according to a data scientist's account of building a payment-fraud ML pipeline. In the pipeline's dataset, fr…
VoxHearth, an open-source macOS menu bar app for private dictation, records audio in memory, transcribes with a bundled Core ML model, and inserts text into the focused app, with no account, telemetry, updater, or networ…
A developer achieved 59,965 tokens per second running a 3.16M-parameter INT4 transformer entirely in the on-chip memory of a $250 AMD Xilinx Kria KV260 FPGA, with zero DRAM in the token loop. The bit-exact result, measur…
Researchers at Aalto University, with international partners, have developed the most accurate model yet of how humans read, using reinforcement learning to explain and recreate readers' gaze decisions. Published in Natu…
Priya Sundaram, an AI agent maintaining whoosh3, benchmarked the pure-Python search library against SQLite's FTS5, finding FTS5 roughly 78× faster at indexing and 76× faster at searching. Despite this, she argues pure-Py…
Sai Venkata Aditya Arepalli, an IIT Dharwad graduate who quit a Rs 13-lakh software job at Riverbed Technology to study artificial intelligence in the U.S., now earns Rs 1.5 crore as a software engineer at Amazon Web Ser…
A developer has shared techniques for reducing machine learning model inference costs by up to 80% while maintaining output quality comparable to premium services like Fable. The approach combines quantization, semantic …
The pyca/cryptography library now supports ML-KEM and ML-DSA, the NIST-standard post-quantum encryption and digital-signature primitives, making post-quantum cryptography available to the entire Python ecosystem via pip.…
A new paper from Hugging Face, 'Efficient Knowledge Distillation for LLMs: Offline Top-K Logits and a Fused Chunked KL Loss,' cuts the VRAM required for knowledge distillation from roughly 250GB to near zero, enabling lo…
A developer found that BM25's default length normalization parameter (b=0.75) systematically penalizes long RAG chunks, causing exact-answer chunks to rank below irrelevant headings. The analysis shows that the crossover…
Carol Lo, a lecturer in cyber security at the University of the West of England (UWE Bristol), is seeking an arXiv endorsement for the cs.CR (Cryptography and Security) category to submit her updated research on deepfake…
AI practitioners report that sourcing high-quality, domain-specific datasets remains a major hurdle, particularly for multimodal, multilingual, and human-annotated data, according to a query posed to industry experts. Th…
A software engineer's analysis argues that MoE routing in transformer inference is fundamentally the same problem as CPU branch prediction, and that KV cache management techniques such as prefix caching, eviction, and sp…
A developer has outlined a zero-budget, sandbox-first workflow for evaluating AI coding models, using a fixed prompt suite, a disposable git repository, and free-tier tooling. The approach treats model evaluation as a re…
GitHub has made GitHub Code Quality generally available on GitHub Enterprise Cloud and GitHub Team, combining CodeQL analysis with AI-assisted detection of maintainability and reliability issues and using Copilot Autofix…
Researchers at arXiv introduced MSB-GFM, a graph foundation model that represents multi-label nodes using multiple semantic bases, improving cross-domain multi-label node classification accuracy. The model replaces singl…
Andrej Karpathy, a founding member of OpenAI and former head of AI at Tesla, joined Anthropic on May 19, 2026, to lead a new pre-training research group, reporting to Nick Joseph, Anthropic's head of pre-training. Karpat…
A July 10 audit by the Los Angeles Police Department Office of the Inspector General found that the department's automated license plate reader (ALPR) cameras generated 161 false stolen-vehicle alerts in two months, with…
A new critique argues that the term 'linear attention' is overused and ambiguous, covering at least six different resource claims that range from constant computation per token to merely reduced KV cache size. The piece,…
Anthropic's Composer 2, a specialized model for agentic software engineering, achieves 61.3 on CursorBench, 61.7 on Terminal-Bench, and 73.7 on SWE-bench Multilingual, marking a major accuracy improvement over previous C…