Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
Tuning Jev, TypeSafe's structured decision model, as a safety judge on the BeaverTails dataset produced a worse result than Part 1's quality-judge run: escalating every flagged case to a reasoning model drops F1 by two f…
Anthropic member of technical staff Yifan Yang reported that a CUDA-core fp32 GEMM kernel on an Nvidia V100 GPU reached 10.7 TFLOPS, or 88% of cuBLAS's 12.1 TFLOPS, using only two optimizations over a naive implementatio…
FreeCodeCamp.org published a 1-hour YouTube tutorial that walks viewers through building, pre-training, and fine-tuning a 25-million parameter language model on a standard CPU. The course covers hybrid linear/sparse atte…
MLCommons released MLPerf Inference v6.1 on September 16, adding two new benchmarks: an End-to-End RAG test that times a full retrieval pipeline (embedder, vector database, reranker, and LLM stages) and an Edge Agentic b…
Researchers at the Weizmann Institute of Science in Israel, led by professor Michal Irani, developed an AI model called Brain-IT that reconstructs and predicts images a person is viewing from functional MRI brain scans, …
A developer built GroundSignal, an open-source, local-first field tool that uses phone motion sensors and the TabPFN tabular model to generate private, uncertainty-aware reports on how a sidewalk or path physically feels…
A zero-cost reweighting technique that keeps the attention sink at token position 0 exact during activation-whitened SVD compression lowered the loss increase in 18 of 20 model-and-budget cells across six model families …
A Hacker News user posting as 5CEOs claims to have solved all 400 puzzles of the ARC AGI 1 benchmark in around 2 minutes on a Mac M1 without using an LLM. The post, submitted 16 minutes before the snapshot, states "No pr…
Google's AI Edge Team open-sourced ML Drift, a cross-platform on-device GPU compute engine for AI/ML inference, under the Apache 2.0 license. ML Drift abstracts OpenGL ES, OpenCL, Metal, and WebGPU hardware APIs, serves …
A benchmark of four local System One decision models and the hosted Jev 1.13.0 reference on 18 synthetic customer messages found that only Jev 1.13.0 classified all 18 correctly, while the local models scored 15/18, 13/1…
A community research effort claims it beat a prior OpenAI benchmark by a factor of 500,000, with verified findings published on a short timeline, according to a research summary. The work targets a 124M-parameter GPT-2 v…
An AI agent named Max rebuilt a small feedforward neural network, originally authored by its human as a KNIME workflow, as roughly 80 lines of dependency-free Python. The script, gehirn_mini.py, trains a 22-16-22 network…
Bright reintroduced its STAR Fuzzer, an AI-guided structural fuzzer originally released in 2018 as NeuraLegion's NexPloit, after running it for two days in an automated pipeline against multiple open-source libraries and…
A developer built Acoustic Atlas, an offline-first Progressive Web App that uses Google's open-weight YAMNet audio classifier via TensorFlow.js to score how "alive" a location sounds. The app processes microphone audio i…
Operant Labs founder reports that swapping both the embedding model and the comparison text raised first-message task-cluster matching from 21.1% to 92.1% on 92 demo-workload cases, so the run cannot isolate which change…
Samsung Research's LittleBit quantization technique compresses Llama2-13B to under 0.9GB and Llama2-70B to under 2GB by binarizing low-rank latent factors to roughly 0.1 bits per weight, a nearly 31x memory cut, accordin…
Researchers submitted RoboJEPA, an 8B-parameter latent world model built on the Joint Embedding Predictive Architecture (JEPA) and trained on a dataset spanning 12 robotic embodiments, to arXiv on 7 Oct 2026. The team re…
A paper submitted to arXiv on 7 October 2026 by Aleksandar Armacki shows that clipped decentralized SGD (DSGD) achieves order-optimal convergence rates under heavy-tailed noise for smooth non-convex costs with bounded p-…
SiMa.ai published a six-page migration guide for moving machine learning models and applications from an NVIDIA/CUDA environment to its Physical AI platform on the Modalix MLSoC. The guide details a five-step model migra…
Researchers at ETH Zurich's Soft Robotics Lab turned an off-the-shelf Wuji Technology robotic hand into a self-contained walking robot that moves on its fingertips, using reinforcement learning trained in simulation to o…