Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
TypeSafe AI launched Jev, a 40B-parameter "System One" transformer trained exclusively on synthetic data via RLCD that returns typed values with calibrated probabilities instead of free text, in early access on September…
Paradigma released Limit 1B Violetto, a 1 billion parameter dense transformer math-specialist model trained from scratch on under 300 billion curated tokens with a 131k context window, under Apache 2.0 weights and a cust…
Researchers from Carnegie Mellon, MIT, NYU, and Stanford built an AI called Ataraxos that beat four-time Stratego world champion Pim Niemeijer 15 games to one with four draws, using a second neural network that estimates…
Justine Tunney wrote 84 new matrix multiplication kernels for llamafile that make prompt evaluation 30% to 500% faster than llama.cpp on CPU with F16 and Q8_0 weights, according to benchmarks published March 31, 2024. On…
Cactus Compute released Whistle, a 16.9 MB on-device speech recognition model that runs on CPU with no dependencies and loads into the same C++ engine as Needle. Whistle handles transcription in seven languages, word tim…
Fixstars CTO Takuro Iizuka reported at Autonomous Vehicles & AI USA 2026 that an AI-agent-assisted rewrite of an autonomous driving training codebase cut training time 1.7x, with the change spanning roughly 10,000 lines …
Researchers at Meta and the University of Washington trained roughly 1-billion-parameter byte-based student models on up to one trillion bytes distilled from a Llama 3 8B teacher, using an end-of-token (<eot>) marker to …
A flu forecasting model built with Google AI was the most accurate of 39 eligible models at predicting U.S. flu-related hospital admissions during the 2025-26 season, according to the CDC's end-of-season FluSight evaluat…
Moorcheh Edge's 1536-dimensional vector benchmark on two ARM64 edge boards found that at 1 million vectors the Ventuno Q (about 15 GB RAM) was roughly 3× faster on search (269 ms vs 833 ms median) and about 4× faster on …
A developer detailed the memory and GPU VRAM bandwidth bottlenecks that dominate LLM inference engineering, focusing on the KV-Cache's linear growth with context length and batch size. The writeup explains how PagedAtten…
A developer outlined an architecture for forest monitoring that fuses satellite imagery, LiDAR, IoT ground sensors, weather data, and machine learning models into a layered pipeline spanning ingestion, processing, spatia…
DynamicTune, a Show HN project, reports that closed-form weight surgery can transfer trajectory dynamics from a 4B-parameter Qwen3.5 (d=2560) into a 0.8B student (d=1024) and from GPT-2 XL (d=1600) into GPT-2 small (d=76…
A paper submitted to arXiv on 1 October 2026 introduces a Markov chain Monte Carlo (MCMC) sampling algorithm that progressively transforms off-policy traces into more on-policy data for finetuning, allowing supervised fi…
Google has launched the first Project Suncatcher prototype satellite, built with satellite operator Planet and carried to orbit on SpaceX's Transporter-18 rideshare mission, and confirmed it has established contact and i…
A researcher proposes dedicated robotics evaluations that measure the capabilities supported by a robot's hardware rather than its control software, arguing that current policy evals conflate the two. The proposed evals …
TikTok launched Attribution Source Optimization, a capability that lets app advertisers choose which attribution source informs both campaign reporting and optimization, and it is currently available with Adjust, AppsFly…
A systematic study of distillation dynamics finds that existing comparisons between supervised fine-tuning and reinforcement learning vary too many factors at once, making it impossible to isolate the contribution of rol…
A developer behind CardGrade, a phone-photo trading-card pre-grading app, described an architecture in which a border-detection model warps each card side to a flat canvas and slices it into 16 inspection zones, with a v…
A hand-written Blackwell GEMM kernel in CUTLASS's CuTeDSL reached 1401 TFLOP/s on a 4096-cubed problem on a single B200, 97% of cuBLAS, up from 164 TFLOP/s (11% of cuBLAS) in its first correct version, according to the a…
The Allen Institute for AI (Ai2) released Olmo-core 3, an open-source training framework that scales mixture-of-experts large language models to over one trillion parameters while growing the expert pool from 8 to 128 an…