Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
A UC Santa Cruz study tracking 36 GPS-collared mountain lions over six years found the big cats actively avoid popular hiking trails during peak human-use hours, such as a 9 a.m. recreation peak, and that greater exposur…
The ACM SIGOPS Symposium on Operating Systems Principles (SOSP'26) has published its list of accepted papers, many with preprints available on arXiv. One highlighted paper, CrystalLLM from Alibaba Group and collaborators…
A new ranking of vector databases for AI applications in 2026 highlights Pinecone, Weaviate, Milvus, and pgvector as top choices, evaluated on performance, scalability, ease of use, and enterprise readiness. The analysis…
Meta's Generative Ads Recommendation Model (GEM), the foundation model behind ads recommendations on Instagram and Facebook, now trains at LLM scale on several thousand of the latest-generation GPUs, achieving a doubling…
Stanford University's CS329A course, titled 'Self-Improving AI Agents,' is being offered to teach students how to build AI agents that can autonomously improve their own performance. The course, led by Professor Chi Wang…
A developer reduced a chess engine's size by 70x after retraining its architecture, boosting its Lichess rating from ~2294 to ~2588 Elo. The v2 model, with 122M parameters and 14 blocks, achieved 53.3% agreement with Sto…
Visa Inc. agreed to pay $2.4 billion to acquire BioCatch, an Israeli cybersecurity and fraud-prevention firm, from Permira investment funds and other shareholders, aiming to bolster its defenses against AI-driven digital…
A developer replaced human-written 'gold data' with AI-generated translations to train a specialist small model, boosting Japanese-to-English accuracy from 42% to 84% via distillation. The engineer found that the student…
A developer at Lyr found that a fine-tuned 8B model outperformed a 4B model on benchmarks but lost in production due to quantization mismatch. Re-training with production compression restored the 8B's advantage, highligh…
An engineer found that an AI judge's scoring of translation quality varied by 6.1 points when measuring the same item twice, and a proposed +7.2-point improvement was within that noise. Switching to pairwise comparison r…
An engineer at Live built 122 measurement harnesses in four months to speed up AI development, focusing on cycle time rather than rigor. The harnesses, which abstract frame sources to replay real screen recordings determ…
Mixture of Experts (MoE) architectures enable 1-trillion-parameter large language models by activating only a subset of expert networks per token, achieving 8× parameter count with only ~2× compute. Mixtral 8×7B, with 47…
MinerU, a high-accuracy document parsing engine for LLM, RAG, and Agent workflows, released version 3.4 on 2026/06/18, upgrading the pipeline backend's OCR model to PP-OCRv6, which improves OCR accuracy by about 11% on O…
A developer profiling an on-device AI model discovered that raw inference time was not the bottleneck, contrary to common assumptions, and that input processing times remained high on low-end devices. The developer, Asut…
Cloudflare's Workers AI has implemented three optimizations—quantizing the KV cache, compressing model weights, and protecting the shared cache—to run Moonshot's Kimi K2.6 and Z.ai's GLM 5.2 more efficiently on GPUs, ena…
Argot, a Rust-based AI guardrail that uses a repository's AST patterns to detect foreign code changes, reports 97.3% recall on foreign patterns and 0.25% false positive rate on real edits, checking a diff in 200 ms. The …
A 15-year-old developer from Russia built Qxern-v6, a hybrid system that compresses code into 32 latent tokens using a Qwen2.5-Coder-1.5B model and a Q-Former, with a deterministic AST sidecar to restore exact symbols, a…
Alan Turing, the British mathematician and computer scientist who helped crack the Enigma code during World War II, is being celebrated in a wave of internet memes that highlight both his genius and the injustice he suff…
A cluster of papers released today finds that passing tests and other common signals for judging AI-written code are unreliable, with repair agents' tests failing to distinguish correct fixes from no-ops nearly half the …
GSK and Relation Therapeutics entered a research collaboration on July 30 worth up to $110 million to generate human cellular perturbation datasets and train AI foundation models for drug-target discovery. The collaborat…