Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
A paper submitted to arXiv on 9 Oct 2026 presents Minions, a dynamic and distributed virtual GPU cloud that aggregates idle Apple Silicon GPUs across macOS devices to run large-scale machine learning inference workloads.…
Poetiq said its Metasystem optimized a custom large language model router in 12 hours, producing configurations it claims beat OpenRouter's Auto Router and TypeSafe's JEV on its own tests. The company said one engineer s…
Multiverse Computing released Comet 60B on October 15th, a compressed version of NVIDIA's Nemotron 3.5 Super VL that roughly halves the model's total parameter count while retaining text, image and video input, with down…
MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL) ran a study in which 13 AI agents nicknamed Eko, each powered by Claude Code and given a virtual lobster body, were left to roam a bounded virtual isl…
A new research paper posted to arXiv on October 12, 2026 introduces GRAFT (Gradient Routing Adapter Fine Tuning), a technique that uses a deletable "forget adapter" to curb reward hacking in AI models, with hacking rates…
Mistral released Timestral, a model targeting time-series and temporal reasoning workloads, according to the company's announcement and GitHub repository. Mistral simultaneously acquired New Material (newmaterial.ai), a …
Liquid AI released d1-ultra on October 15th, a multimodal model that takes text, images or both and returns probabilities for defined decisions in one pass, priced at $0.042 per million input tokens with no charge for ou…
A developer built AI Grand Prix, a browser-based demo that races five small open-weight decision models on a closed circuit, where each model receives a JSON state 20 times per second and returns probabilities over five …
A developer outlined a four-layer architecture for AI- and IoT-based water infrastructure monitoring, separating sensor, connectivity, data processing, and application layers to keep systems maintainable. The approach us…
Researchers released GhostFaceMPC, described as the first end-to-end privacy-preserving face recognition system to evaluate a deep convolutional neural network extractor and matching under secure multi-party computation,…
ModelFront announced OpenLoc, an open translation management system built for AI that works with any LLM and can be managed through MCP, with code to be released on GitHub on Monday under BSL 1.1. DeepL also launched an …
Liquid AI released d1-ultra, a multimodal decision model it says outperforms all other decision models on the Artificial Analysis Decision Index by at least 8 points and beats Jev 1.14 across all tasks, based on 112 uniq…
Apple's MLX framework, in a release published today, runs the 816-billion-parameter Muse Spark Open model at over 36 tokens per second on a cluster of four Mac Studios connected via the RDMA over Thunderbolt driver. Appl…
Researchers at the University of Amsterdam and TNO have proposed a differentiable soft-pruning method that turns a single pretrained transformer encoder into an elastic model runnable at any compute budget from 5% to 70%…
Mechanistic interpretability researcher Chris Olah's collaborator reports that the canonical evidence for polysemanticity and superposition in neural networks was misread, and that what looked like arbitrary concept pack…
A developer outlined a practical approach to industrial AIoT predictive maintenance, describing a six-step pipeline — collect, validate, establish baseline, detect, alert, and evaluate — for turning machine sensor readin…
Bhaskar Gurram released numguard, an open-source tool that tests numerical array code for edge cases including NaN, ±inf, overflow, subnormals, cancellation and float16/32/64 precision, comparing results against high-pre…
OpenAI released a claimed Lean proof of L = RL = BPL in its GitHub math repository, according to complexity theorist Lance Fortnow's Computational Complexity blog post on formalizing complexity-class results. Fortnow had…
Researchers at the University of Illinois Urbana-Champaign (ulab-uiuc) posted a paper on arXiv on 14 Oct 2026 introducing CoPE, a language-agent training method that adds multi-step action targets to agentic reinforcemen…
A paper submitted to arXiv on 14 Oct 2026 argues that first-price auctions used to allocate user prompts among bidding AI agents suffer less welfare loss than the general worst case, because agentic markets tend to have …