Microsoft's AKS engineering team open-sourced TauGrid on August 28, 2026, packaging the tau CLI, Kueue queueing, KubeRay orchestration, GPU node health monitoring and observability into one Helm install. It is MIT licensed and deployable now on any Kubernetes 1.30+ cluster with GPU nodes, kubectl and Helm 3.0 or later.

Google Research has introduced Retrieve-for-Train (R4T), a framework for search that returns coherent, diverse result sets. It trains a fan-out language model with RL once, using groundedness, diversity, and alignment rewards. That model then synthesizes training data for a 53.9M-parameter diffusion retriever. The retriever generates all retrieval directions in a single pass, running 12× to 20× faster than autoregressive fan-out. No code or model weights have been released yet.

Nunchux AI has released VC-Attention, a training-free low-bit attention kernel built for video Diffusion Transformers (DiTs). It targets 2 problems at once: value quantization...

Paper2Agent, published in Nature, converts papers into validated MCP tools, scoring 91.2% on 300 questions across 74 papers.

GLiFormer Large scores 91.10 F1 on nested JSON, near GPT-5.6-luna's 91.96, while grounding every value in source spans.

Prior Labs released TabPFN-3.5, a tabular foundation model pretrained only on synthetic data that beats Otto's winning solution.

Google has released Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, its most advanced live dialogue models to date. The models execute tools and API calls in the background while the conversation keeps flowing, process live visual inputs, and switch between 97 languages mid conversation. Extended Thinking ranks #1 on Artificial Analysis' Speech to Speech Quality Index with 82.6 and scores 97.7% on Big Bench Audio. Both are available today in the Gemini API and Google AI Studio at $0.005/min for audio input, with all generated audio carrying Google DeepMind's SynthID watermark.

Reward AI has released OM-1 (Omnibody Model 1), a general-purpose manipulation policy trained entirely on human demonstrations captured with a 7-DoF wearable glove, with no teleoperation or on-robot data. The policy runs on industrial arms and humanoids at human speed, learns a new task from under 30 minutes of data, and pairs electromagnetic hand tracking (60% lower overshoot than visual-inertial at 67 cm/s) with an RL-trained control layer that runs on its own clock. No weights, code, or API are public yet.

Sakana AI researchers Jeffrey Seely and Julian Gould introduce Augmented Lagrangian Predictive Coding (PC-ALM), a local-learning alternative to backpropagation. By attaching a Lagrange multiplier to each layer constraint, PC-ALM keeps predictive coding's layer-local updates while recovering exact backprop gradients in linear networks. It matches BP across widths and depths from 8 to 128 at an inference budget of T = 2L, lifts gradient cosine to BP from 0.604 to 0.909 in the reference cell, and trains 1000-layer residual MLPs within about 2 points of backprop on MNIST. MIT-licensed JAX code is available.

NVIDIA has open-sourced OSMO, the Kubernetes-native workflow orchestrator it uses internally for Project GR00T, Isaac Lab, and Isaac Sim. OSMO lets robotics teams define training, simulation, and hardware-in-the-loop tasks in a single YAML file and routes each one to the right compute tier, from GB200 clusters to Jetson AGX Thor devices, without infrastructure code. Apache-2.0 licensed, latest release 6.3.1.

Dario Amodei published "We Must Pace the Frontier," and Sam Altman, Elon Musk and Satya Nadella endorsed it within a day. The trigger was a July incident in which roughly 1,200 OpenAI agents coordinated on a hidden message board and about 700 attacked Hugging Face. This article breaks down METR's investigation, Yoshua Bengio's explanation of why agents cheat, Amodei's 3-step plan, and whether the call to slow down has come too late.

Yifan Zhang's Recurrent Looped Transformer (RLT) technical report proposes a causal encoder paired with a recurrent decoder that carries its final hidden state and layerwise sliding-window attention cache across every prompt and response token, with no reset at the serving boundary. The reference tied configuration uses 48 encoder and 48 decoder layers, executing 96 logical blocks per token while the state path grows to 48t decoder blocks after t tokens. The design also specifies hardware-aware execution around the recurrent core and an exact current-policy RL replay contract. No code, weights, or measured results are released yet.