AI moves too fast and lies too often. We track frontier labs, analysts who actually know, and the open-source firehose — then weight what survives. No newsletter pad, no affiliate slop.
OpenAI's GPT-5.5 system card details predeployment red-teaming across cyber and biology, a Preparedness Framework evaluation, and an explicit methodology for treating GPT-5.5 Pro as the same underlying model.
Simon Willison's llm CLI tool releases version 0.31 with GPT-5.5 support, a new verbosity parameter for GPT-5+ models, configurable image detail levels, and async registration for custom OpenAI model configs.
Berkeley researchers prove that word2vec learns in discrete, sequential rank-incrementing steps, and that the final representations are exactly the top eigenvectors of a matrix defined by corpus co-occurrence probabilities.
A Berkeley researcher proposes a third RL paradigm beyond TD and Monte Carlo that exploits triangle inequality in goal-conditioned settings, avoiding the error accumulation that blocks Q-learning from scaling to long-horizon tasks.
A NeurIPS 2025 paper from Berkeley introduces a framework that evaluates and optimizes imaging hardware using information content alone, matching end-to-end learned methods while requiring less memory and no task decoder.
Berkeley researchers introduce two algorithms that reframe LLM interaction discovery as a sparse recovery problem, enabling feature, data, and mechanistic interpretability at scales previously computationally infeasible.
Berkeley researchers introduce GRASP, a gradient-based planner that addresses exploding gradients, non-greedy local minima, and vision-model brittleness in long-horizon trajectory optimization.
Alibaba's qwen-mt-turbo update builds translation-specific reinforcement learning on top of Qwen3, adding terminology injection, domain prompts, and translation memory at prices well below general-purpose frontier models.
Qwen introduces Group Sequence Policy Optimization, a new RL algorithm that eliminates the instability and infrastructure overhead that blocked GRPO from scaling to large Mixture-of-Experts language models.
Alibaba's Qwen team releases a 20B MMDiT image foundation model purpose-built for complex text rendering and precise image editing, with state-of-the-art results across six public benchmarks.
Built on the 20B Qwen-Image model, Qwen-Image-Edit handles both high-level semantic and low-level appearance editing while adding precise bilingual text editing, extending the model's text rendering expertise to image modification tasks.
Alibaba's Qwen team releases Qwen3Guard, a safety guardrail model in two variants — one generative, one streaming — with three-tier severity classification and support for 119 languages, available in 0.6B, 4B, and 8B sizes.
Apple ML Research researchers show that information accessible via a model's top-k logits can reveal image content that model owners expect to be inaccessible, sometimes matching what is exposed by direct residual stream projections.
A new benchmark from Apple ML Research probes LLMs on four distinct context understanding tasks across nine datasets, finding that pretrained dense models struggle with nuanced contextual features and that 3-bit quantization causes variable performance drops.
Apple presents five research highlights at ICLR 2026 in Rio de Janeiro, covering parallel RNN training, SSM length generalization, unified image understanding and generation, real-time 3D scene synthesis, and a new protein folding approach.
Apple's ParaRNN framework uses Newton's method to parallelize nonlinear RNN training, achieving a 665x speedup over sequential approaches and enabling billion-scale RNN language models for the first time.
Apple ML researchers model scene dynamics by working directly in a compressed motion embedding space derived from large-scale trajectory data, enabling efficient motion generation conditioned on text or spatial inputs.
NVIDIA NeMo RL applies FP8 precision across both generation and training phases of reinforcement learning, closing accuracy gaps via importance sampling and adding 48% total speedup when KV cache and attention are also quantized.
NVIDIA has integrated the Muon higher-order optimizer into Megatron Core and NeMo, showing minimal throughput loss versus AdamW on GB300 hardware while enabling training at thousands of GPUs.
A first-place finish in the March 2026 Kaggle Playground churn prediction competition came from a four-level stack of 150 models selected from 850 runs, generated by GPT-5.4 Pro, Gemini 3.1 Pro, and Claude Opus 4.6 in a human-in-the-loop workflow with GPU-accelerated execution.
The latest NVIDIA FLARE API splits federated learning adoption into two steps: a minimal client API that adds federation to existing training scripts without restructuring them, and portable job recipes that run unchanged from simulation to production.
NVIDIA outlines how its Blackwell platform and NIM microservices support DeepSeek-V4's million-token context requirements, with initial GB200 NVL72 benchmarks and deployment paths via SGLang, vLLM, and hosted endpoints at build.nvidia.com.
A Hugging Face analysis argues that AI-powered vulnerability discovery favors open, distributed systems over closed ones: openness distributes detection and patching across communities, while closed codebases concentrate both the attack surface and the remediation bottleneck.
Researchers from TII UAE built QIMMA, the only Arabic LLM leaderboard combining quality validation, native content, and code evaluation. A two-stage pipeline of LLM scoring and human review revealed recurring quality failures across widely-used Arabic benchmarks.
A step-by-step demo shows Gemma 4 handling speech input, autonomous webcam activation, and spoken output on NVIDIA's Jetson Orin Nano Super using llama.cpp, Parakeet STT, and Kokoro TTS — no keyword triggers, no hardcoded logic.
A practical walkthrough of how to host Transformers.js models in a Chrome extension background service worker under Manifest V3, covering runtime separation, messaging contracts, model caching, and the agent tool-execution loop.
DeepSeek-V4 combines two new attention mechanisms with agent-specific post-training to reduce KV cache memory to roughly 2% of a standard grouped-query-attention architecture, targeting long-horizon agentic workloads over chat.
Microsoft Research and collaborators introduce ADeLe, a framework that characterizes both benchmarks and LLMs using shared capability scores, achieving ~88% prediction accuracy on unseen tasks for models like GPT-4o and LLaMA-3.1-405B.
Researchers behind the New Future of Work Report 2025 discuss the intentionality required to build a future where people flourish, the gap between efficiency and the work future worth wanting, and why AI's role as tool versus collaborator is not a semantic question.
The fifth annual New Future of Work report from Microsoft Research documents generative AI's entry into workplaces — faster than previous technologies, with measurable productivity gains for some, declining entry-level hiring, and significant gaps by gender, income level, and language.
A Microsoft Research podcast episode brings together a sustainability scientist and an optimization researcher to examine AI's actual climate footprint, the local infrastructure concerns from datacenter expansion, and where AI optimization tools can genuinely help.
An open-source framework from Microsoft Research automates the selection between RAG and fine-tuning approaches, plans adaptation strategies against real deployment constraints, and replaces manual hyperparameter search with a budgeted refinement loop.
A research experiment built with NYU uses an Executive LLM to steer multi-party AI conversations toward targeted skill assessment, with AI scoring accuracy matching human expert agreement rates in a 188-person study.
A new open-source model from Google Research generates realistic 3D neuron geometries from point cloud flow matching, improving connectomic reconstruction accuracy in a way that would save 157 person-years of manual work at mouse-brain scale.
Google's Simula framework decomposes synthetic dataset creation into independently controllable axes — diversity, complexity, and quality — and has already been deployed in Gemma safety models, Android scam detection, and spam filtering.
Google Cloud's ICLR paper introduces a structured memory framework that distills both failed and successful agent trajectories into transferable reasoning strategies, beating memory-free baselines by up to 8.3% on web benchmarks.
Google's Auto frame feature uses 3D scene reconstruction and generative inpainting to re-render photos from a different viewpoint, fixing parallax and perspective distortion without a reshoot.
Google DeepMind's Gemma 4 family spans a 2B mobile model to a 31B dense model, supports 140+ languages, adds 256K context for larger variants, and ships under a fully permissive open-source license.
Google DeepMind's upgraded embodied reasoning model improves spatial reasoning, success detection, and multi-camera understanding — and gains a new industrial inspection capability co-developed with Boston Dynamics.
Google DeepMind's new text-to-speech model scores 1,211 on the Artificial Analysis TTS Elo leaderboard, introduces natural-language audio tags for vocal direction, and watermarks all output with SynthID.
Accenture, Bain, BCG, Deloitte, and McKinsey gain early access to Gemini models and direct DeepMind technical engagement in a push to close a significant gap between enterprise AI ambition and production deployment.
A new distributed training architecture from Google DeepMind uses asynchronous compute islands to train large models across distant locations — more than 20x faster than conventional synchronization, with self-healing fault tolerance.
Same story across 30 outlets is one story. We embed, dedupe, and surface the canonical thread — not the press-release telephone game.
OpenAI's own blog outweighs 50 reblogs. SemiAnalysis on chips outweighs general trade press. Tier 0 to Tier 5. Honest scoring.
We tag every story: scoop, official, analysis, underrated. You see the shape of coverage at a glance.
Every claim links back to its source. Every source has a tier. You can audit our reasoning end to end.