Daily selection of the latest in AI: research, tools, papers and practical tips.
We read it. We curate it. Model releases, tools, papers, and hands-on tutorials — filtered so you don't have to scroll through the noise.
2026-11-18
Tencent's Hy4 Flash is designed for fast, cost-efficient everyday workflows, with internal blind-test satisfaction scores nearing Claude Haiku 5.5.
2026-11-18
Hugging Face Inference Providers now routes language-model requests through OpenRouter, while one access token works across supported open-weight models.
2026-11-18
Perplexity's new research initiative aims to build algorithms, models, and evaluations in the open toward an automated scientist. Applications are open until December 5, with 10,000 GPUs supporting work released as papers, blog posts, model weights, and open repositories.
2026-11-18
Reliable automated checks enable teams to run dozens or hundreds of coding agents in parallel while maintaining confidence in their output.
2026-11-18
ZERO-TIM produces bitwise-identical log probabilities in TPU samplers and trainers by standardizing arithmetic operations and reduction orders, improving stability in on-policy reinforcement learning.
2026-11-18
A new TBATS implementation supports multiple seasonalities, missing values, and promotion or price effects while preventing ARMA errors from amplifying shocks. Early benchmarks show promising results.
2026-11-18
CMU is publishing full computer science classroom lectures on YouTube, beginning with Machine Learning with Large Datasets and Principles of Computing. Courses on Probability and Computing and ML Systems are next.
2026-11-18
Tests indicate that the latest Claude models were trained on protein sequences at scale. They can predict the next amino acid in little-studied bacterial proteins, outperforming all tested open protein language models, but fail when sequences are shuffled.
2026-11-18
Netflix details the engineering decisions behind AI Discover, a TV app tab that uses members' viewing history and new AI-driven approaches to help them find something to watch.
2026-11-18
A new framework enables automatic evaluation and reinforcement learning for agents performing tasks in GTA V, using 103 online tasks and 47 image-based shooting tasks. Tests show the best model reaches 70.9% accuracy, while training on generated navigation tasks improves performance and transfer to new tasks.
2026-11-17
d1.2 and d1.2-ultra process audio, text, and images for tasks such as extraction, passage selection, visual localization, numerical estimation, and multi-option classification. d1.2-ultra ranks first on the Artificial Analysis Decision Index v2.1, with both models available through the Liquid API at usage-based rates.
2026-11-17
Gemini Notebook now transforms uploaded sources into shareable interactive sites, making source-based information easier to present and access.
2026-11-17
The Rubin HPC GPU has more than 50,000 cores to accelerate simulations for AI agents in science, engineering, and analytics, with availability expected in 2028.
2026-11-17
An analysis of Apple's on-device knowledge graph, including its 144-feature XGBoost classifier, retained data, and an open-source tool for inspecting your own graph.
2026-11-17
A case study in probabilistic record matching, data-quality challenges, and evaluation without reliable ground truth, including where LLMs help and traditional statistical methods remain stronger.
2026-11-17
MIT researchers developed WorldParticle, an AI system that learns how particles move and interact across different physical processes. A unified approach could make simulations more flexible for soft robotics, medicine, and materials design.
2026-11-17
Reusable SKILL.md files, project configuration, and automated checks let teams share workflows across GitHub Copilot, Claude Code, and Cursor. Examples cover code verification and data pipelines with clear responsibilities, reliable testing, and human approval.
2026-11-17
A simulated trial shows how analyzing only survivors can create a highly significant treatment effect even when treatment has no impact, revealing how selection bias can undermine otherwise reassuring sanity checks.
2026-11-17
A new serving framework adapts compute allocation and KV-cache management to changing agent workloads. Tests on coding, math, and computer-use tasks reduced mean request latency by up to 4.1× and P95 time to first token by up to 14× compared with NVIDIA Dynamo.
2026-11-17
Mozilla's agentic security pipeline prioritizes high-risk code, tests vulnerability theories in isolated Firefox environments, verifies findings, detects related bug patterns, and generates candidate fixes. Variant scanning and stronger cyber models have helped uncover higher-severity issues, contributing to an approximately 8x increase in bug discovery and fixes from 2025 to 2026.
2026-11-16
Qwen4-Max handles creative and professional tasks across 3D modeling, game development, manufacturing, finance, law, healthcare, and scientific discovery. It can use Blender and game engines for modeling, scene assembly, gameplay logic, rendering, and playable game creation. Smaller Qwen4 models are planned for open-source release.
2026-11-16
A tutorial for building offensive-agent benchmarks graded by deterministic code instead of LLM judges, based on Wiz's 327-challenge arena spanning 26 models, four harnesses, and 59 agents.
2026-11-16
A production report covering more than 100 trillion monthly gateway tokens finds decision models cost about one-seventieth as much per request, while Anthropic wins new teams by a 2:1 margin.
2026-11-16
Quantization errors can affect language-model accuracy during 4-bit training. This technical deep dive compares BF16, MXFP8, and NVFP4 formats and the challenges of using them throughout training.
2026-11-16
Robust evaluations help determine how AI models should be trained and deployed safely.
2026-11-16
Always-on agents manage context and take action through a core loop supported by event inboxes, timers, memory, permissions, and outboxes.
2026-11-16
AI may bring extraordinary gains in intelligence, but economic growth remains constrained by slower-moving sectors. Historical technology revolutions in electricity, communications, computing, transportation, and agriculture have coincided with roughly 2% annual GDP-per-capita growth for nearly 150 years.
2026-11-16
Quark details an end-to-end language-model training stack, from data and architecture to pretraining, context extension, reinforcement learning, evaluation, and release, using about 1,300 H100 hours.
2026-11-16
Summaryformer keeps recent video frames in full and older frames as 96-token summaries, maintaining a fixed-size memory buffer for real-time gameplay. The technical note details its eviction process, rotary position techniques, inference steps, and open implementation.
2026-11-16
openpilot is an open-source robotics operating system that upgrades driver assistance systems across more than 300 supported cars. The repository includes installation guidance, development tools, safety documentation, testing details, and contribution resources.
2026-11-16
A framework called HEELO uses reusable agent-native primitives, where each resource includes an LLM that applies it in context. Across several benchmarks, this approach improves task completion, including Terminal-Bench 4.0 from 37.3% to 55.5% and ToolBench from 55.8% to 75.1%.
2026-11-16
A new method compresses tool outputs into continuous latent tokens instead of raw text for LLM agents. Tests on web-search tasks reduced token use by 70–89% and runtime by 24–56%, while improving task performance.
2026-11-16
Nvidia's driving world model transforms a single camera image into driving video and a bird's-eye view of the surrounding scene.