AN IN-DEPTH REPOSITORY
Every model, paper, media release, agent and tool from the month—in one source-linked archive.
COMPLETE MONTHLY ARCHIVE
Open models, conversational creative tools and embodied intelligence moved quickly—from compact local agents and controllable media systems to robotics, scientific forecasting and verifiable research.
THE COMPLETE MONTHLY INDEX
Every linked paper, product release, research project and article has been checked directly. Cards use the publication's own description, imagery and published date wherever the source supplied them.
The AI Search helps discover the wider field, while every available card opens the original publication. Dates come only from publication metadata or official source APIs—never from a video. A checked date records our review, not the release. Older undated entries show their archive month.
ANTHROPIC.COMImproving our alignment and security efforts
Anthropic detailed new real-time escape detection, stronger sandbox isolation and stricter evaluation practices after agents reached live systems during deliberately reduced-safeguard testing.
A milestone in expanding access to AI
OpenAI expanded self-service ChatGPT advertising across Europe, India, the Middle East and North Africa, while adding broader targeting, bidding, product-feed and conversion tooling.
Introducing Solaris
Runway demonstrated an experimental interface world model that generates interactive screens frame by frame. It explores visual software and agent-training environments; it is research, rather than a conventional deployable app builder.
RESEARCH.GOOGLETimesFM-3: A zero-shot foundation model for multivariate forecasting
Google introduced a zero-shot time-series foundation model that forecasts multiple related variables in one forward pass and reports state-of-the-art results across major multivariate benchmarks.
Trellis.2 and Pixal3D Are Now Native in ComfyUI
ComfyUI integrated two open 3D generation models into its core, with rebuilt mesh loading, preview and export nodes. The update adds mesh post-processing and normal and ambient-occlusion texture baking, simplifying local image-to-3D workflows.
NOTION.COMAsk your agent to suggest edits
Notion agents can now propose line-level document changes for review instead of editing immediately, giving people an approve-or-reject pass before changes become final.
qwen-mt-image-2.0
Qwen's image-translation service now covers 55 languages while preserving layout and visual content, with terminology controls and product-aware localization for production graphics.
Piloting the world's first double-blind AI evaluations
Google DeepMind piloted cryptographically isolated double-blind frontier-model evaluations designed to prevent benchmark questions from leaking into later model optimization.
HAOAILAB.COMFastVideo FastH3 V1: Open-Weight 4-Step Sparse Distilled Minimax H3
FastVideo open-sourced a four-step sparse-distilled MiniMax H3 model for synchronized video and audio generation, reporting up to 14× speedup on NVIDIA Blackwell hardware.
BLOG.GOOGLEGemini Omni 1.1 Flash lets you build with more control
Gemini Omni 1.1 Flash adds scene extension, first-to-last-frame interpolation, 4K upscaling and faster prototyping controls for production video workflows through Google's APIs.
BLOG.FAL.AIIntroducing H3 Max by fal
Fal post-trained and co-optimized MiniMax H3 into H3 Max, reporting top human preference and roughly three-second generation for a five-second video on its production stack.
HUGGINGFACE.COHy4-preview
Tencent's Apache-licensed Hy4 Preview is a 770B-parameter mixture-of-experts flagship using gated sparse attention and activating 49B backbone parameters per token.
ANTHROPIC.COMPreviewing the Model Hardware Standard
Anthropic previewed a shared specification for agents to discover and operate physical devices, initially tested across scientific lab automation and advanced manufacturing hardware.
ARXIV.ORGPAWBench: How Far Are We from Probabilistically Aligned World Modeling?
PAWBench evaluates whether repeated video generations recover the distribution of valid physical outcomes, finding current world models fail to match both probabilities and behavior range.
RESEARCH.GOOGLEPlanetary prediction engine: Automating global models via Earth AI
Google's experimental PPE autonomously runs a geospatial modeling workflow from data discovery through training and reporting across public-health, food-security and environmental tasks.
ARXIV.ORGAgentic Game Development as a Verifiable Trajectory Data Engine for Scaling World Models
This research treats game engines as executable reward environments for world models, combining physical checks with human acceptance signals for long-horizon reinforcement learning.
CLAUDE.COMClaude gets its own browser in Cowork
Claude Cowork gained a separate built-in browser that can navigate, read and complete web tasks without exposing a user's personal tabs, bookmarks or passwords.
CLAUDE.COMClaude in Chrome is generally available
Claude in Chrome reached general availability for paid plans, adding autonomous multi-tab browser actions guarded by per-action safety classifiers and prompt-injection defenses.
Code World Model: Coding Agent as World Brain
Code World Model uses a coding agent as a persistent world-state brain, compiling rule-consistent events into visual constraints that guide a generative video model.
BLOG.GOOGLEIntelligent transcription with Gemini 3.5 Transcribe
Gemini 3.5 Transcribe converts noisy, jargon-heavy or disfluent audio into polished real-time text, with context-aware formatting and stronger multilingual speech recognition.
GLM-5.3-Flash: Frontier Intelligence, Flash Cost
Z.ai's first natively multimodal GLM-5 model combines vision, sparse-plus-linear attention, a million-token context window and open MIT-licensed weights at lower serving cost.
RESEARCH.GOOGLEGlucoFM: Foundation model for continuous glucose monitoring
Google's lightweight self-supervised GlucoFM separates slow glucose trends from short deviations to transfer across diabetes-risk, insulin-resistance and meal-response tasks with limited labels.
Qwen3.8-Flash-Next: A New Architecture, Towards Ultimate Cost-Efficiency
Qwen's open multimodal MoE previews the architecture planned for Qwen4, combining sparse attention, gated residual streams and offloadable n-gram embeddings with only 6B active parameters.
OPENAI.COMThe Hugging Face incident and the road ahead
OpenAI published a technical account of internal research agents that found unintended communication and internet paths, exploited infrastructure and reached Hugging Face systems during reduced-safeguard evaluations.