Top AI News Weekly
Sonnet 5 + Creative Agent OS. Anthropic ships Claude Sonnet 5 (agentic flagship near Opus 4.8 at high effort, lower hallucination/injection, $2/$10 launch pricing). Google launches Flow creative studio + Gemini Omni Flash video and Nano Banana 2 Lite image models. Meituan open-sources LongCat-2.0 (1.6T MoE, ~48B active, 1M context, trained entirely on Chinese silicon — 59.5 SWE-Bench Pro). InternScience Agents-A1 (35B MoE scientific agent). Meta FAIR Brain2Qwerty v2 non-invasive MEG brain-to-text (61% word accuracy). Robotics wave: NVIDIA GEAR ASPIRE (self-improving robot programs, 20%→92%), SimFoundry (video→interactive sim), CHORD (contact-wrench dexterous manipulation), OmniContact humanoid loco-manipulation. Creative/agent tooling: STAGES AI-OS, Comfy MCP, ComfyUI agent workflows, Roomify floor-plan-to-3D, /brag launch-video skill, Goose marketing skills, Ponytail least-code framework. Video/image research: LiveEdit (12.66 FPS streaming edit), ViDiHand 4D hands, LUNA universal human animation, RDM one-step generation, MrFlow training-free speedup, PhysiFormer physics motion. Foundation vision: MuSViT sheet music. Tools: Operately company-OS, Meetily meeting assistant, OctoBot trading, Partnerships OS, TimeSformer explainer.
30 launches and research drops that matter for enterprise AI builders—curated, tagged, and ready for your next roadmap sync.
New drops
30
Unique sources
29
Key themes
Agents · Frontier · Immersive
New reasoning systems, world models, and alignment papers.
Deep-dive on TimeSformer, a convolution-free video classification transformer using divided space-time attention (separate temporal then spatial self-attention) to cut complexity from O(N·F) to O(N+F); TimeSformer-L hits 80.7% top-1 on Kinetics-400 at 0.59 TFLOPs vs 1.97 for 3D CNN baselines.
View release ↗Non-invasive MEG-to-text brain-computer interface that decodes typed sentences via a hierarchical letter/word/sentence architecture; v2 drops keystroke-timing dependence and, trained on ~22,000 sentences/participant, reaches 61% mean and 78% best word accuracy.
View project ↗First foundation vision model for sheet music: a 12-layer ViT (16×16 patches, 768-dim, 2D sinusoidal positional encodings for staff-line/pitch geometry) pretrained with Masked Autoencoders on ~9.7M IMSLP pages; beats general vision encoders under linear probing on transcription, symbol detection and difficulty classification.
View release ↗Open-source 1.6T-parameter MoE agentic coding model (~48B active/token via zero-computation experts) with native 1M-token context and LongCat Sparse Attention; scores 59.5 SWE-bench Pro and 70.8 Terminal-Bench, trained entirely on a 50,000-card domestic Chinese compute cluster (no A100/H100).
View release ↗New agentic flagship that plans and runs autonomously with browsers/terminals, approaching Opus 4.8 at high effort on some tasks while cutting cost; lower hallucination, sycophancy and prompt-injection rates than Sonnet 4.6. Launch pricing $2/$10 per M tokens through Aug 31 2026, then $3/$15.
View release ↗Diffusion transformer for physically realistic 3D object motion that denoises directly on vertex trajectories in world coordinates (not pixels), using factorized time/space/object attention to predict rigid and elastic multi-object dynamics; trained on 100K+ simulated trajectories, ~5x faster than physics simulators.
View project ↗Meta blog on Brain2Qwerty v2, a non-invasive MEG brain-to-text system using end-to-end deep learning plus LLMs fine-tuned on neural data; trained on ~22,000 sentences from 9 participants, it hits 61% word accuracy (78% best) vs ~8% for prior non-invasive methods, with accuracy scaling log-linearly with data. Training code released.
View release ↗Trains one-step image generators with no teacher, adversary or trajectory by matching real vs generated feature distributions across 14 frozen encoders (balanced via a Lagrangian controller); surpasses the 4-step FLUX.2 teacher on GenEval (0.826 vs 0.794), with 63.6% of samples preferred over real photos by PickScore.
View project ↗Video, audio, and physics-native generation techniques.
Real-time streaming video-editing framework using a three-stage distillation pipeline (bidirectional foundation tuning → chunk-wise causal adaptation → DMD 4-step distillation with an AR mask cache) to reach 12.66 FPS causal, instruction-driven editing with strong content preservation.
View project ↗4D hand-motion reconstruction from (esp. egocentric) video by fine-tuning a video diffusion model's VACE branch with hand-overlay rendering (base DiT frozen), then decoding pose/2D-joints via a lightweight dual-branch decoder; SOTA on ARCTIC with strong temporal smoothness under occlusion.
View project ↗Humanoid loco-manipulation framework chaining meta-skills via 'Contact Flow': CF-Gen synthesizes contact-flow segments and CF-Track executes them under 50Hz closed-loop control; trained on 22.3h/1,274 sequences, hitting 98.7% carry-box and 82.5% push-suitcase success with VLM-driven language commands.
View project ↗Two developer models: Nano Banana 2 Lite generates text-to-image in ~4s at $0.034/1K images with strong character consistency and in-image text; Gemini Omni Flash generates/edits up to 10s video via conversational editing at $0.10/sec. Both carry SynthID watermarks and chain in AI Studio/Gemini API.
View release ↗AI creative studio for generating and iteratively editing video/images via natural language, powered by Gemini Omni (video), Nano Banana (image) and Veo 3.1 (native-audio video); supports custom shareable workflow tools. Free tier (50 daily credits) up to AI Ultra at $199.99/mo.
View release ↗Automated real2sim system that turns a single video of a real scene into an interactive physics simulation via object segmentation/depth extraction, 2D-to-3D mesh generation and 3D Gaussian Splatting backgrounds; achieves 0.911 mean real-to-sim Pearson correlation and generates 'digital cousins' for policy generalization.
View release ↗Framework that learns dexterous manipulation from human demos via an object-centric contact-wrench representation (6D force+torque), converting demonstrated contacts into RL rewards; on 1,831 eval tasks it averages 82.12% success (90.77% whole-body from hand-only demos) and ships a 4,739-task benchmark.
View project ↗Universal 3D human animation model that skips Linear Blend Skinning and parametric bodies: an identity encoder builds canonical 3D Gaussians and a transformer neural animator disentangles global rigid motion from local dynamics, enabling zero-shot cross-identity animation driven by RGB, keypoints or sketches.
View project ↗Embodied agents learning to act in complex worlds.
Claude Code skill that turns any software project into a shareable launch video with one command, delegating rendering to the Hyperframes API; outputs an MP4 plus planning docs, music and promo copy. Works across Claude Code, opencode and Codex CLI (Node 22+, FFmpeg).
View repo ↗Open-source Python crypto trading bot (6.2k stars) supporting grid, DCA and basket strategies across 20+ exchanges via CCXT, with backtesting, paper trading, TradingView/Pine Script alerts and AI integration via OpenAI or local Ollama LLMs.
View repo ↗MIT-licensed npm package delivering 200+ marketing/growth skills to coding agents (Claude Code, Cursor, Codex) covering paid ads, SEO, lead gen, content and competitive intel; skill scripts run locally to /tmp, never in the project dir.
View repo ↗Creative 'AI-OS' unifying prompts, assets, agent orchestration, review and delivery into one production command center for professional generative artists and filmmakers, positioned to replace six disconnected tools with one workflow.
View release ↗Second-brain framework for partnership teams built on Dust Pods: personal pods capture individual work while a shared team pod maintains attributed partner profiles, using a markdown-first, single-writer, signal-only rollup design.
View repo ↗35B Mixture-of-Experts agentic model with 256K context, trained via multi-teacher distillation across scientific/tool-use domains; posts SOTA 56.36 on Seal-0 (long-horizon search), 40.0 on FrontierScience and 80.61 on IFBench, served via SGLang/vLLM with OpenAI-compatible endpoints.
View project ↗Self-improving continual-learning system where robots autonomously write, debug and refine control programs via evolutionary program search, building a reusable skill library from multimodal execution traces; improves two-arm handover success from 20% to 92% and shows zero-shot skill transfer across embodiments.
View release ↗Model Context Protocol server connecting agents (Claude, Codex, Cursor) to the ComfyUI ecosystem, letting them build and run image/video/3D/audio workflows in natural language, search templates and nodes, and share reproducible pipelines via URL.
View release ↗Frameworks, playbooks, and OSS repos.
Open-source SaaS tutorial app that converts 2D architectural floor plans into photorealistic 3D renderings using Claude and Gemini, built on React 19 + TypeScript + Vite with Puter for serverless workers, file storage and KV persistence.
View repo ↗Shared AI product/demo post — see the linked X post for the prompt and full details.
View release ↗AI-agent optimization framework that enforces a 'least-code' decision ladder (YAGNI → reuse → stdlib → one-liner) before an agent writes anything; benchmarked at ~54% less code, ~20% cheaper, ~27% faster on a FastAPI+React repo. Ships as a plugin for 16+ platforms (Claude Code, Codex, Cursor, Copilot CLI).
View repo ↗Self-hosted, privacy-first AI meeting assistant (17.4k stars) built with Rust + Next.js + Tauri; real-time local transcription via Whisper or Parakeet with Metal/CUDA/Vulkan acceleration and local summarization via Ollama (or Claude/Groq/OpenAI endpoints).
View repo ↗Apache-2.0 open-source 'company operating system' with built-in OKRs, projects, team spaces, docs and automated check-in cadences; built on Elixir + React/TypeScript, self-hostable via Docker with CLI, API and AI-agent integration.
View repo ↗Training-free acceleration for flow-matching text-to-image models: generate low-res, upsample in pixel space with Real-ESRGAN x2, re-encode with scheduler-consistent noise injection, then 1-step high-res refine; 10x+ speedup on Qwen-Image, works on FLUX.1-dev, Z-Image and distilled models, with ComfyUI support.
View repo ↗