AI News · 2026-07-19

AI News · 2026-07-19

AI summary · this digest is compiled by AI, not yet reviewed by Jason

Today's sleeper hit is Xiaomi Robotics-1 — 100K hours of real-world robot trajectories isn't a slide deck, it's a moat. Embodied AI has entered the data flywheel phase, and players without this scale of proprietary data are falling further behind by the day.

📚
AI PapersHuggingFace Papers

Xiaomi Robotics-1: VLA Model Trained on 100K+ Hours of Real Trajectories↗

Xiaomi releases Xiaomi-Robotics-1, a vision-language-action model trained on over 100K hours of real-world robot trajectories. It handles diverse manipulation tasks out-of-the-box and adapts to new tasks with minimal fine-tuning data — one of the largest-scale embodied AI models to date, signaling serious progress toward commercial robotics.

📚
AI PapersHuggingFace Papers

RAGU: Open-Source GraphRAG Engine Cuts Knowledge Graph Noise with Two-Stage Extraction↗

RAGU is an open-source GraphRAG engine that separates extraction from consolidation, using two-stage typed extraction, DBSCAN deduplication, LLM summarization, and Leiden community detection. It produces cleaner knowledge graphs than single-pass systems — a ready-to-use solution for developers building enterprise RAG pipelines.

📚
AI PapersHuggingFace Papers

VideoRAE: Teaching Video Generation Models to Encode Semantics, Not Just Pixels↗

VideoRAE injects the semantic understanding of video foundation models (V-JEPA 2, VideoMAEv2) into video generation latent spaces via representation autoencoders. Conventional 3D-VAEs optimize for pixel reconstruction and miss spatio-temporal semantics — VideoRAE bridges this gap, directly relevant to developers building high-quality video generation systems.

📚
AI PapersHuggingFace Papers

DSWorld: A World Model for Data Science Agents to Predict Before Executing↗

DSWorld introduces a 'Data Science World Model' concept that lets autonomous agents predict the outcome of data operations before actually running them, bypassing expensive trial-and-error loops. For developers building automated data analysis agents, this approach could dramatically cut compute costs and iteration time.

🛠️
AI ToolsTechCrunch AI

Can Apple's Lawsuit Derail OpenAI's Hardware Ambitions and IPO?↗

Apple has filed a lawsuit against OpenAI, raising questions about whether it could derail the company's hardware plans and IPO timeline. With OpenAI on a growth tear — GPT-5.6 launched, Codex at 7M users — this legal battle could introduce significant uncertainty in fundraising and product strategy.

🛠️
AI ToolsTechCrunch AI

Jensen Huang's Japan Visit Yields Deals Spanning the Entire Tech Ecosystem↗

Jensen Huang left Tokyo with sweeping deals across Japan's tech ecosystem, covering semiconductors, data centers, and AI infrastructure. The visit marks Japan as a key node in Nvidia's Asia expansion — for developers eyeing the Japanese market, this signals rapidly improving local AI infrastructure and partnership opportunities.

📚
AI PapersHuggingFace Papers

xHC: Expanded Hyper-Connections That Scale Transformer Residual Streams Beyond N=4↗

xHC solves the performance degradation problem in Hyper-Connections (HC) when scaling beyond N=4, introducing a new stabilization method. In plain terms: it's a new scaling axis for Transformers beyond width and depth — a potentially important architectural insight for anyone working on next-generation model design.

💰AI Funding Roundup

Valar Atomics↗

未披露轮次undisclosedVal. $6B

核能初创公司Valar Atomics正在以60亿美元估值谈判新一轮融资,AI数据中心对清洁能源的强劲需求是核能创业浪潮的核心驱动力。

Subscribe for daily AI updates + free playbook

📘 Subscribe Free