Kaiming He's All-Undergrad Team Achieves Text-to-Image With Only 258M Parameters
量子位 QbitAI39 days agoPaperA new research paper from Kaiming He's lab — notable for having an all-undergraduate team — demonstrates that high-quality text-to-image generation can be achieved with just 258 million parameters. This challenges the prevailing assumption that competitive image synthesis requires multi-billion-parameter models. The work signals a push toward leaner, more accessible generative vision architectures.
World's First Universal Humanoid Robot Cerebellum Achieves Zero-Shot Generalization
量子位 QbitAI39 days agoPaperA team has released what it claims is the world's first universal "cerebellum" for humanoid robots — a general-purpose low-level motion-control module. The system was trained on the largest known human motion dataset, comprising 20,000 hours of recorded actions. The result is a controller capable of zero-shot generalization, meaning it can drive robot motion on new tasks or platforms without task-specific retraining.
AIEC 2026: Taichu Yuanji Shares Domestic AI Compute Practices to Scale Token Services
量子位 QbitAI39 days agoHardwareAt the AIEC 2026 conference, Chinese AI infrastructure firm Taichu Yuanji shared its hands-on practices for leveraging domestic AI computing power. The session focused on translating that compute capacity into reliable, production-ready Token-based LLM services. The talk reflects China's broader push to build self-reliant AI infrastructure independent of Western chip supply chains.
ABot-Earth0.5 Tops Three Hugging Face Paper Leaderboards, Earns Praise from Graphics Expert Chen Baoquan
量子位 QbitAI39 days agoPaperABot-Earth0.5, a newly released AI model or research paper, has reached the top position across three concurrent Hugging Face paper ranking lists. The achievement drew public praise from Chen Baoquan, a respected international authority in computer graphics. The milestone signals growing recognition for the project within both the research and graphics communities.
GLM-5.2 Passes Vibe Check; Z.ai Forecasts Open Fable by December
Latent Space39 days agoBenchmarkZhipu AI's GLM-5.2 has passed broad informal community vibe checks, drawing favorable comparisons to GPT-class models and signaling a meaningful quality leap for open-weights AI. Z.ai, the company behind GLM, is additionally forecasting release of an open frontier-tier model — dubbed Open Fable — by December 2026. Together, these developments suggest open models are genuinely competing at the frontier rather than perpetually trailing closed proprietary systems.
Are You in the Weights? — Tool to Check If Your Content Trained AI Models
Hacker News (AI keywords)39 days agoNew Tool"Are You in the Weights?" is a newly launched service submitted to Hacker News by its creator, letting people investigate whether their content contributed to AI model training. The phrase "in the weights" refers to how training data becomes encoded in a model's parameters. The tool addresses growing public demand for transparency around AI data provenance and consent.
MosaicLeaks: Can Your Research Agent Keep a Secret?
Hugging Face Blog39 days agoBenchmarkServiceNow researchers introduce MosaicLeaks, a benchmark evaluating information-leakage risks in AI-powered research agents. The work asks whether agentic systems—given access to proprietary or sensitive documents—might inadvertently expose confidential content in their outputs. It targets a growing enterprise security concern as agents move from single-turn Q&A into multi-step workflows spanning private knowledge bases.
Build Your Own Vulnerability Harness
Cloudflare Blog39 days agoTutorialCloudflare has published a technical breakdown of an AI-assisted vulnerability discovery pipeline built around multiple processing stages and an automated triage loop. The architecture addresses false positives through adversarial review, where the system challenges its own findings before surfacing them to humans. The post also covers state control strategies and techniques for routing around the context-window limits inherent to large language models.
Stellantis Eyes Role as Robotaxi Platform Supplier for Uber via Wayve Partnership
INSIDE 硬塞 AI40 days agoBusinessUber, Wayve, and Stellantis have signed a memorandum of understanding to jointly deploy Level 4 autonomous robotaxis. The partnership integrates Uber's ride-hailing network, Wayve's AI driving system, and Stellantis vehicles. The deal reflects Uber's broader strategy of building a multi-partner, technology-diversified autonomous vehicle ecosystem.
China's Domestic AI Compute Is Entering the Token Standardization Era
量子位 QbitAI40 days agoBusinessChina's domestic AI computing sector is undergoing a structural shift, adopting Token throughput as its standard measurement and pricing unit — mirroring the model established by global API providers. This standardization signals a maturation of the Chinese AI supply chain, moving away from raw hardware metrics toward output-centric benchmarks. The transition has implications for how domestic chip vendors, cloud platforms, and AI service providers compete and interoperate.
G7 Nations Want US AI but Fear Overnight Access Cutoffs, Digital Sovereignty Anxiety Rises★ 72
INSIDE 硬塞 AI40 days agoRegulationUS restrictions on AI model exports have raised alarms among G7 allies who fear unilateral access cutoffs to critical technology. The issue has escalated to a digital sovereignty concern, with leaders exploring 'trusted partner' mechanisms to ensure stable supply of key AI infrastructure. The debate reflects a deepening tension between dependence on American AI and national security autonomy.
France Advances Europe's AI Future With NVIDIA Technologies
NVIDIA Blog40 days agoBusinessA year after France unveiled its national AI ambitions at NVIDIA GTC Paris during VivaTech, the infrastructure is moving from blueprint to reality. AI factories, national compute capacity, open frontier models, and industrial platforms are coming online. AI agents are now running in production, and French startups are actively deploying applications across the ecosystem.
Hermes Agent Integrates Stripe: AI Agents Can Initiate Payments but Cannot Self-Authorize
INSIDE 硬塞 AI40 days agoNew ToolHermes Agent has integrated with Stripe, enabling autonomous AI agents to participate in end-to-end payment transactions. While this marks a significant step toward fully automated commerce, the system deliberately prevents agents from self-authorizing transactions. The development highlights growing industry pressure to establish authorization, spending-limit, and audit standards for agentic financial workflows.
Midjourney Medical: Scan Your Organs as Easily as Stepping on a Scale
Latent Space40 days agoBusinessMidjourney, widely known as the only bootstrapped frontier AI lab, has announced its second product: Midjourney Medical, a system designed to make organ scanning as accessible as stepping on a household scale. The launch marks a dramatic pivot beyond image generation into consumer health diagnostics. It represents one of the most unexpected product expansions from a major AI company in recent memory.
GLM-5.2 Takes the Top Spot Among Text-Only Open-Weights LLMs★ 72
Simon Willison's Weblog40 days agoReleaseZ.ai has released GLM-5.2, a 753B-parameter MIT-licensed open-weights model with a 1-million-token context window. Independent benchmark site Artificial Analysis ranks it first among open-weights models on their Intelligence Index v4.1, ahead of MiniMax-M3, DeepSeek V4 Pro, and Kimi K2.6. It also places second on Code Arena's WebDev leaderboard behind only Claude Fable 5, despite being text-only, and is available on OpenRouter at $1.40/$4.40 per million input/output tokens.
AI Coding Agents Can Autonomously Direct Robot Training, NVIDIA Demonstrates
Ars Technica AI40 days agoPaperNVIDIA has introduced a self-improvement program for robots that delegates training direction to teams of AI coding agents rather than human engineers. The system enabled robots to learn precise physical tasks, including installing GPUs and cutting zip-ties. The approach signals that agentic AI paradigms developed for software are now being applied to embodied robotics training pipelines.
World Model Startup Odyssey Reaches $1.45B Valuation With Amazon Backing
TechCrunch AI40 days agoBusinessOdyssey, a startup building world models — AI systems that simulate and predict physical environments beyond the scope of text-based LLMs — has closed a funding round valuing the company at $1.45 billion. Amazon is among the prominent backers. The raise positions Odyssey as a leading contender in what many see as the next frontier of AI development.
Adam (YC W25) – Open-Source AI CAD Tool Launches on GitHub
Hacker News (AI keywords)40 days agoNew ToolAdam, a Y Combinator Winter 2025-backed company, has announced its open-source AI CAD tool via a Hacker News Launch post. The project is hosted on GitHub under Adam-CAD/CADAM and targets the computer-aided design space with AI capabilities. As a YC-pedigreed open-source entry, it signals growing momentum toward AI-native design tooling for engineers and hardware builders.
MolmoMotion: Language-Guided 3D Motion Forecasting
Hugging Face Blog40 days agoPaperAllen Institute for AI has released MolmoMotion, a new model that adds language-guided 3D motion forecasting to the open-source Molmo family. By conditioning spatial trajectory predictions on natural language, the system enables more flexible, human-interpretable motion anticipation. The work targets applications in robotics, video understanding, and embodied AI where predicting movement in 3D space is safety-critical or operationally essential.
From Hugging Face Hub to Robot Hardware with Strands Agents and LeRobot
Hugging Face Blog41 days agoTutorialA Hugging Face blog post co-authored with Amazon demonstrates how to take AI models from the Hugging Face Hub all the way to running on physical robots. The integration combines Amazon's open-source Strands Agents agentic framework with Hugging Face's LeRobot robotics library to create an end-to-end pipeline. The result is a practical path for developers to deploy Hub-trained policies and models onto real robot hardware using agent-based orchestration.
High-Res Neural Cellular Automata
Hacker News (AI keywords)41 days agoNew ToolA researcher has shared an open project — cells2pixels — showcasing high-resolution neural cellular automata (NCA), a technique where neural networks encode local update rules that cells apply iteratively to produce emergent, self-organizing images. The work extends prior NCA research by targeting higher output resolutions. Shared as a 'Show HN' post, it invites community feedback on the approach and implementation.
GLM-5.2 Claims Top Open-Weights Spot on Artificial Analysis Intelligence Index
Hacker News (AI keywords)41 days agoBenchmarkGLM-5.2, the latest open-weights model from Zhipu AI, has claimed the top position on the Artificial Analysis Intelligence Index among all openly available models. This marks a notable shift in the open-weights leaderboard, which tracks quality, speed, and price across dozens of frontier and community models. The result signals continued momentum from Chinese AI labs producing competitive open-weights alternatives to proprietary frontier systems.
One Model Controls Hands, Feet, and Torso: Robots Achieve Whole-Body Coordination for Fine Manipulation
量子位 QbitAI41 days agoPaperResearchers have developed a unified model that simultaneously controls a robot's hands, feet, and torso, enabling full-body coordination. This approach allows robots to perform fine, dexterous tasks that previously required fragmented, limb-specific control systems. The advance represents a meaningful step toward humanoid or multi-limbed robots that can handle complex real-world manipulation with integrated motor intelligence.
Six Consecutive Wins! WeRide Breaks China Autonomous Driving Competition Record
量子位 QbitAI41 days agoBenchmarkChinese autonomous driving company WeRide has secured its sixth consecutive championship at a major China autonomous driving competition, extending its own record for the longest winning streak in the event's history. The achievement underscores WeRide's sustained technical edge in China's increasingly competitive intelligent-driving landscape. The milestone comes as domestic AV players intensify efforts ahead of anticipated commercial scaling in the second half of 2026.
A Chinese Domestic Player Finally Cracks the Vicious Cycle Trapping Medical AI
量子位 QbitAI41 days agoBusinessMedical AI in China—and globally—has long been trapped in interlocking chicken-and-egg problems: you need clinical data to build good models, regulatory approval to deploy, and deployment evidence to earn approval. A Chinese domestic player, highlighted by QbitAI, is reported to have navigated this full end-to-end gauntlet. If validated, the breakthrough could offer a replicable commercial template for the broader medical AI sector in China.
Zhipu's Open-Source GLM-5.2 Claims Top AI Coding Rank, Second Only to Fable-5
量子位 QbitAI41 days agoBenchmarkZhipu AI has released GLM-5.2, an open-source large language model that has claimed the top position in AI coding benchmarks among all models except Anthropic's Fable-5. The result marks a significant milestone for the open-source community, showing that the gap between proprietary frontier models and open-source alternatives in code generation continues to shrink. For developers seeking capable, self-hostable coding models, GLM-5.2 now represents the strongest open-source option available.
Top Embodied-AI Brain Startup Raises Hundreds of Millions; 15 VCs Back World-Model Route
量子位 QbitAI41 days agoBusinessA top-tier startup specializing in embodied-AI brain systems has secured another funding round worth hundreds of millions of dollars, drawing fierce competition from 15 venture capital firms. The company pursues a world-model approach—building internal representations of physical environments to enable more generalizable robot reasoning. The deal underscores surging investor conviction in world-model architectures as the dominant path to scalable embodied intelligence.
GLM-5.2: Built for Long-Horizon Tasks
Hugging Face Blog41 days agoReleaseZhipu AI has published GLM-5.2 on Hugging Face, framing the release around strong performance on long-horizon tasks — problems requiring sustained reasoning and planning across many dependent steps. The model continues the GLM lineage, one of China's most prominent open-source large-language-model families. By centering the announcement on long-horizon capability, Zhipu AI signals a strategic shift toward agentic and autonomous AI workflows rather than single-turn benchmark performance.
GLM 5.2 Performance Benchmarks — Artificial Analysis
Hacker News (AI keywords)41 days agoBenchmarkArtificial Analysis, an independent AI evaluation platform, has released benchmark results for Zhipu AI's GLM 5.2 language model. The evaluation covers the standard Artificial Analysis methodology, which typically assesses output quality, inference speed, and price-per-token. GLM 5.2 represents the latest iteration of Zhipu AI's flagship model series, positioning it against leading frontier models on a common scoring framework.
GLM-5.2: World's Top Open Frontend Coding Model + IndexShare Speculative Decoding
Latent Space41 days agoReleaseGLM-5.2 has claimed the leading position worldwide among open models on frontend coding benchmarks, marking a significant milestone for the open-source AI ecosystem. The release is accompanied by IndexShare, a new method targeting speculative decoding to improve inference throughput and reduce serving latency. Together, the two developments advance both capability and deployment efficiency for teams building with open models.