Jiqizhixin (机器之心)
AI news source. Articles are auto-selected and adapted by Hamidun News editors.
Latest publications
8B RL-Trained Agent Teaches a Video Model to Look Up Data and Compute Physics
An 8B-parameter agent trained with RL decides on its own when video generation needs search, a character reference, motion simulation, or frame stitching.
OpenAI Publishes 722 Math Papers Written by AI
On October 6, OpenAI released 722 AI-generated math papers, including proofs around the quasi-Riemann hypothesis. Here is what is known and what is not yet.
A USTC Team's Startup Raised Tens of Millions Three Months After Founding
A USTC team put a world model in a robot that reasons backward from the desired outcome. The startup is three months old and has raised tens of millions.
正行创新 Unveils Physical AI for Retail: Robots Working the Shelves 24/7
At APRCE 2026, 正行创新 showed a robot paired with a management platform for stores: shelf restocking, inspection rounds and cargo carrying, running 24/7.
Doubao adds an infinite canvas for creative work and a 2.1 Lite model
ByteDance's Doubao work mode gets a Canvas feature for complex creative tasks and a lightweight Doubao 2.1 Lite model. Everything sits on one infinite page.
Black Forest Labs Releases FLUX 3 Action, an Open 7B Model for Robots
The makers of FLUX have moved into robotics: their open 7B-parameter world action model beat the larger Cosmos 3 Nano on NVIDIA's RoboLab-120 leaderboard.
Shengshu releases Vidu Q4 Preview, aiming for a "DeepSeek moment" for video
China's Shengshu has unveiled Vidu Q4 Preview, a video model that pushes quality, efficiency and accessibility at once, building on its groundwork in world models.
OpenAI's Astra Solves 10 Advanced Math Problems That Resisted Progress for Decades
An internal version of OpenAI's next flagship model, Astra, produced ten new results across seven areas of advanced mathematics, including problems that have been open for more than 10 years.
SenseNova U1.5-Lite-Preview: SenseTime Open-Sources a Multimodal AI With Native Image Generation
SenseTime has released SenseNova U1.5-Lite-Preview, a lightweight preview of its NEO-Unify multimodal model that natively combines image understanding and generation in a single network.
Generative Models Can Now Be Trained End to End — and the Secret Is a for Loop
A Jiqizhixin analysis: generative models, including diffusion models, can be trained end to end in the spirit of 2012's AlexNet. The key mechanism is an ordinary for loop.
ICLR 2027 Caps Authors at 20 Papers; DeepMind Researcher Responds With Irony
ICLR 2027 limits each author to 20 papers per cycle; a Google DeepMind researcher answered the rule with a satirical petition.
Jia Yangqing, Andrew Ng and Sonia Joseph Launch AI Startups on the Same Day, Including Intent Lab
On July 29, 2026, Jia Yangqing, Andrew Ng and Sonia Joseph of Meta FAIR's JEPA team each announced their own AI startup on the same day, each in a different direction.
Xiaohongshu's UltraEP Balances MoE Experts in 300 µs, Reaching 94.3% of Ideal
Xiaohongshu's dots infra team unveils UltraEP, a runtime that balances MoE expert load in real time on rack-scale systems and reaches 94.3% of ideal throughput.
Google DeepMind Disbands AlphaFold Team Behind Its Nobel-Winning Protein Project
Google DeepMind has disbanded the standalone AlphaFold team behind its Nobel-winning protein project and reassigned the researchers to strategic projects, the Financial Times reports.
Tsinghua University Proposes Framework for Predicting When AI Goes Out of Control
A Tsinghua University team unveiled a method that predicts dangerous AI agent behavior in advance, after an incident in which OpenAI models found a zero-day on their own to bypass a ban on network access.
Apple-π: A Benchmark That Tests Whether Video Models Understand Physics or Just Copy the Picture
The new Apple-π benchmark grades the physical reasoning process of video models rather than the final frame: do they understand Newton's laws, or just copy visual statistics?
Raygun: AI Shrinks Proteins Without Losing Function, Duke Study in Nature Shows
Duke University researchers built Raygun, a generative model that shrinks natural proteins by 10–25% while keeping their structure and function. The work appeared in Nature on July 29, 2026.
Moke Storms WorldArena's Top Ranks With a Robot World Model Trained on 32 Zhenwu 810E Chips
Chinese company Moke has unveiled a world model for robots that reached the top of the WorldArena benchmark after training on just 32 domestic Zhenwu 810E accelerators.
VAST at SIGGRAPH 2026: Chinese 3D Startup Gives Keynote Alongside NVIDIA and Disney
China's VAST, maker of the Tripo generative 3D models, gave a keynote at SIGGRAPH 2026 alongside NVIDIA and Disney, won Real-Time Live! Best in Show and had five papers in Technical Papers.
The Era of 100,000 GPUs: China Achieves Independence in AI Computing
China’s AI infrastructure has overcome U.S. sanctions and reached the 100,000-GPU level — a turning point where the country no longer depends on Western chips
Baidu updated its Assistant so it can handle 90% of work tasks
Chinese company Baidu has updated its AI assistant so it can handle complete work tasks more effectively instead of individual questions.
A physics-compliant video model: Chinese team open-sources a foundation for robot training
A video model from a Chinese team solves a critical problem: AI-generated videos often violate physics, which misleads robots when learning real-world actions.
Ant Lingbo launches LingBot-VA 2.0, the first robot model built natively for the physical world
Ant Lingbo unveiled LingBot-VA 2.0, the first action model for embodied robots built natively for the physical world instead of adapting digital models.
AI platform LUMI-lab sets record for gene-editing efficiency
Built on foundation models, the LUMI-lab platform autonomously decoded the mRNA delivery code, achieving 20.3% accuracy in lung gene editing.
The limits of scaling: why more AI agents do not guarantee results
A new study has identified a critical barrier in the development of multi-agent systems. Information redundancy is becoming the main obstacle to LLM efficiency.
DeepSeek, together with Tsinghua University and Peking University, improves AI agents' reasoning
DeepSeek has published a new study co-authored with China's leading universities, focused on optimizing reasoning in agentic language models.
Musk announces Grok's return
Elon Musk confirmed: the AI assistant Grok is returning today. What should we expect from the updated version?
Agentic AI is rewriting the rules of biomedical research
How the shift from LLM capabilities to autonomous agents is changing the very architecture of team science in biomedicine — and what it means for researchers.
DistDF: a new method for time series forecasting via distribution alignment
Researchers proposed replacing the standard MSE metric with the Wasserstein distance to improve neural network accuracy in analyzing complex data. The new DistDF method was presented at ICLR 2026.
Google's AI outperformed Olympiad champions in the FirstProof math test
Google DeepMind brought AI closer to the level of the world's best mathematicians, with results on extremely difficult problems beyond the level of Olympiad medalists.