Jiqizhixin (机器之心)
AI news source. Articles are auto-selected and adapted by Hamidun News editors.
Latest publications
RoboFoundry Lifts GPT-5.5 on EmbodiedBench From 56.9% to 72.7% Without Fine-Tuning
RoboFoundry evolves the system around a frozen GPT-5.5 and raises its average success rate on EmbodiedBench from 56.9% to 72.7%. The model's weights did not change.
PhAI Labs Launches RSI-Anything: A World Model Drives AI Self-Evolution Across 5 Tasks
PhAI Labs proposes RSI-Anything: a world model predicts and selects edits while the executor stays the same. The framework was tested on five task types, from bugs to immunology.
OpenClaw v2026.10.1 Lands 2425 PRs From 402 Contributors, and Users Are Tired
Open-source agent OpenClaw shipped v2026.10.1: 402 contributors, 2425 pull requests and 3055 commits, with updates to Skills and MCP Apps. Some users want the pace to slow down.
AI Analyzes 70 Million Cells to Identify Regulators of Autophagy and STING
Genetic screening meets microscopy: a neural network reads the morphology of 70 million cells to find genes that affect autophagy and the STING signaling pathway.
A Survey of 200 Papers Systematizes Affordance Learning in Embodied AI
Jiqizhixin reviewed a survey of nearly 200 studies on affordances: how a robot can tell what it can touch, where, and how, rather than just recognizing an object.
Google lets candidates use Gemini in engineering interviews
Google has launched a pilot interview format that allows Gemini: instead of purely solving algorithms, it tests how candidates read someone else's code and find bugs in it.
Karpathy Estimates Only About 5,000 People Have Access to the Most Advanced AI
Andrej Karpathy believes only about 5000 people can actually work with the closed systems inside frontier labs, where clusters of thousands of agents run for weeks.
Tsinghua Team Pushes Generation Down to a Single Step, From Images to Streaming Video
Tsinghua researchers are developing "sample distribution training" to enable one- or few-step generation, from still images to real-time video.
模思智能 Raises Over 1 Billion Yuan in Pre-A for a Real-Time Omni-Modal Model
Shanghai startup 模思智能 closed a Pre-A round of over 1 billion yuan, about $140 million. The money will fund a foundation model for real-time omni-modal interaction.
New Benchmark Tests Whether AI-Generated Images Obey the Laws of Physics
An image can score high on aesthetics and prompt adherence while showing distorted hands and objects passing through each other. A new test measures "world consistency."
UNITAS: A 3D World Model for Robots Claims the Lead on Several Benchmarks
A Chinese team unveiled UNITAS, which it calls the first 3D world model built for robots. It claims the lead on several benchmarks, but the numbers have not been disclosed.
HIL-UMI: Dong Hao's Team Finds a Robot's Weak Spots Without the Robot
Dong Hao's team proposes HIL-UMI: find a robot policy's weak spots from human demonstrations with a handheld gripper, without running the manipulator at every fine-tuning step.
OpenRSI Index Hits 133,000 Views: Measuring AI Against Scientists
OpenRSI Index tries to measure how far AI is from the level of frontier researchers. Its launch post on X drew more than 133,000 views and made it into Today's News.
ChatGPT Co-Creator Liam Fedus: After OpenAI's 722 Manuscripts, AI Should Move Into the Lab
OpenAI posted 722 math manuscripts across 17 fields, some with Lean proofs. Liam Fedus argues that after mathematics, AI should move into the laboratory.
LIBERO-MAX Runs Robots Through 224,000 Simulations With a Moving Object
The LIBERO-MAX benchmark tests how a robot reacts when an object is moved mid-grasp: 224,000 simulations instead of the usual static scenes.
Terence Tao Gives Caltech Lecture "Math 2.0" on Machine-Assisted Proofs
Terence Tao has published his Caltech lecture "Math 2.0." Over a year ago it was conceived as "Machine Assisted Proof"; it then slipped from May to October.
China's CXMT Plans to Show 4F² DDR5 RDIMM by Year-End
CXMT president Cao Kanyu announced server DDR5 RDIMM modules with a 4F² cell and hybrid wafer bonding. The cell area is theoretically one-third smaller than 6F².
OpenAI Says It Has an 'Intern Researcher' and Targets an Automated Scientist by 2028
OpenAI says it has reached the "automated AI intern" stage and promises a full AI researcher by March 2028. Jiqizhixin links the path to RSI to agent memory.
Claude Science Fills In the Missing Third of the Ultraviolet Sky Map
The first complete all-sky map in ultraviolet is here. Claude Science filled in the third of the celestial sphere that observational data did not cover.
AgentForesight: A 7B Auditor That Catches Multi-Agent Errors While the Task Is Still Running
Researchers from Rutgers, UT Austin and Purdue built a 7B-parameter model that monitors a multi-agent system and flags the critical error before the task fails.
Karpathy Releases autoresearch: AI Agents Start Doing AI Research
In March 2026, Andrej Karpathy open-sourced autoresearch on GitHub. Jiqizhixin sees it as a symptom of a year in which AI increasingly takes part in building AI.
ROMA Teaches Robots to Touch, Shake and Squeeze Objects to Understand Them
Researchers unveiled ROMA: a robot that recognizes when it lacks information about an object and fills the gap with its hands, much as a person weighs and shakes a box.
Six Months In, Li Hongyang's Team Shows a Robot That Fetches Its Own Step Stool
Six months after launch, Li Hongyang's startup has shown its first real-world results in 全身智能 (WBI): the robot couldn't reach a toy, so it brought a stand and climbed onto it.
Robot-Staffed Shop in Hefei Takes Around 100 Orders a Day
A shop staffed only by robots opened in a Hefei art district over the National Day holidays. It handled about 100 orders a day, a figure that offers a real-world check on the technology.
NVIDIA Physis-Lang Describes Physics in Words and Takes First Place on Physics-IQ Verified
NVIDIA's Physis-Lang uses text descriptions of physical processes, such as melting butter, to help a video model reproduce physics more accurately. NVIDIA says it ranks first on Physics-IQ Verified.
Westlake Robotics Unveils WR1: A Humanoid That Folds Clothes on Its Own With a Dexterous Hand
Westlake Robotics has unveiled WR1, a "general-purpose brain" for humanoids. The robot moves between household scenes without stopping and, the company claims, is the first to fold clothes by itself with a dexterous hand
OpenAI Fires Three Safety Researchers, Who Respond With an Open Letter
OpenAI says it fired three employees for sharing confidential data with an outside AI safety evaluator. They responded with a joint open letter.
Aether AI shows a manipulator with a causal core in a seven-minute video full of interference
Aether AI released a seven-minute video of a manipulator working with a coffee machine and a microwave while people interfere at every step. The company is betting on a causal core over correlations.
Protein Structure Model Fine-Tuned on Pharma Companies' Private Data Without Transferring It
A protein structure AI model was improved with pharma data that never leaves company perimeters. The algorithm goes to the data, not the data to the algorithm.
OpenAI releases GPT-6.1 Sol Ultrafast on day 4 of its 28-day marathon
A day after opening GPT-6 to all users, OpenAI shipped a fast GPT-6.1 Sol Ultrafast and updated its model control system. This is day 4 of 28.