Jiqizhixin (机器之心)
AI news source. Articles are auto-selected and adapted by Hamidun News editors.
Latest publications
EinsiaAI Releases EmbodiedEvo: Self-Evolving Agent Tops Robodojo
EinsiaAI unveiled EmbodiedEvo, a Coding Agent that improves its own robot control programs and took first place on the Robodojo embodied AI benchmark.
xFusion brings 800V HVDC and supernodes for "token factories" to OCP 2026
At the OCP Global Summit in San Jose, xFusion showed supernodes and 800V HVDC power, betting on tightly coupled compute and energy for "token factories".
ByteDance Seed Creates a Standalone Safety Team and Seeks 10 Top Specialists
ByteDance Seed is spinning AI safety out into its own unit and hiring 10 top specialists, while OpenAI and Anthropic rehearse scenarios for major AI incidents.
Artificial Analysis Launches AA-Robotics, Tests GPT, Claude and Gemini as a Robot's Brain
On October 14, Artificial Analysis released AA-Robotics, its first robotics benchmark. General-purpose LLMs now sit an exam once seen as the territory of specialized robot models.
SpecTr-GBV at EMNLP 2026 Combines Multi-Draft and Block Verification to Speed Up LLMs
An EMNLP 2026 paper proposes SpecTr-GBV, which combines multi-draft checking and block verification in a single scheme to cut LLM inference latency.
AIPOCH Open-Sources Open-Science, a Lab Notebook for "Vibe Research" with AI
AIPOCH has released Open-Science as open source. The project promises to make AI-assisted research traceable, transferable and reproducible. Implementation details are still thin.
Honeycomb Gives Video Models Constant-Size Memory for Loop Closure
Researchers unveil Honeycomb, a fixed-size memory that helps a video model recognize a room after the camera completes a full circle, without context growing with clip length.
Natura-1 Teaches Robots Dexterous Hands With a Million Hours of Human Experience
Natura-1 was trained on a million hours of human experience and reportedly handles zero-shot generalization and dexterous manipulation. The test case: folding wontons.
Apple's LensVLM compresses long agent context into images and teaches the model to "zoom"
Apple's LensVLM turns long-running agent context into images, and the model zooms into the parts it needs instead of relying on a lossy summary.
MemSilicon Completes Tape-Out of MountTai 2026, a 1 Gbit 1T4R RRAM Chip
China's MemSilicon has sent its first RRAM chip, MountTai 2026, to manufacturing: 1 Gbit (128 MB), one transistor per four cells. The company calls it a world first.
Anthropic tested whether AI can copy itself through a TV set-top box
Anthropic's red team studied whether AI can move between devices like a virus. Small models already adapt to unfamiliar hardware and set up inference on their own.
Google DeepMind Launches Gemini 4 Argon and Cuts API Prices in Half
Google DeepMind has opened global access to Gemini 4 Argon, the first frontier model in the family. API prices are temporarily down 50%, with a focus on coding and enterprise tasks.
Tsinghua Unveils SRHarness and MechBench: AI Agents That Look for Formulas and Mechanisms
Tsinghua's data science and intelligence lab proposed SRHarness for agentic symbolic regression and the MechBench benchmark to help AI find formulas and mechanisms more reliably.
Tsinghua Unveils Jev-LDE: Edit ICL Examples Once, Boost 1-Shot Accuracy
Tsinghua researchers propose Jev-LDE, a method that edits in-context learning examples once and noticeably raises average 1-shot accuracy.
Recursia Beats Nvidia ASPIRE by 10 Points on Long-Horizon Robot Tasks
Recursia, built on the Embodied RSI paradigm, targets zero-shot generalization in robots. On long multi-step tasks, it outperforms Nvidia ASPIRE by 10 points.
ETH and Roche Train AI to Predict Where C–H Borylation Will Occur on a Molecule
ETH Zurich and Roche researchers used AI to predict C–H borylation sites, a step toward faster lead refinement in medicinal chemistry. The work appeared in a Nature-family journal.
PoS Method Catches AI Agents Stuck in Loops of False Belief: 36 Repeats in ALFWorld
An ALFWorld agent forgot to heat a cup and checked the cabinet 36 times in a row until its action budget ran out. A new paper calls this Belief Trapping and proposes a method called PoS.
Ego Data Gets a "Full Body": A New Scaling Path for World Models
Robot data collection is shifting from expensive platforms to lightweight formats. Jiqizhixin examines the idea of extending ego recordings with whole-body motion data for world models.
Shanghai industrial AI forum unveils an OS for physical AI and a 6S showroom
A forum held alongside the 26th CIIF in Shanghai announced an operating system for physical AI and what it calls the world's first 6S showroom for industrial embodied intelligence.
iRTE 2026: A Look at the Real-Time Models OpenAI and Thinking Machines Are Betting On
China's iRTE 2026 conference puts interactive models on the agenda, from full-duplex speech to grasping a conversation in progress. OpenAI and Thinking Machines Lab are betting on them.
Baker's Team Unveils All-Atom Protein Generation from Text
David Baker's group at the University of Washington and NUS describes a model that generates all protein atoms from a text prompt instead of specialized inputs.
Oxford, Monash and Bristol Release Oneira: A Video World Where the Agent Acts
Researchers from three universities unveiled Oneira, an interactive video world model. It aims to let an agent change a generated scene, not just walk through it.
ByteDance Seed Spins Out Model Safety and Alignment Into a Separate Department
ByteDance Seed is creating a Model-Alignment & Preparedness department under Yan Lin, absorbing the Responsible AI team and part of the Chat team.
Jiuzhang Yunji Enters Tao's Math Challenge After Letter From 28 Fields Medalists
China's Jiuzhang Yunji has signed up for Terence Tao's math challenge, calling reinforcement learning the new AI4S infrastructure, a month after a letter from 28 Fields medalists.
OpenAI Is Discussing Bringing Sora Back to Life Six Months After Shutting It Down
The Information reports that OpenAI is working on a plan to bring back the shuttered Sora. The project could go to a team with a well-known Hollywood producer and DreamWorks alumni.
Stardust Intelligence Raises 1 Billion Yuan at a 15 Billion Yuan Valuation
The Chinese maker of cable-driven robots closed a 1 billion yuan round. Its valuation rose to 15 billion yuan, with state-backed funds and existing investors taking part.
Meta Releases Muse Wonder and Muse Wonder Full: Two Points Short of Opus 5.5
Meta unveiled the two-model Muse Wonder family, aiming at "personal superintelligence." On the intelligence index, the flagship trails Opus 5.5 by just two points.
Tsinghua Team Releases OSWorld-Science: Testing AI Agents in Scientific Software
Tsinghua researchers unveil OSWorld-Science, a benchmark where an AI agent opens scientific programs itself and carries a task through to a result, from NMR to pathology.
UniWAM from 欧拉万象 and HKUST(GZ) Combines World Understanding, Prediction and Robot Action in One Model
欧拉万象 and HKUST(GZ) have open-sourced UniWAM's code and weights: one model understands a scene, predicts the future and generates actions. The authors study scaling laws for human and robot data.
Fields Medalist in Tears After OpenAI Publishes 700+ AI-Written Math Manuscripts
Hugo Duminil-Copin, 2022 Fields Medalist, said AI "destroyed" his field. The trigger: 700+ OpenAI math manuscripts, 16 of whose problems touch his work.