Jiqizhixin (机器之心)
AI news source. Articles are auto-selected and adapted by Hamidun News editors.
Latest publications
LightSprint Raises $6.1M for a Multi-Agent Product-to-Engineering Loop
Startup LightSprint has raised $6.1M to build a system of multiple AI agents that aims to connect product teams and engineering without long approval cycles.
Matrix Game 4.0 Hits 60 FPS on a Single Gaming GPU and Beats Nvidia's Rival Model
China's interactive world model Matrix Game 4.0 generates a playable image in real time at 60 FPS on a single consumer GPU.
CodeActionBench: GPT-6 Astra in Codex CLI Hits 73.3% Success on Two-Handed Tasks
The new CodeActionBench tests whether a general-purpose agent that looks at an image and writes code can control a robot. The leader so far posts a 73.3% attempt success rate.
Stanford and Together AI Teach LLMs to Write Text Starting From an Outline: Generation Up to 31x Faster
Stanford and Together AI researchers brought VAR's "next-scale prediction" to text: the model writes an outline first, then fills in details. The claimed speedup is up to 31x.
DeepSeek credits users with tokens as compensation after multi-hour outage
On the morning of October 15, DeepSeek's web chat and API, including V4 Pro and V4.1 Flash, were down for several hours. The company made a rare move for the market: it issued tokens.
Meta to Use Muse Spark 1.3 in Medical Research, Promises a 10x Speedup
Meta is applying Muse Spark 1.3 to observational studies on electronic health records and claims a tenfold efficiency gain. The Muse Wonder model moves into medicine.
Apple and CMU Speed Up AI Agents 10x by Replacing Reasoning with Selection
Apple and Carnegie Mellon researchers propose dropping free-form "thinking" from ReAct agents and replacing it with a choice among options. The claimed speedup is up to 10x.
KTransformers Moves Part of the MoE Computation to the CPU and Eliminates Weight Transfers During Fine-Tuning
The KTransformers team proposes running MoE experts directly on the CPU during fine-tuning, aiming to tune open-source giants like DeepSeek-V4 and Kimi-K3 without a GPU cluster.
Kling 4.0 Launches With Native 30-Second Video Generation and Combat Scenes
Kling 4.0 launched on the evening of October 15. The model generates 30 seconds of video in a single pass and takes on battle scenes with cavalry, sieges and crowds.
WorldOracle Tests World Models: The Robot Missed, but the Model Showed Success
A team including Princeton researchers introduced WorldOracle, a test of whether world models predict a robot's real outcome or just paint the familiar success.
HKU's EgoWild2Dex Trains a Robot Hand on 538.9 Hours of "Messy" Video and One Hour on Hardware
The University of Hong Kong trained robotic manipulators to unpack parcels, use a glue gun and make cold drinks from 538.9 hours of noisy first-person footage.
Manifold AI Tops RoboDojo, Beating GPT-6-Astra and GR00T-N1.7
China's Manifold AI unveiled R²SI and took first place in the RoboDojo Agent Track, ahead of GPT-6-Astra and NVIDIA GR00T-N1.7, three months after releasing its world action model.
Meta's Muse Agent Hits 5M US Downloads in 22 Days, Faster Than ChatGPT
Sensor Tower estimates Meta's personal AI agent Muse reached 5M US downloads 22 days after its September 8 launch, faster than historical figures for ChatGPT, Grok and Claude.
VAST Open-Sources a World Model: Anime Version of Flash Step Runs in Real Time
China's VAST has open-sourced a world model, showcased in a viral anime version of "Flash Step". Alongside it comes the story of a Little Prince game built in 16 days.
Yuandong Tian's Recursive_SI Releases Flint-40B-A2B: Model Built in Under 4 Months
Yuandong Tian's startup unveiled its first model, Flint-40B-A2B, built for local agentic coding. The Eureka Machine system built it from scratch in under four months.
AI Chatted With Scammers 2.5 Million Times and Pulled Out 250,000 Leads
A Chinese project turned scam calls into a data source: an AI held 2.5 million conversations and collected 250,000 pieces of intelligence on fraud schemes.
Tsinghua's LoopWAM Tops RoboChallenge Table30 and Goes Open Source
LoopWAM, a world action model from Tsinghua, 原力灵机 and 极寰时代, took first place on the RoboChallenge Table30 benchmark. The code, paper and project page are already public.
Ant Shuke Releases LingDT-Decision-1.0: 4B and 27B Models for Fast Decisions
Ant Shuke unveiled LingDT-Decision-1.0, a decision model for structured tasks in 4B and 27B versions. The company claims higher accuracy than Microsoft's solutions and cheaper tokens.
EinsiaAI Launches Quant Arena: Half of 9 LLMs Lose Money on China's A-Share Market
Nine models from six vendors wrote and refined trading strategies for the Chinese market. On held-out data, only GPT and Opus 5.5 stayed profitable; the best result was about 10% annualized.
ReGDiff Teaches Diffusion Models to Design Metamaterials Without Copying or Producing Defects
ReGDiff first organizes the latent space, then pushes the model away only from its nearest training samples during sampling. The resulting structures are novel and do not fall apart.
Volcano Engine Opens Seedance 2.5 4K to 20 Studios for Film Projects
Volcano Engine has launched an enterprise pilot of Seedance 2.5 4K. The first access batch went to 20 studios and directors, with big-screen cinema as the stated goal.
OpenAI: GPT‑5.6 Sol model finds critical flaw in PHP, which runs 70% of websites
OpenAI says GPT‑5.6 Sol found a critical vulnerability in PHP. For two months, the company coordinated the fix with maintainers, researchers and security vendors.
Wujie Dongli Open-Sources World Model MWA-Phoenix, Claims Top Spot on Six Benchmarks
China's Wujie Dongli has open-sourced MWA-Phoenix, a latent world model. The company claims first place on six benchmarks, from simulation to real robots.
DeepSeek open-sources its DeepInfer inference engine, and analysts calculate an 82% margin
DeepSeek has released its internal DeepInfer framework and the deepseek-recipe component under Apache 2.0. Third-party estimates put the gross margin of its AI service at 82%.
RVCBench Tests 18 Voice Cloning Models on Noise, Compression, and Anti-Cloning Protection
UBC researchers present RVCBench at NeurIPS 2026: 18 voice cloning models tested on noisy recordings, phone channels, and "anti-clone" protection.
OpenAI's Interpretability Founder Says 90% of Neurons Have a Single Function
OpenAI's Nick Cammarata argues that with agentic tools even a trillion-parameter model is nearly readable, and 90% of neurons serve one function. The essay is already drawing debate.
Kunlun Wanwei launches Matrix-Game 4.0: a video model that plays in 480p at 60 FPS on a single GPU
Kunlun Wanwei has released Matrix-Game 4.0, a game video model that outputs 480p at 60 FPS on a single consumer GPU. A game engine runs the world; a neural network draws the picture.
Claude Traces the Origin of Five NIST Curve Seeds in Half an Hour
Anthropic says an internal Claude model worked out where the seeds of five NIST elliptic curves came from in under 30 minutes. The question had stayed open for more than 20 years.
Anthropic: Claude Designed 318 484 Proteins in 64 Hours, and 17 Passed Lab Validation
Claude generated 318 484 protein designs against a scorpion neurotoxin in 64 hours. In the lab, 17 candidates were confirmed: AI-driven pharma has moved from simulation to the test tube.
CAS Institute of Software and Kuaishou Build First Benchmark for LLM User Simulation
The CAS Institute of Software and Kuaishou introduced a NeurIPS 2026 benchmark testing whether LLMs can convincingly play a real user across the entire behavior chain.