Jiqizhixin (机器之心)
AI news source. Articles are auto-selected and adapted by Hamidun News editors.
Latest publications
Tsinghua Team Releases HandsOn: 66 Tasks to Test Robot Dexterity
Tsinghua researchers built HandsOn, a 66-task benchmark for bimanual manipulation with multi-fingered robot hands, designed around egocentric recordings of human activity.
Google Open-Sources TorchTPU: PyTorch Gets TPU as a Native Device
At PyTorch Conference 2026, Google open-sourced TorchTPU, a stack for running PyTorch on TPUs. In Hugging Face Transformers, one line is enough: device_map="tpu".
Altman Outlines Three Endgames for the AI Race: From a Global Regulator to Paralyzing Rivals' GPUs
At an internal OpenAI meeting, Sam Altman described three ways the AI race could end. The harshest: the leader gains the ability to disable competitors' hardware.
VentureBeat: 79% of Companies Leave GPUs Idle at Least Half the Time
VentureBeat Intelligence's continuous Q3 2026 survey: 79% of companies say their GPUs sit idle at least half the time. The figure did not change across all three months.
Moonshot AI Unveils Kimi K3.1 Pro, the First Pro Model in the Lineup, After the Kibitz Test
China's Moonshot AI has confirmed that Kibitz, the anonymous model from Arena's blind battles, is Kimi K3.1 Pro. It is the first Pro model in the lineup, and it draws SVGs, websites and 3D scenes.
Inception Releases Mercury 3: a Diffusion LLM 8x Faster Than GPT-6 Luna
Diffusion LLM Mercury 3 is over 40 times larger than its predecessor, runs 8 times faster than GPT-6 Luna and costs $0.35 per million output tokens. Benchmarks favor it too.
Alibaba Qwen Names Three Pillars of Its Personal Agent: Model, Memory and Ecosystem
At CCF 2026, Qwen algorithm lead Jiang Guanjun explained how Alibaba is building a personal agent: the model, memory and ecosystem work only in combination.
X Square Robot to Invest 300 Million Yuan in Hefei Headquarters for Its Robots
Chinese embodied AI startup X Square Robot has signed a headquarters project in Hefei's high-tech zone worth about 300 million yuan, a bet on moving from demos to deliveries.
Anonymous Models Kibitz and Kalmia Surface in Code Arena: Kimi K3.1 and Qwen 4 Suspected
Two unknown models have appeared in blind coding battles on Arena.ai. The community links them to upcoming Chinese releases: Kimi K3.1 from Moonshot and Qwen 4 from Alibaba.
Anthropic is testing different Claude limits on identical plans: a gap of up to 20%
SemiAnalysis found that Anthropic is running an A/B test on Claude subscriptions: at the same price, some users get a five-hour quota about 20% smaller than everyone else.
Alibaba Qwen Unveils TRACE: FP4 Rollout Runs Up to 5.4x Faster
Qwen's TRACE method for RL on large models runs weights, activations and the KV-cache in FP4 during Rollout, delivering up to a 5.4x speedup over BF16.
Codex goes beyond code: 5M weekly users in June, 40M by October
Knowledge workers are growing on Codex three times faster than developers. Together with ChatGPT Work, the OpenAI product reached 40 million active users by October.
iFlytek Unveils Spark X2.6, the Xingchen TaaS Platform and the SparkAudio 1.0 Speech Model
At the 1024 festival in Beijing, iFlytek showed Spark X2.6 with a device-plus-cloud architecture, the Xingchen TaaS platform and the SparkAudio 1.0 speech foundation model.
Online Convex Optimization Went from Theory to Adam and LLMs in 20 Years
Jiqizhixin looks back at 20 years of online convex optimization: from Online Newton Step with logarithmic regret to the adaptive optimizers that train neural networks.
Open-Source Vision Backbone Learns to Change Its Own Learning Rule on the Fly: 1000+ GitHub Stars
Developers have open-sourced a vision backbone that changes its own learning rule as it runs. It has passed 1000 stars on GitHub and ranks 2nd on Hugging Face.
Stanford PhD Derives the Transformer From Scratch, Starting With One Memory Problem
A Stanford graduate's blog post derives the Transformer from one question: how to store a lot and retrieve any item. Nine years after the 2017 paper, the post is spreading widely online.
Zhejiang University and Peking University Propose SPADER for Multi-Answer QA
SPADER, accepted to EMNLP 2026 Main, is an RL method with step-wise GRPO and a diversity reward that trains models to find all answers, not just the most common ones.
Xiaohongshu and Fudan University Build DR-Rubric for RL on Open-Ended Tasks
Xiaohongshu and Fudan University present DR-Rubric at NeurIPS 2026, a way to build evaluation criteria for RL on tasks that have no reference answer.
OpenAI exposes three Aurora models in its API: red, green and blue alpha
Three unfamiliar identifiers appeared in OpenAI's API model list on October 18: aurora-red-alpha, aurora-green-alpha and aurora-blue-alpha. There is no official information yet.
DynStream from Beijing builds a dynamic 3D scene from video without camera poses
Peking University, VGI Labs and the University of California present DynStream: feed-forward Gaussian reconstruction of moving scenes from streaming video, with no camera poses.
PhAI Labs: Claude Opus 5 Ranks Last in Scientific Hypothesis Test With 2.3%
PhAI Labs checked hypotheses from five model groups against later publications. Claude Opus 5 scored 2.3% and finished last; GPT got 2.5%. Here is what is behind the numbers.
Alibaba DAMO Unveils First Automotive RISC-V Roadmap for XuanTie Chips
At the RISC-V China Summit in Shenzhen on October 19, Alibaba DAMO Academy unveiled its XuanTie automotive roadmap and claimed gains in Agentic AI and servers.
Zhejiang University Publishes First Survey of Egocentric AI for Smart Glasses
A Zhejiang University team has published the first survey of first-person intelligence for smart glasses, tracing the path from scene recognition to memory, dialogue and action.
Tongyan Lingzhi moves AI personalization from prompts into the model layer itself
Chinese startup 通衍灵智 wants to grow a separate model for each person, project and company instead of one for everyone. We unpack the thesis and the open questions.
Baidu, Tsinghua and Renmin Open-Source LoopVL, Teaching a Loop Transformer to See
Baidu, Tsinghua University and Renmin University's Gaoling School have open-sourced LoopVL, a vision-language model built on a Loop Transformer with shared weights.
GPT-6-Astra Maxes Out Robotics Benchmarks but Stumbles on Real Robots
GPT-6-Astra embodied AI test: OpenAI's model tops classic benchmarks but regularly fails on physical robots. What this means for VLA and WAM.
The Information: Anthropic Is Preparing a Model Stronger Than Any Current Claude
The Information reports Anthropic is preparing a new model whose reasoning and "intuition" rank above every commercial Claude. No figures or name yet.
Tencent Hunyuan and Fudan Launch ExplorationBench to Measure Exploration in LLMs
Fudan University and Tencent Hunyuan unveil ExplorationBench, a benchmark for how AI systems explore an unfamiliar environment, not just how well they answer questions.
WorldMemBench Tests Whether World Models Remember Rooms and Events
The WorldMemBench benchmark scores world-model memory on two axes: whether previously seen scenes persist, and whether events keep progressing while the camera looks elsewhere.
University of Pittsburgh Looks for a 3D Map Inside a VLM: How a Model Sees a Room
Haoming Wang and Wei Gao test whether vision-language models store a room's objects as a map of adjacency relations or just as a bag of colors and shapes.