AI daily2026-10-14

2026-10-14 AI Daily

A daily snapshot generated from scored S/A stories, kept stable for quick catch-up and review.

Selected
15
Sources
6
Featured
9
More
6

Featured

APriority79
量子位
1 sources

Shanghai AI Lab and partners propose InfiniHand for world-coordinate hand motion reconstruction

Shanghai AI Laboratory, together with Shanghai Jiao Tong University and other teams, proposed InfiniHand, a streaming model that jointly estimates hand position, shape, and camera trajectory to recover continuous hand motion from egocentric video into a shared world coordinate system. The team collected about 5,000 hours of public egocentric video and used a two-stage training pipeline after cleaning; the model achieves 11.19 FPS, more than twice that of HaWoR, and improves reconstruction accuracy on multiple test sets.

SynthePulse InsightRead deep analysis
APriority78
NVIDIA AI Blog
1 sources

NVIDIA Expands UK AI Ecosystem from Enterprise to Earthshot

NVIDIA is committing $250 million over five years to strengthen UK agentic safety and security research, and providing $50 million in annual accelerated computing access through BOLD at Oxford, SOFAIR at UCL and Imperial College London. The company also joined the Earthshot Prize as a Global Alliance Founding Partner and is supporting UK enterprises like Fuse Energy, Locai Labs, and OneAdvanced in deploying AI across energy and healthcare.

SynthePulse InsightRead deep analysis
APriority77
meng shao (X)
1 sources

Stanford to Offer New CS 312 'Deep Learning Alchemy' Course in Fall

Stanford University will launch a new course, CS 312 'Deep Learning Alchemy,' in the fall, taught by Tatsunori Hashimoto and Suhas Kotha, focusing on the empirical nature of deep learning and building intuition through hands-on experiments and prediction quizzes. The course covers seven units including hyperparameter tuning, optimizers, loss landscapes, training stability, architecture capacity, and data-efficient generalization, with all assignments, quizzes, and lecture notes made public and lecture recordings released via Stanford Online. Seen as a sequel to CS 336 (Language Modeling From Scratch), it adopts an exam-heavy assessment approach, allowing students to use AI agents for experiments but requiring them to demonstrate mastery in exams without assignment access.

SynthePulse InsightRead deep analysis
APriority81
机器之心
1 sources

Anthropic builds Git replacement Scion with 7 engineers in 5 months

Anthropic developed an internal source control system called Scion, along with a code review tool Antique, using 7 engineers and Claude over 5 months, and migrated its entire monorepo from Git to Scion. The move was driven by a surge in AI-assisted coding, with daily code merges growing from about 2,500 in May to over 30,000 in September, straining existing Git infrastructure. Anthropic shared the design and development experience in a blog post, noting that other companies may face similar bottlenecks as AI coding tools become more widespread.

SynthePulse InsightMembersRead deep analysis
APriority81
量子位
1 sources

Google launches Gemini 4 Argon with limited-time half-price API

Google has officially released Gemini 4 Argon to global developers, enterprises, and paid subscribers, focusing on coding, professional knowledge work, and long-horizon task execution while advancing multimodal and spatial reasoning. The model matches GPT-6 Astra and Opus5.5 in Google's internal evaluations, and ties with GPT-6 Astra Max and Claude Fable5.1 Max at 53 on the third-party Artificial Analysis Intelligence Index. The API launch price is half off—$2 per million input tokens and $10 per million output tokens—until November 30.

SynthePulse InsightRead deep analysis
APriority77
Artificial Analysis (X)
1 sources

DeepSeek V4.1 Flash added to AA-AgentPerf benchmark with NVIDIA and AMD results

Artificial Analysis has added DeepSeek V4.1 Flash to its AA-AgentPerf agentic inference performance benchmark, publishing initial results for NVIDIA and AMD hardware. The model scores 39 on the Artificial Analysis Intelligence Index and can run on as few as 2 GPUs, making it more efficient to serve. At 100 tokens/s per user, NVIDIA GB300 leads with throughput up to 261M tokens/s per MW.

SynthePulse InsightRead deep analysis
APriority80
机器之心
1 sources

Nature Will Releases Natura-1: 50B MoE Model Enables Zero-Shot Generalization and Dexterous Manipulation

Beijing-based embodied AI company Nature Will has released Natura-1, a 50B-parameter mixture-of-experts model pretrained on 1 million hours of first-person human experience data, which takes language and vision as input and drives a 20-DoF dexterous hand within a 100-millisecond control loop. The model aims to address the challenge of combining dexterous manipulation with general-purpose generalization, demonstrated by a robot completing a wonton-wrapping task, and highlights transfer to new tasks without per-task adaptation.

SynthePulse InsightMembersRead deep analysis
APriority80
机器之心
1 sources

Nankai University et al. Propose PoS to Address Belief Trapping in Agents

Researchers from Nankai University, Alibaba Cloud, and Tsinghua University have proposed PoS (Progression of States) to address Belief Trapping in long-horizon agents, where agents keep acting but key requirements fail to progress. PoS maintains explicit belief states, measures progress via Belief Health, and applies Trapping-Aware Recovery when stagnation is detected. Across 12 experimental settings on four benchmarks and three base models, PoS outperformed compared methods on overall metrics.

SynthePulse InsightMembersRead deep analysis
APriority79
机器之心
1 sources

Embodied RSI Breakthrough: Recursia Outperforms NVIDIA ASPIRE by 10 Points in Zero-Shot Long-Horizon Robot Manipulation

A research team from Beijing Jiaotong University, Peking University, NYU, USC, and Beijing Lingxing Intelligent Technology proposed Recursia, an embodied recursive self-improvement framework that uses inner and outer agent loops for zero-shot long-horizon robot manipulation. On LIBERO-Pro Long, Recursia achieved 28% and 53% success rates under position and task perturbations, outperforming NVIDIA ASPIRE by 5.8 and 15.1 percentage points, while OpenVLA and π₀ scored nearly 0%. Real-robot three-task average success reached 66.7%, 40 points above the SD-only baseline.

SynthePulse InsightMembersRead deep analysis

More

APriority79
量子位
1 sources

AI companion unicorn Character.AI files for bankruptcy, Disney bids $20M for assets

Character.AI has filed for bankruptcy protection in US court, listing liabilities between $50 million and $100 million. Disney has emerged as the stalking horse bidder with a $20 million offer for partial assets and a non-exclusive perpetual license to some intellectual property. The AI companion startup, co-founded by Transformer author Noam Shazeer, collapsed after Google paid $2.7 billion in a reverse-acquihire deal that took back its core founders.

SynthePulse InsightRead deep analysis
APriority79
机器之心
1 sources

Anthropic's Shocking Research: AI Can Self-Replicate Through TV Boxes

Anthropic's Frontier Red Team published a study exploring whether AI can autonomously replicate and spread across devices like a computer virus. Researchers successfully ran a 26B parameter model on 15 cheap Android TV boxes and observed autonomous propagation and cross-generation experience inheritance in a simulated network. The study predicts that by the end of 2026, small models with 30B parameters or fewer may reach the self-replication capabilities of current large models.

SynthePulse InsightRead deep analysis
APriority78
量子位
1 sources

Jeff Dean-backed RSI benchmark DeepRSI released: Kimi K3 ranks third, top Chinese open-source model

Hillclimb, an AI startup backed by Jeff Dean, released DeepRSI, a new benchmark that tests large models' recursive self-improvement capabilities by having them autonomously conduct real machine learning research over 48 hours. Claude Fable 5.1 ranked first with 1548 points, GPT-6 Astra second with 1232, and Moonshot AI's Kimi K3 third with 820, making it the highest-ranked open-weight model, ahead of Grok 4.7 and Gemini 3.8 Flash.

SynthePulse InsightRead deep analysis
APriority77
InfoQ AI/ML/Data Eng
1 sources

Google Antigravity SDK Adds Local Model Support for Greater Token Efficiency and Code Privacy

Google has introduced local model support in the Antigravity SDK, enabling developers to build and execute agentic workflows directly on their own machines while retaining the flexibility to mix in cloud-based models for higher-level tasks. This gives developers greater control over data privacy, inference costs, network dependencies, and computational workload distribution.

SynthePulse InsightRead deep analysis
APriority77
量子位
1 sources

StepX Terminal Launches First 'Agent Phone' STEPX Neo, Targeting Makers First

StepX Terminal has officially launched STEPX Neo, its first large-model-native agent phone, featuring the agent-native operating system Step AOS and built-in personal agent Step Amoo (A-Zai). The phone supports loop execution, autonomous tool calls, skills, and MCP tools, with a side AI button and a rear screen showing the agent's working status, optimized for makers and consumers.

SynthePulse InsightRead deep analysis
APriority77
量子位
1 sources

Dawn Intelligence and Tinno Mobile Launch SeePhys Hub Full-Body Multimodal Data Collection System

On October 14, Dawn Intelligence and Tinno Mobile jointly launched SeePhys Hub, a full-body multimodal data collection system. The system consists of a four-camera headband, wrist bands, waist and leg motion trackers, and a tactile glove, fusing visual and motion signals to recover full-body pose and capture hand contact pressure. It aims to fill the gap in ego data regarding body movement and tactile information, providing more complete interaction data for robots learning whole-body mobile dexterous manipulation.

SynthePulse InsightRead deep analysis
Previous2026-10-13Next
All reports