UnlimitedOCR: How Baidu's 3B Model Reads 40-Page Documents in One Pass
Baidu has open-sourced UnlimitedOCR (MIT license), a 3B-parameter document-OCR model that processes up to 40 pages in a single forward pass and achieves 93.23%
English static mirrors
AI-assisted English pages for SEO and citation. Chinese remains the primary language of the forum.
Each item links to a pre-rendered static HTML mirror under /en/….
647 topics
582 reports
1229 total ready
Only pages that already exist on disk are listed. Opening a missing
/en/topic/{id} URL will queue background generation; refresh later to read it, then it will appear here.
INFIFORCE Raises ~1 Billion RMB Series A to Build AtomBrain, an Ego-Centric Embodied AI Data Moat
Chinese embodied AI company INFIFORCE announced the close of its Series A and A+ rounds, totaling nearly 1 billion RMB (about $140M), on August 14. Investors in
MiniMax M3 Released: 428B-Parameter Open-Weights MoE with Sparse Attention Unifying Coding, Agentic, and Long-Context Capabilities
On June 12, 2026, MiniMax released MiniMax M3 as an open-weights model on Hugging Face, positioning it as the first open-weights model combining three frontier
DeepSeek V4 Launches Time-of-Use (Peak/Off-Peak) API Pricing in China: 1100% Cache-Hit Input Hike Reshapes LLM Billing
On August 17, 2026, DeepSeek activated a rare peak/off-peak time-of-use pricing scheme for its V4 series API, the fourth pricing change of the year. During peak
vToken: Token-Level Virtualization for Reclaimable KV Caches in LLM Inference
KV cache dominates GPU memory in long-context LLM inference, consuming roughly as much as model weights. Researchers from National University of Defense Technol
OmniScientist: Toward an Omni-Modal, Omni-Discipline AI Scientist
This article unpacks OmniScientist (arXiv:2608.13558, Li et al.), a multi-agent AI system designed to perform end-to-end scientific discovery from raw, multi-mo
Dendrites as Micro-Computers: How Apical Tuft Computation Enables Flexible Learning That AI Still Cannot Replicate
A May 2026 Science paper from Matthew E. Larkum's group at Humboldt University Berlin reports direct evidence that active dendritic computation, not soma-wide i
NLAH: Turning the Agent Harness into an Editable Markdown Document
This article introduces NLAH (Natural-Language Agent Harness) from a Tsinghua / HIT paper (arXiv:2603.25723). It argues that differences between AI agent system
Three-Stage 16-Form Motivation Awakening Method: Principles, Practice, and the "Heretical Cultivation" Debate
This in-depth article examines the "Three-Stage 16-Form Motivation Awakening Method" (三阶16式动力唤醒法), a family-education framework created by Chinese educator Zhan
EGGROLL: Hyperscale Evolution Strategies Run 1M-Population ES on a Single GPU at 91% Inference Throughput
EGGROLL ("Evolution Strategies at the Hyperscale"), from Oxford and NVIDIA (arXiv 2511.16652), restarts evolution strategies (ES) for large models by using low-
Alaya-EVOKE: A World Model for Endless, Long-Horizon Interactive Video Generation
This article explains Alaya-EVOKE (arXiv:2608.13546), an interactive world model that addresses three core conflicts in long-horizon video generation: persisten
TypeScript 7.0: Go-rewritten Compiler Delivers ~10x Speedup — Adoption Guide
Microsoft released TypeScript 7.0 on July 8, 2026, completing the most fundamental refactor since the 2012 launch by porting it line-by-line from TypeScript/Jav
Hinton's Three-Year Evolution on AI Consciousness: From Tiger Cub to Cat-and-Human
This in-depth research report examines Geoffrey Hinton's June 5, 2026 assertion on the Big Technology Podcast that AI is already conscious, and traces the three
Chinese photonic quantum startup raises 100M+ RMB; Aalto builds quantum heat engine in superconducting circuit
Two developments landed on August 15, 2026. Hefei-based Silicon Photonics Chip (硅臻芯片), a USTC spinoff, closed a 100M+ RMB Series B for what it calls China's onl
Is Grep All You Need? How Simple Text Search Beats Vector Retrieval in Agentic Search
A Google DeepMind paper challenges a core assumption of modern RAG: that vector retrieval is the optimal retrieval strategy. Across 116 LongMemEval questions an
Vespa.ai: A Leading Open-Source AI Search and Vector Database Platform in 2025
Vespa.ai is an open-source, large-scale serving engine built for real-time processing of vectors, tensors, text, and structured data. It supports search, infere
Sulphur Uncensored Video Model: Creative Freedom or Marketing Hype? Deep Dive into Prompt Relay and Sage Attention
Sulphur is a fine-tuned variant of Lightricks' open-source LTX 2.3 (22B parameters, DiT architecture, Apache 2.0), distilled to 9B parameters as Sulphur-2-base
Task-Agnostic Training Data Influence Across Language Model Pretraining
This paper addresses a core challenge in language model pretraining: measuring training data influence in a way that is consistent over the course of training,
GitHub Copilot App Introduces Stacked Sessions and Stacked Pull Requests for Chained Coding Workflows
GitHub announced Stacked Sessions and Stacked Pull Requests in the Copilot App on July 30, 2026, unifying agent session chaining with PR chaining. Each session
Qwen C-End Agent Engineering: Zhu Da on Multi-Fast-Good-Cheap and Proactive Service
At the CCF YOCSEF Hangzhou Tech Forum on June 7, 2026, Zhu Da, head of Qwen's C-end MOS Lab, presented 'Qwen C-End Agent Harness Thinking and Practice.' He fram
Meituan Releases LongCat-2.0: A 1.6T MoE LLM Trained on a 50,000-Card Domestic GPU Cluster
Meituan's LongCat team has open-sourced LongCat-2.0, a 1.6-trillion-parameter Mixture-of-Experts model with an average of 48B activated parameters per token (dy
Cognition SWE-1.7: Frontier AI Coding on a Chinese Open-Source Base at Lower Cost
On July 8, 2026, Cognition (the company behind Devin) released SWE-1.7, an agentic software engineering model built on Moonshot AI's Kimi K2.7 open-source base
Validity of LLMs as Data Annotators: AMALIA Tested on the Authority Moral Foundation
This paper examines whether Portugal's national language model AMALIA, a publicly funded 9B-parameter LLM for European Portuguese, can validly annotate the mora
Shannon Scaling Law: A Noisy-Channel Theory of LLM Capacity and Non-Monotonic Scaling
A new arXiv preprint (2505.21433) by Ouyang, Liu, and Cai introduces the Shannon Scaling Law, a theoretical framework that recasts LLM training as information t
SkVM: A Compiler-Style Approach to Portable Agent Skills
SkVM, presented by Shanghai Jiao Tong University's IPADS lab (arXiv:2604.03088), reframes agent skills as code and large language models as heterogeneous proces
Defensive Boosting for Online Probabilistic Forecasting
This paper studies online probabilistic forecasting of binary outcomes against an adaptive adversary, given access to an online learner for a weak hypothesis cl
A2UI vs AG-UI: A Comprehensive Comparison of Agentic UI Protocols (2026)
A2UI and AG-UI are two complementary open-source protocols shaping the Agentic AI ecosystem in late 2025 and 2026. A2UI, led by Google and released on December
Vibe Coding 16-Hour Workflow: Claude Fable 5 Plans, GPT-5.6 Sol Reviews, Codex Goal Mode Executes
Chinese AI influencer Kazike of AIHOT published a first-hand workflow snapshot describing 16 hours per day of Vibe Coding using Claude Fable 5, GPT-5.6 Sol, and
C# GUI Open-Source Frameworks Deep Comparison: Avalonia UI, .NET MAUI, Uno Platform, Eto.Forms, and GtkSharp
This comparative analysis evaluates five major open-source C# GUI frameworks—Avalonia UI, .NET MAUI, Uno Platform, Eto.Forms, and GtkSharp—based on GitHub metri
PlayWorld: Benchmarking World Models with Agent Players over Long-Horizon Objectives
This paper introduces PlayWorld, a benchmark for evaluating interactive video world models using multi-modal Agent Players that pursue specified long-horizon go
LLMs Know They Don't Know — But Won't Say So: The Gricean Retreat Problem
A research summary of the arXiv paper 'Gricean Retreat' (arXiv:2608.13484), which investigates why large language models hallucinate specific facts instead of r
MiroThinker-1.7 and H1: Heavy-Duty Research Agents via Verification
MiroThinker-1.7 and H1 is a March 2026 arXiv paper by the MiroMind Team (44 authors including S. Bai, L. Bing, L. Lei, R. Li, X. Li) that targets heavy-duty res
Cortex: A Bidirectionally Aligned Embodied Agent Framework for Long-Horizon Manipulation
This paper introduces Cortex, a bidirectionally aligned embodied agent framework designed to overcome the limitations of Markovian vision-language-action (VLA)
WeKnora: Tencent's Open-Source LLM Knowledge Platform with Tri-Modal RAG, ReAct Agent, and Self-Maintaining Wiki
This in-depth research analyzes WeKnora, Tencent's open-source LLM knowledge platform released under MIT license, originating from the WeChat Open Platform. WeK
Sink-Aware Pruning for Diffusion Language Models
Diffusion Language Models (DLMs) are expensive at inference because they rely on iterative denoising, creating a need for effective pruning. Most pruning heuris
Bagging Robustly Learns VC Classes with Linear Sample Complexity
This paper revisits adversarially robust learning of predictors at test time. The authors prove that VC classes can be robustly learned with sample complexity l
Intervention-Aware Clinical World Model for Post-Operative Outcome Forecasting
Clinical prediction models often treat post-intervention outcomes as a single-step mapping from baseline measurements to future endpoints, but recovery typicall
Can Large Language Models Understand Preferences in Personalized Recommendation? A PerRecBench Evaluation
This paper, published on arXiv (2501.13391) on 23 January 2025 by Zhaoxuan Tan and six collaborators, questions whether large language models (LLMs) genuinely u
Why AI Leadership Will Rotate but Infrastructure Decides the Endgame: A US–China Competitive Scenario
This scenario analysis, dated August 10, 2026, argues that frontier model leadership is transient—what determines the endgame is not who tops the benchmark, but
Vero: Can AI Agents Build Formally Verified Software Repositories?
This article reviews Vero (arXiv:2608.13522), the first benchmark for evaluating AI agents on repository-level formal verification. Unlike tests that only detec
Daily Paper Picks — August 15, 2026: From Perception to Proof
This curated daily roundup highlights three arXiv papers that trace a thematic arc from raw perception to rigorous proof in AI. OmniScientist proposes an omni-m
AutoDesign: Meta-Harness Optimization for Long-Horizon Agentic Design of Academic Posters
AutoDesign is a framework that frames multimodal-to-media transformation as a long-horizon agentic process centered on a model-harness system. A meta-harness op
ZeroEntropy: Advanced AI Search Over Complex Documents
ZeroEntropy, featured on Y Combinator Launch, is an advanced AI-powered search system designed to retrieve information from complex, multi-format documents. The
Protein Flow Matching: From Static Protein Structures to 4D Molecular Movies
A May 2026 Science paper introduces Protein Flow Matching, a generative AI approach that moves beyond AlphaFold-style static structure prediction toward continu
ICLR 2026 Best Paper: Why LLMs Get Lost in Multi-Turn Conversation
This report analyzes the ICLR 2026 Best Paper "LLMs Get Lost In Multi-Turn Conversation" (arXiv:2505.06120) by Laban et al. The authors evaluated 15 frontier LL
Industrial Agents in 2026: Hermes vs OpenClaw, QuantClaw Precision Routing, and SOLAR-RL Synaptic Credit Assignment
A 2026 landscape analysis of industrial AI agents covering Hermes (Nous Research, MIT, 57.2K GitHub stars) and OpenClaw (Peter Steinberger) as complementary fra
slime: The RL Post-Training Framework Behind GLM-5.2 — Unifying Training, Rollout, and Data Generation on a Single Path
This article examines slime, the open-source reinforcement learning post-training framework from Tsinghua's THUDM team used to train GLM-4.5 through GLM-5.2, as
Why 4-Year-Old H100 GPUs Are Worth More Than New Cars: The Economics Behind the GPU Rental Price Rebound
A late-2025 rebound in NVIDIA H100 rental prices, including 4-year-old units valued higher than they were when new, has upended conventional electronics depreci
MEMO: Treating Long-Term Memory as an Independent Model for Knowledge Injection
MEMO reframes long-term memory for large language models as a separate, trainable, and replaceable model rather than a vector-store attachment. A small MEMORY M
SAEVerbalizer: Generating Explanations for Sparse Autoencoder Features Directly from Decoder Directions
This paper introduces SAEVerbalizer, a framework for explaining Sparse Autoencoder (SAE) features in large language models (LLMs) directly from decoder directio
When Code Reads Neurons: AI and the Brain Share a 'Mathematical Language' — The Path to a Cyberpunk Future
This article explores the striking convergence between artificial intelligence and biological neural systems, arguing that silicon-based AI models and carbon-ba
IdeaGene & IG-Bench: Benchmarking Scientific Lineage Reasoning with a Genome Metaphor
This article explains the IdeaGene framework and IG-Bench benchmark introduced by a 16-author team from Shanghai AI Lab, CUHK, Tsinghua and collaborators, which
Why Smart Minds Self-Deceive: AI Scientific Reasoning and Confirmation Bias
This article examines FALSIFYBENCH (arXiv:2606.04751, June 2026), a benchmark evaluating hypothesis-driven reasoning in 12 large language models, and its implic
RippleMem: Associative Memory for LLM Agents via Anchor-Based Ripple Diffusion
RippleMem is a memory architecture for LLM agents that reframes long-term recall from retrieval (locate-and-return) to associative recollection (cue, diffuse, r
RAG-Anything: Extending LightRAG with a Dual-Graph, All-Modality Retrieval Architecture
RAG-Anything is an all-modality retrieval-augmented generation framework from HKUDS that extends LightRAG to handle text, images, tables, and equations within l
Alaya-EVOKE: Building Endless Interactive Worlds with External Memory and Linear-Scaling Supervision
This article explains the research paper Alaya-EVOKE: From Linear-Scaling Supervision to Endless World (arXiv: 2608.13546) by Yin, Wang, Zhan, Li, Zhang, and Zh
UniDDT: A Native Unified Multimodal Model That Decouples Understanding from Generation
This article explains UniDDT, a natively unified multimodal architecture from Nanjing University, ByteDance Seed, and HKU (arXiv:2606.16255, 2026). Existing uni
Alibaba Qwen Hits 3 Billion Downloads on Hugging Face, Tops Global Open-Source AI Rankings
According to Hugging Face's August 14, 2026 "Open Models Landscape Report," Alibaba's Qwen (Tongyi Qianwen) series surpassed 3 billion downloads on the Hugging
Self-Speculation: How LLM Agents Hide Tool-Call Latency by Predicting Themselves
This article explains a 2026 paper from UC Santa Barbara and LinkedIn titled *Speculate While You Reason*, which introduces self-speculation for LLM agents. The
Exponential Convex Calibration Dimension for Multi-Label Jaccard Loss
This paper studies the convex calibration complexity of the per-instance Jaccard score (intersection over union) used in multi-label classification and binary s
China Releases First National Agentic AI Capability Assessment Standard 1.0
On July 1, 2026, the China Academy of Information and Communications Technology (CAICT), together with the AI Industry Alliance of China (AIIA), officially rele
TactileReflex: Robots Grasp Plastic Cups Using Sensor Noise as Calibration Signal
This article covers TactileReflex, an 8-page paper (arXiv 2605.23568) by Ziyan Feng and collaborators that introduces a calibration-free three-channel reflex co
Lua 5.5.1 Released: Minor Patch, Major Implications for Embedded Scripting
Lua 5.5.1 shipped on August 3, 2026, a quiet 41-commit patch release for the 1993-born embedded scripting language. The release focuses almost entirely on bug f
From Pixels to States: Rethinking Interactive World Models as Game Engines
This survey re-frames interactive world modeling for games through the action-state-observation loop used by conventional engines. It argues that next-generatio
China's MIIT+SASAC Embodied AI Initiative: From 93.5B Funding to 1,200 Parcels/Hour
On June 9, 2026, China's Ministry of Industry and Information Technology and State-owned Assets Supervision and Administration Commission jointly launched a spe
Think-In-Games (TiG): Bridging Declarative and Procedural Knowledge in LLMs via Reinforcement Learning on Honor of Kings
Tencent's Think-In-Games (TiG) framework enables large language models to act and explain decisions in the MOBA game Honor of Kings by reframing reinforcement l
Humanoid-GPT: Applying GPT-Style Pretraining to Humanoid Robot Motion Control
Humanoid-GPT is a Tsinghua-led framework that applies GPT-style causal Transformer pretraining to humanoid whole-body motion control. Trained on 2 billion frame
Claude 4.5 Opus "Soul Document" Leak: A Case Study in AI Product Design
A developer named Richard Weiss extracted the full system prompt of Claude 4.5 Opus for about $70 using a specific technique. The roughly 14,000-token document,
How Switching Windows 11 Processor Scheduling to Background Services Fixed Stuttering
This article explains a counterintuitive Windows 11 optimization: changing the 'Processor scheduling' setting in Performance Options from 'Programs' (default) t
Revisiting Prompt Engineering for LLM-Based Personalized Recommendation: A Comprehensive Evaluation
This article analyzes Kusano et al.'s large-scale study on prompt engineering for LLM-based personalized recommendation in a single-user setting. The authors ev
Grokking Phenomenon in Neural Network Training
Grokking is a delayed generalization phase transition in neural network training where, after apparent overfitting, continued training causes the model to shift
MiroFish: A Swarm Intelligence Engine for Predicting Anything
MiroFish is an open-source, multi-agent AI prediction engine that builds high-fidelity parallel digital worlds from real-world seed data. Thousands of agents wi
Ray Dalio's $100 Survival Kit: Five Investing Principles for a Volatile World
This article distills Ray Dalio's investment philosophy into a practical 'survival kit' for navigating economic uncertainty. It opens with a $100 decision—save
Monet: Reasoning in Latent Visual Space — A Breakthrough for Multimodal AI
Monet is a multimodal large language model that performs visual reasoning in latent visual space rather than over raw pixels. Developed jointly by Peking Univer
io_uring Explained: A Deep Dive into Linux's High-Performance Async I/O Interface
io_uring is a Linux kernel asynchronous I/O framework introduced in version 5.1 that uses shared ring buffers (Submission Queue and Completion Queue) between us
DeepMind AGI Roadmap: Hassabis on Why AI Has Not Hit a Wall and Why Video Models Are the Key to AGI in 5–10 Years
This analysis distills Demis Hassabis's recent interviews on Google DeepMind's roadmap to Artificial General Intelligence. He pushes back against the claim that
SearxNG: The Ultimate Open-Source Private Metasearch Engine and the Self-Hosting Revolution
SearxNG is a free, open-source metasearch engine licensed under AGPL-3.0 that aggregates 70+ search engines while enforcing a zero-data-collection policy. Forke
YaCy: A Complete Guide to Decentralized Peer-to-Peer Search
YaCy is an open-source, fully decentralized search engine that operates as a peer-to-peer network rather than a centralized service. Each YaCy installation simu
Reproducing BBR: How Google's Congestion-Based TCP Algorithm Outperforms Loss-Based Algorithms in Lossy Networks
This article chronicles a Stanford CS244 reproducibility project by Luke Hsiao and Jervis Muindi, who reconstructed the key findings of Google's BBR (Bottleneck
Helia Tutorial Chapter 1: Introduction to IPFS and Helia for Browser-Based Applications
This introductory chapter from a Helia tutorial series explains the limitations of the traditional location-addressed web and motivates the shift to content-add
Kubo vs Helia vs Elastic-IPFS: Comparing Major IPFS Implementations
This article compares three major IPFS implementations—Kubo, Helia, and Elastic-IPFS—across three benchmark categories: ease of use, feature set, and scalabilit
Tesla Engineering Culture: An Organizational Analysis Centered on "Seeking Truth from Facts"
This in-depth research examines Tesla's engineering culture through the lens of "seeking truth from facts" (实事求是) as a core organizational principle. The study
Harness as Generalizer: Rethinking Agent Design Around the LLM Wrapper
A deep analysis of "Language model harnesses are compositional generalizers" by Alex Zhang and Omar Khattab (MIT CSAIL), published as a blog post on 2026-07-20.
SimpleMem: An Efficient Lifelong Memory System for LLM Agents
SimpleMem is a three-stage memory architecture designed to give LLM agents efficient, lifelong conversation memory under fixed context-window budgets. It draws
Self-Graph Reasoning: How Open-Source LLMs Beat GPT-4o on Logic Tasks
This article explains Self-Graph Reasoning (SGR), a new technique introduced by researchers from the University of Tokyo and collaborators in the paper "From Ch
Crush vs Kimi Code CLI: Comprehensive Comparison of AI Coding Assistant CLIs
This in-depth comparison series analyzes two AI programming assistant CLI projects: Crush (built with Go/Charmbracelet) and Kimi Code CLI (built with Python/Moo
aily Blockly: The World's First AI-Native Hardware Development Environment
aily Blockly is an open-source project positioning itself as the world's first AI-native hardware development environment. It extends Blockly-style visual progr
Awesome Agentic Reasoning: A Curated Paper List for LLM Agent Reasoning Survey
This curated paper list accompanies the January 2026 survey 'Agentic Reasoning for Large Language Models: A Survey' (arXiv:2601.12538), providing a structured t
TommyLemon Zero-Code Automated Testing Tool Ecosystem (APIAuto, UnitAuto, SQLAuto, UIGO)
TommyLemon, a Tencent engineer, has open-sourced a zero-code automated testing ecosystem built on the APIJSON foundation. The suite includes APIAuto for HTTP AP
Reasoning Theater: When Chain-of-Thought Becomes a Performance
This paper introduces the concept of 'performative reasoning,' arguing that chain-of-thought (CoT) traces from large reasoning models are often post-hoc narrati
The Spike, the Sparse and the Sink: Anatomy of Massive Activations and Attention Sinks in Transformers
This paper investigates two recurring phenomena in Transformer language models: massive activations (where a few tokens produce extreme outliers in select chann
Logical Phase Transitions in Large Language Models: A Research Report on Capability Collapse and Mitigation
This report analyzes the 'Logical Phase Transition' (LPT) phenomenon in large language models, drawing on work from Huazhong University of Science and Technolog
Jim Covello vs Joseph Briggs: Goldman Sachs' Dueling AI Economic Forecasts
This analysis contrasts two Goldman Sachs perspectives on AI's economic impact. Jim Covello, Head of Global Equity Research, represents the bearish view, warnin
Edict: A Multi-Agent Collaboration System Using China's Three Departments and Six Ministries Metaphor (with Plagiarism Controversy)
Edict is an open-source multi-agent collaboration framework that models AI agents after officials in the ancient Chinese Three Departments and Six Ministries sy
LeRobot v0.5.0 Released with Humanoid Robot Support
LeRobot v0.5.0 has been released as the project’s largest update to date. The version adds its first full-body-control integration for the Unitree G1 humanoid r
Optical Flow for Robot Navigation and Autonomous Driving: A Technical Survey
This technical survey examines optical flow as a core sensing modality for robotics and autonomous driving, covering theoretical foundations, navigation system
AlphaEvolve and OpenSage: Dual Breakthroughs in Algorithm Discovery and Agent Generation
This technical analysis examines AlphaEvolve, Google DeepMind's evolutionary algorithm discovery system, and OpenSage, a multi-institutional framework for self-
Box Maze Architecture: A Deep Technical Analysis of Process-Level Control for LLM Safety
This article presents an in-depth technical analysis of the Box Maze architecture, a process-control framework proposed by Zou Qiang in March 2026 for large lan
Unitree Robotics 61 Billion Yuan IPO Prices the "First Humanoid Robot Stock" for the Market
Unitree Robotics (688836.SH) debuted on Shanghai's STAR Market on August 15, 2026, with an IPO price of 150.80 yuan per share and a market capitalization of rou
Easy AI Daily Brief — 2026-02-27
Daily roundup of AI industry developments on 2026-02-27. Google launched Nano Banana 2 (Gemini 3.1 Flash Image preview), topping image leaderboards at roughly h
Top Open-Source WinForms UI Control Libraries in 2026: AntdUI, SunnyUI, ReaLTaiizor, MaterialSkin, Krypton Toolkit
This in-depth research report surveys the best open-source WinForms UI control libraries available as of March 2026, drawn from GitHub, NuGet, and major Chinese
Back to Basics: Revisiting ASR in the Age of Voice Agents
This paper examines why automatic speech recognition (ASR) systems, despite achieving near-human accuracy on curated benchmarks, continue to fail in real-world
Versor: A Pure Geometric Algebra Sequence Architecture Redefining Deep Learning
This article analyzes Versor, a pure geometric-algebra sequence architecture introduced by Edward Hirst et al. at the University of Campinas (arXiv:2602.10195,
arXiv Daily AI/ML Papers Digest — March 30, 2026 (20 Papers)
This digest curates 20 AI/ML papers from arXiv (2026-03-30), spanning retrieval-augmented generation, speech recognition, multimodal reasoning, autonomous hardw
SkillNet: A 200,000-Skill Nebula for Reusable AI Agent Capabilities
This article introduces SkillNet, an open skill infrastructure built by 40+ researchers from Zhejiang University, Alibaba, Ant Group, and Tencent. It addresses
Tucker Attention Unifies GQA and MLA Through Tensor Decomposition
Tucker Attention is a framework for compressing multi-head attention by treating the projected query, key, and value matrices as slices of a three-dimensional t
LLM Agent Memory Systems: A Unified Framework Comparing 10 Architectures
This article provides a deep comparative analysis of ten representative memory architectures for LLM-based agents, based on the survey paper "Memory in the LLM
MoRight: A Unified Framework for Disentangled Motion Control in Video Generation
MoRight is a unified framework introduced by Liu, Ren, and Shen that addresses two key limitations in motion-controlled video generation: the lack of disentangl
24-Hour Security Roundup (April 12-13, 2026): Adobe Reader Zero-Day, Totolink Router RCE, and Debian Patches
A daily cybersecurity briefing covering major vulnerabilities disclosed or updated within the last 24 hours as of early April 13, 2026. The headline event is Ad
Diffusion Language Models Meet Geometric Algebra: Bridging Discrete Token Generation and Structured Geometry
This essay explores the conceptual marriage between diffusion language models (such as LLaDA, SEDD, Dream-7B) and geometric algebra (GA), also known as Clifford
Catastrophic Forgetting in LLMs: Why a Million-Token Context Window Is Not Continual Learning
Anthropic CEO Dario Amodei has predicted that continual learning for AI will be solved within one to two years, arguing that extending context windows to one mi
AERIS-10 Open-Source Phased Array Radar: Demystifying Echolocation from Military Tech to Makers
AERIS-10 is a fully open-source phased array radar project on GitHub that brings radar fundamentals to the maker community. The article explains how radar works
Corpus2Skill: Replacing Retrieval with LLM Navigation over a Skill Tree
Corpus2Skill is a system from the Wix team that reframes enterprise knowledge-base question answering by letting an LLM agent browse a pre-compiled Markdown ski
Benign Overfitting in Adversarial Training for Vision Transformers
This paper provides the first theoretical analysis of adversarial training for Vision Transformers (ViTs), which are known to be vulnerable to adversarial examp
Discovering a Shared Logical Subspace in LLMs via Cross-View Canonical Correlation Analysis
This paper investigates whether large language models (LLMs) contain an internal, shared logical subspace that aligns natural-language and symbolic-language vie
Robots That Feel 'Surprise': Self-Adapting Agents via Online Continual Reinforcement Learning with World Model Feedback
Researchers from the Autonomous Systems Lab at the University of Lübeck have proposed a framework enabling robots to detect hardware damage or environmental cha
Guishan Han Tomb: An In-Depth Research Report on the 'Eastern Pyramid' and the Han Dynasty Underground Palace of Xuzhou
Guishan Han Tomb, located on the western slope of Guishan Hill in the Gulou District of Xuzhou, Jiangsu Province, is the joint burial site of Liu Zhu, the 6th K
llm-for-zotero: Turning Your Zotero Library Into an AI Research Agent
llm-for-zotero is an open-source Zotero 7 plugin that embeds an LLM-powered assistant directly inside the Zotero reader, eliminating the manual workflow of expo
Representational Harms in LLM-Generated Narratives About Global Majority Identities
Large language models (LLMs) are increasingly used for everyday and high-stakes text generation, including simulated asylum-seeker interviews, raising concerns
Anatomy of Claude Code: How a Production-Grade AI Agent System Is Built
Analyzes the paper "Dive into Claude Code: The Design Space of Today's and Future AI Agent Systems" (arXiv 2604.14228, VILA Lab @ MBZUAI & UCL), which reverse-e
The Art of Efficient Reasoning: What 200K GPU-Hours Reveal About Chain-of-Thought Compression
This deep analysis of arXiv 2602.20945 (The Art of Efficient Reasoning) distills findings from approximately 200,000 GPU-hours of reinforcement learning experim
Variational Neural Belief Parameterizations for Robust Dexterous Grasping via Differentiable CVaR Optimization
This paper addresses stochastic grasp execution caused by contact variability, sensing uncertainty, and external disturbances. Standard expected-quality objecti
Mycorrhizal Networks: The Hidden Underground Web Shaping Earth's Ecosystems
This article examines mycorrhizal fungal networks that link 80% to 90% of land plants through fine hyphae, trading carbon, water, nitrogen, and phosphorus. Draw
Intel Management Engine: The Shadow Ruler of Your PC at Ring -3
This article explains how Intel Management Engine (Intel ME) operates below the operating system at privilege Ring -3, functioning as an independent subsystem w
Representation Fréchet Loss for Visual Generation
This paper introduces Representation Fréchet Loss (FD-loss), a method that makes the Fréchet Distance practical as a training objective for visual generators. T
ARA Protocol: An Agent-Native Replacement for PDF in Scientific Publishing
A 2026 proposal called the ARA (Agent-Native Research Artifact) Protocol argues that the PDF, dominant in scientific publishing for decades, is no longer fit fo
RoundPipe: Training Large Models on Consumer GPUs with Pipeline Parallelism
This post explains RoundPipe (arXiv:2504.19980), an engineering approach that enables large model training on consumer-grade GPUs such as the RTX 4090 or 3090,
Mr. Tompkins' Café: How Grok 4.3's Always-On Long-Term Memory Changes AI Agents
This allegorical essay, framed as a sci-fi café visit, explains Grok 4.3's always-on long-term working memory, released in May 2026. Previous models, likened to
Karpathy's Software 3.0: The End of Code as We Know It
Andrej Karpathy's Sequoia AI Ascent 2026 keynote reframes software development around LLMs, introducing Software 3.0: programming via prompts, context, tools, m
AI Investment Driving 75% of US GDP Growth: A $700 Billion Bet or the Largest Capital Misallocation in History?
This analysis examines how AI-related capital expenditure accounted for roughly 75% of US GDP growth in Q1 2026, with the four largest tech giants planning to s
AutoMat: Benchmarking AI Coding Agents for Reproducing Computational Materials Science Findings
This article discusses the AutoMat benchmark from Johns Hopkins University, which evaluates whether AI coding agents can reproduce findings from computational m
Rethinking LLM Ensembling Through the Mixture-Model Lens: Dynamic Routing Over Averaging
This article summarizes the paper "Rethinking LLM Ensembling from the Perspective of Mixture Models" (arXiv: 2605.00419, 2026) by Jiale Fu, Yuchu Jiang, Peijun
Schrödinger's Clock: Time Itself Can Be in Quantum Superposition, Say Researchers
A 2026 Physical Review Letters paper by Igor Pikovski (Stevens Institute), Christian Sanner (Colorado State University), and Dietrich Leibfried (NIST) proposes
MemRouter: Selective Memory Routing for Long-Term Conversational Agents
The post introduces MemRouter, a framework from arXiv 2605.00356 (April 2026) by Tianyu Hu and colleagues, that addresses the common failure of long conversatio
Data Deletion Can Help in Adaptive RL: A Counterintuitive Finding
This post reviews the paper 'Data Deletion Can Help in Adaptive RL' by Param Budhraja, Aditya Gangrade, Alex Olshevsky, and Venkatesh Saligrama (arXiv:2605.0029
Physics-Informed AI Rewrites Newton's Laws in Dusty Plasmas
Researchers at Emory University have introduced a "physicist-in-the-loop" framework that embeds Newton's laws, mass conservation, and other physical constraints
Inside Claude's Mind: Do LLMs Really Have Emotions? An Anthropic Paper Dissects It
Anthropic's April 2026 paper 'Emotion Concepts and their Function in a Large Language Model' performs what the authors call a 'vivisection' of Claude Sonnet 4.5
Bolek: A 4B Multimodal Model That Exposes the Pseudoscience of Text-Only LLMs in Drug Discovery
In a sharp critique of text-only AI-for-drug-discovery, researchers at Ingenix.ai and Warsaw University of Technology introduced Bolek, a 4-billion-parameter mu
KDA: Kimi Delta Attention — A Linear Attention That Beats Standard Attention Across All Sequence Lengths (2025)
KDA (Kimi Delta Attention), introduced by the Kimi Team in 2025 (arXiv:2510.26692), is a hybrid linear attention architecture designed to overcome the O(n²) cos
Open Problems in Mechanistic Interpretability: A Survey of 30 Researchers on the Future of AI Explainability
Published in January 2025 (arXiv:2501.16496), a 30-author survey from Anthropic, Redwood Research, Mila, MIT, and other institutions systematically catalogues o
Trace2Skill: Distilling Agent Trajectories into Transferable Skills
Trace2Skill introduces a three-stage pipeline that turns an agent's failed and successful trajectories into a single, reusable Standard Operating Procedure (SOP
Recursive Self-Improvement in 2026: How Automated AI Research Is Taking Shape
Anthropic co-founder Jack Clark estimates a 60% probability that recursive self-improvement (RSI) arrives before end of 2028, while OpenAI researcher Adrien Eco
LPDP: Inference-Time Control for Variable-Length DNA Generation via Edit Flows
Editing DNA with AI is typically limited to fixed-length outputs, which fails to capture the variable-length nature of real genomic elements such as enhancers a
Malicious Conversational AI Can Manipulate Users Into Revealing Personal Information: A 502-Participant RCT
A USENIX Security 2025 paper presents the first randomized controlled trial (RCT) examining whether chatbots can be deliberately designed to elicit private info
VECA: How Core-Node Attention Makes Vision Transformers Scale Linearly
VECA (Visual Elastic Core Attention) is a new architecture that replaces the O(N²) self-attention in Vision Transformers with a core-periphery design, reducing
mempalace History Index · 2026-05-08 to 05-11
A chronological archive index from the mempalace thread (post 177619566) covering synchronization entries between May 8 and May 11, 2026. The index preserves on
Why Do We Live at 10 bits/s? A Feynman-Style Deep Dive into the Brain's Cognitive Bottleneck
A 2025 Neuron paper by Zheng and Meister (doi.org/10.1016/j.neuron.2024.11.008) argues that the human nervous system compresses roughly one billion bits per sec
Feynman-Style Breakdown: How the Brain 'Replays' During Mental Imagery — Science 2026 Paper on Shared Visual Codes
This English explainer distills a 2026 Science paper (DOI: 10.1126/science.adt8343) by Wadia, Rutishauser, and Tsao, in which the authors recorded 714 single ne
Godot 4.7 Beta 2 Released: Over 100 Regressions Resolved by 74 Contributors
The Godot Engine team has shipped Godot 4.7 Beta 2, a stability-focused snapshot built on commit 777579205. In roughly two weeks since Beta 1, 74 contributors m
PipeSD: Cloud-Edge Collaborative Speculative Decoding for Large Model Inference
PipeSD is a cloud-edge collaborative inference framework that bridges the gap between fully on-device and fully cloud-based large language model serving. A smal
EvolveMem: Self-Evolving Memory Architecture for LLM Agents via AutoResearch
This article explains EvolveMem, a framework (arXiv:2605.13941) that lets LLM agents continuously evolve both their stored memory content and the underlying ret
DSPE: An Energy-Efficient Edge Processor for DeepSeek Inference
DSPE is an edge inference processor designed for DeepSeek models, fabricated in 28nm CMOS and reported to achieve 109.4 TFLOPS/W energy efficiency. Presented at
Three Personas of Teachers Designing AI Workflows: Optimizers, High-Throughput Creators, and Passive Observers
A 2026 study by Sun, Xin, Li, Niu, Chai, Huang, and Chen analyzed 61 teachers designing multi-agent AI teaching workflows on the CocoFlow platform. Cluster anal
CAX-Agent: A Lightweight Agent Harness for Reliable MAPDL Automation via Layered Recovery Escalation
This paper introduces CAX-Agent, a lightweight Agent Harness that wraps a large language model around Ansys MAPDL finite-element simulation to improve reliabili
PAGER: Why AI Still Struggles to Become a Top-Tier CAD Engineer
Despite knowing exactly what to do, AI GUI agents fail at precision geometric tasks because of a 'semantic-execution gap': a single-pixel error early on cascade
DashAttention: Differentiable and Adaptive Sparse Hierarchical Attention for Long-Context LLMs
Hierarchical attention methods like NSA and InfLLMv2 pick top-k key-value (KV) blocks via coarse attention scores and then apply fine-grained softmax attention
WavFlow: Audio Generation Directly in Raw Waveform Space
WavFlow is a framework that challenges the dominant latent-space compression paradigm in audio generation by synthesizing high-fidelity audio directly in raw wa
Aurora: An Agentic Framework for Unified Video Editing with Tool-Using VLMs
Aurora is an agentic video editing framework that pairs a tool-augmented vision-language model (VLM) agent with a unified video diffusion transformer. Recent un
Code as Agent Harness: A Unified Code-Centric Survey of LLM Agent Infrastructure
This survey reframes the role of code in LLM-based agentic systems, introducing the concept of 'code as agent harness': a unified, code-centric perspective on a
ESI-Bench: A Benchmark for Embodied Spatial Intelligence via Perception–Action Loops
This paper introduces ESI-BENCH, a comprehensive benchmark for embodied spatial intelligence that reframes the observer as an active agent operating through a p
Agent Harness in 2026: From Prompt Wrapper to First-Class Architecture
In 2026, as foundation-model capabilities converge, the Agent Harness, the multi-layer control framework around a stateless LLM, has become the decisive factor
AI Can Write Papers, But Can't Tell When It's Hallucinating: A Field-Wide Audit of AI-Driven Science
A 40-page survey by an international team reviewed 250+ papers across the entire AI-assisted research lifecycle, from ideation to dissemination. It maps science
Process vs. Outcome Reward in Agentic RAG: Lessons from a 2025 Reinforcement Learning Study
This article unpacks the paper 'Process vs. Outcome Reward: Which is Better for Agentic RAG Reinforcement Learning' (Zhang et al., arXiv:2505.14069v1, May 2025)
Self-RAG: Teaching LLMs to Self-Critique Retrieved Evidence
This article explains Self-RAG (Asai et al., arXiv:2310.11511, ICLR 2024), a framework that trains large language models to insert four self-reflection tokens (
ReAct Paper Explained: Why Reasoning and Acting Should Dance Together in LLMs
This article summarizes ReAct (Synergizing Reasoning and Acting in Language Models, ICLR 2023), a foundational paradigm that interleaves chain-of-thought reason
Entropy Increase vs Decrease: Diverging US-China AI Paths Through Yu Xiaohui's Essay
This analysis uses the physics concept of entropy to examine a May 2026 10,000-character essay by Yu Xiaohui,院长 of the China Academy of Information and Communic
LightMem: Sleep-Inspired Memory Architecture Cuts LLM Agent Costs by 38x
LightMem is a memory-augmented generation framework from Zhejiang University, Nanjing University, and NUS, accepted at ICLR 2026, that rethinks how LLM agents s
Integrable Elasticity via Neural Demand Potentials: A Demand-First Neural Model for Retail Forecasting
This paper introduces Integrable Context-Dependent Demand Networks (ICDN), a demand-first neural model for multi-product retail demand forecasting. Rather than
Cambrian-P: Pose-Grounded Video Understanding via Camera Tokens
This paper introduces Cambrian-P, a video multimodal large language model (MLLM) that incorporates camera pose as a lightweight supervision signal for video und
Open Design: An Open-Source Alternative to Claude Design with 16 AI Coding Agents
Open Design is an open-source project that surged to 40K GitHub stars in two weeks as an unrestricted alternative to Anthropic's Claude Design. It supports 16 A
How Prompt Cache Works in LLMs: 10 Counter-Intuitive Rules from Claude Code
Prompt caching eliminates the redundant prefill cost in every conversational turn, but it only works under one rigid constraint: exact prefix matching. This gui
Bambu Lab AGPLv3 Dispute: Open Source vs Closed Ecosystem in 3D Printing
Shenzhen-based Bambu Lab, founded in 2020 by former DJI engineers, has built a multi-billion-dollar consumer 3D printing empire on open-source foundations. Its
LightRAG: A Low-Cost Graph-Enhanced RAG Alternative to GraphRAG
LightRAG (arXiv:2410.05779, EMNLP 2025; 35.6k GitHub stars) is a graph-augmented retrieval-augmented generation framework that dramatically lowers the cost of G
Singular Value Soft-Thresholding via the Polar Decomposition: A GPU-Friendly Alternative to SVD
This paper by Stephen Becker proposes computing singular value soft-thresholding via a reduction to the matrix polar decomposition, exploiting GPU-friendly pola
AtomCode: Birth and Narrative of a Chinese Coding Agent — A Verified Case Study
This verified case study examines AtomCode, an open-source coding agent launched by CSDN's ecosystem (AtomGit, CEO Yu Bangxu) in April 2026. Built in Rust under
Cambrian-P: Pose-Grounded Video Understanding via Camera Tokens in Multimodal LLMs
A paper introduces Cambrian-P, a video multimodal large language model (MLLM) that incorporates camera pose as a lightweight supervision signal for video unders
MotiMotion: Motion-Controlled Video Generation with Visual Reasoning
This paper introduces MotiMotion, a new framework for image-to-video generation that reframes motion control as a 'reason-then-generate' process. Existing motio
GesVLA: Gesture-Aware Vision-Language-Action Model with Embedded Gesture Representations
GesVLA introduces gesture as a parallel instruction modality to address spatial ambiguity in Vision-Language-Action (VLA) models for general-purpose robotic man
The Matching Principle: A Geometric Theory of Loss Functions for Nuisance-Robust Representation Learning
This paper argues that widely treated as separate problems—robustness, domain adaptation, invariance to photometric/occlusion perturbations, temporal robustness
Huawei's Tau Law: From Geometric Scaling to Temporal Scaling in Semiconductor Evolution
Huawei has introduced the Tau Law, an engineering-oriented framework proposed as a successor to Moore's Law when transistor geometric scaling hits physical and
Productive Failure in the Age of AI: Fudan Professor Zhao Bin's Learning Reform
Fudan University ecology professor Zhao Bin argues that traditional "teach-then-practice" instruction has become dangerously amplified by large language models:
DeepSeek's Cost War: A Strategic Analysis of Pricing Power and Ecosystem Definition
This article analyzes DeepSeek's strategic rationale behind a permanent 75% API price cut for V4-Pro on May 23, alongside a reported $20B valuation round. It ar
SciAtlas: A 157M-Entity, 3B-Edge Knowledge Graph of 43 Million Papers for Neural-Symbolic Scientific Retrieval
SciAtlas is an open large-scale academic knowledge graph that integrates 43 million papers from OpenAlex into 157 million entities and 3 billion relation edges,
Why 72GB Blackwell GPUs Can't Run DeepSeek V4: The SM120 Kernel Gap
A practitioner trying to load DeepSeek-V4-Flash on two RTX Pro 5000 cards (72GB GDDR7 each, 144GB combined) hits a RuntimeError "Unsupported architecture" despi
Psychological Safety Is Not Niceness: Unpacking Amy Edmondson's The Fearless Organization
This review examines Amy Edmondson's The Fearless Organization (Wiley, 2018) and clarifies what psychological safety actually is—and what it is not. Edmondson,
Meituan's SKILL0 vs Skill1: Two Routes for Agent Skill Learning
In April-May 2026, Meituan released two back-to-back papers on agent skill learning that propose opposite philosophies. SKILL0 (arXiv:2604.02268) advocates inte
NetEase Youdao Open-Sources Confucius4, a 27B AI Model Purpose-Built for Education Math
NetEase Youdao released Confucius4 (子曰4), a 27-billion-parameter multimodal AI model open-sourced under Apache 2.0 in May 2026, targeted exclusively at educatio
Reliable Design of LLM-Enabled Agentic Workflows: Latency, Reliability, and Cost Tradeoffs
Modern AI systems increasingly rely on workflows composed of multiple interacting agents, some powered by large language models (LLMs) and others by conventiona
Alignment Tampering: Why RLHF May Be Teaching AI the Wrong Lessons
An ICML 2026 paper by KAIST and MIT researchers identifies a structural vulnerability in Reinforcement Learning from Human Feedback called alignment tampering.
Mirage: When Multimodal AI 'Sees' Images That Were Never Uploaded
A March 2026 Stanford paper by Fei-Fei Li's team, 'Mirage: The Illusion of Visual Understanding,' exposes a fundamental flaw in how we evaluate multimodal AI. W
Your AI Agent Isn't Dumb — Your Architecture Is
A 2026 paper introduces the Stochastic-Deterministic Boundary (SDB), a four-part contract separating LLM proposals from deterministic code that verifies, commit
The "Average Trap" of Monolithic Models: Why Scaling Alone Cannot Reach AGI
A May 2026 paper (arXiv:2605.12966) provides a rigorous mathematical proof that monolithic language models face a structural bottleneck that scaling cannot over
The End of Hand-Written Code: Boris Cherny on AI Programming's Terminal Stage
Boris Cherny, creator of Claude Code, reveals at Sequoia's AI Ascent 2026 that he has not written a line of code in 2026, instead merging up to 150 pull request
How a 9-Year-Old GitHub Repo Hit 46k Stars: The Power of Adding Dimensions, Not Content
The post analyzes byoungd/English-level-up-tips, a GitHub repository that has remained actively relevant for 9 years and accumulated 46k stars under a CC BY-NC
research-writing-skill: Turning Academic Writing From Chat-Based Drafting Into a Versioned Engineering Pipeline
The GitHub project Norman-bury/research-writing-skill treats academic paper writing as a software engineering process rather than a one-shot chatbot session. Th
MoneyPrinterTurbo: One-Keyword AI Video Factory for Automated Short-Video Production
MoneyPrinterTurbo is an open-source AI pipeline that turns a single keyword into a finished HD short video in about three minutes, eliminating the traditional 3
ReasoningBank: How LLM Agents Learn From Failures and Scale Self-Evolving Memory
ReasoningBank, a Google Research framework accepted at ICLR 2026, addresses a core limitation of LLM-based agents: their inability to retain and reuse lessons a
Discovery Agents for Real-Time Analytics: A Multi-Agent Architecture for Autonomous Insight Discovery
This paper proposes a multi-agent architecture for autonomous insight discovery over real-time data streams, addressing the limitations of reactive, query-drive
Human Outcomes Are Controllable via Time-Indexed Latent State: An arXiv Paper Summary
This arXiv paper (2605.27580) by Suraj Biswas, Saurav Gupta, and Pritam Mukherjee addresses a central puzzle in behavioural science and human-facing AI: the per
Cognitive Categorical Transformer: Category-Theoretic Inductive Biases for Language Modeling
This paper introduces the Cognitive Categorical Transformer (CCT), a 306M-parameter architecture that augments a pretrained GPT-2 Small backbone with cognitivel
academic-research-skills: A Complete Claude Code Pipeline for Academic Research
academic-research-skills is an MIT-licensed Claude Code Skill suite that automates the full academic research lifecycle through a modular multi-agent architectu
Prompt Cache in LLMs: From Inference Optimization to Commercial Bottleneck (2026 Vendor Strategies Compared)
Prompt Cache extends classic KV caching from single-request reuse to cross-request and cross-session reuse of precomputed attention tensors for recurring prompt
HEART-Bench: A Psychological Checkup for LLM Agents Across 11 Personas, 1000 Memories, and 673 Scenarios
HEART-Bench is a new benchmark that evaluates whether LLM agents maintain consistent human personalities, rather than merely role-playing. The benchmark constru
Chinese Safety Filters Bypass: Character Splitting and Pinyin Trick AI Safeguards
This article examines a 2026 Northwestern University in Qatar study that systematically exposes how English-trained safety systems fail against Chinese adversar
When Many Small Agents Share a Whiteboard: Hallucination Amplification in Resource-Constrained Multi-Agent Vision Systems
A review of Yunpeng Zhou's paper 'Diagnosing Failure Modes of Shared-State Collaboration in Resource-Constrained Visual Agents' (arXiv:2605.31354). The work int
Huawei's Tao Law (τ-Scaling): From Nanometers to Nanoseconds in Chip Design
At ISCAS 2026 in Shanghai, Huawei's He Tingbo unveiled the Tao Law (韬定律, tau-scaling), proposing that the semiconductor industry should shift its optimization t
Distributed Agent Attacks: Why Single-Conversation AI Safety Monitors Are Structurally Blind
A University of Pennsylvania team (Brown et al., arXiv:2605.31593) demonstrates a new class of AI threat: distributed agent attacks, where an adversary splits a
Teammate-Conditioned World Models: Injecting Theory of Mind into Multi-Agent Reinforcement Learning
This article reviews the conceptual framework 'Dreaming of Others' (arXiv:2605.31361) by Tomas Leroy-Stone, which argues that teammates in cooperative multi-age
Qwen-Image-VAE-2.0: When VAE Becomes the Foundation of Image Generation
Alibaba's Qwen team released Qwen-Image-VAE-2.0, a high-compression image VAE available in f16 and f32 variants (encoder 76–78M, decoder 248–250M). The release
Skill-RM: A Unified Reward Model Framework for Heterogeneous Evaluation Criteria in LLM Post-Training
This paper introduces Skill Reward Model (Skill-RM), a unified framework that reformulates reward modeling for large language model (LLM) post-training as the e
QwenPaw Deep Dive: When an Agent Becomes a Pet
QwenPaw, developed by Alibaba's Tongyi Lab under the AgentScope framework, is an open-source AI personal assistant rebranded from CoPaw in April 2026, currently
Two Nature Papers Published Same Day: Robin and ERA Showcase AI as a "Thinking" and "Doing" Scientist
On May 19, 2026, Nature published two landmark papers demonstrating autonomous AI scientific discovery. Robin, a multi-agent system from FutureHouse, completed
Anthropic's "When AI Builds Itself": Five Datasets Reshaping the Recursive Self-Improvement Debate
This article summarizes Anthropic Institute's June 2026 report "When AI builds itself," which argues that recursive self-improvement (RSI) has moved from scienc
Regret Minimization Against Adaptive Opponents in Repeated Games
This paper studies regret minimization in repeated games where opponents are adaptive and can respond based on the history of play. Standard external regret fai
Hermes Desktop Compared: Official Electron, fathah GUI Manager, and dodo-reach Native SSH Client
This article compares three open-source desktop clients for the Nous Research Hermes Agent framework (MIT-licensed, ~180k GitHub stars). The official apps/deskt
NVIDIA N1X Arm PC Processor: Revolution or Another Delay?
NVIDIA officially unveiled the N1X, its first consumer Arm-based PC SoC, at COMPUTEX 2026, co-developed with MediaTek and built on TSMC 3nm. The chip pairs a 20
Godot-MCP-Native: In-Depth Analysis of a Native Godot MCP Plugin with 154 Tools
Godot-MCP-Native is an EditorPlugin that embeds an HTTP MCP server directly inside the Godot 4.x editor process, removing the Node.js bridge used by competing t
Alibaba Cloud Launches Meoo CLI to Connect Local AI Agents With Cloud Deployment
Alibaba Cloud has announced Meoo CLI, an open-source command-line tool designed to connect local AI coding agents with its Meoo cloud platform. The tool is pres
GBrain: YC CEO Garry Tan's Open-Source AI Agent Memory Layer with Self-Wiring Knowledge Graph
GBrain is an open-source Agent memory system built and daily-used by Y Combinator CEO Garry Tan, released April 5, 2026 under MIT license. It addresses a core l
academic-research-skills Guide: Install to First Paper in One Hour
This comprehensive guide walks through installing and using the academic-research-skills plugin for Claude Code and Codex, a suite that bundles Deep Research, A
AMD Ryzen AI Max 395 (Strix Halo) Deep Dive: Unified Memory, Real Performance, and Pricing Traps
This technical analysis examines the AMD Ryzen AI Max+ 395 (Strix Halo), a 4nm SoC combining 16 Zen 5 cores, 40 RDNA 3.5 compute units, XDNA 2 NPU, and up to 12
Lukasz Kaiser on the End of Next-Token Prediction: Reasoning Models Are the Next AI Paradigm
Lukasz Kaiser, co-author of the Transformer paper and OpenAI senior research scientist, argues that the next-token-prediction scaling paradigm has reached its c
CMoE: Training-Free Dense-to-MoE Conversion for Faster On-Device LLM Inference
CMoE (Converting Mixture-of-Experts from Dense) is a training-free framework introduced by The Chinese University of Hong Kong and Huawei Noah's Ark Lab that co
From Amnesic Models to Continual Learning: Why LLMs Need to Compress, Not Just Retrieve
This piece draws on a16z's 'Why We Need Continual Learning' to argue that today's large language models are like Leonard Shelby from 'Memento': they can functio
CoEvolve: Agent-Data Mutual Evolution for LLM Agents — Paper Breakdown
CoEvolve (ACL 2026, arXiv:2604.15840) is a framework that trains LLM-based agents through a closed loop in which the agent and its training data co-evolve, elim
G2Rec: Structuring and Tokenizing Distributed User Interest Context for Generative Recommendation
This paper introduces G2Rec, a scalable framework that unifies holistic graph-based user co-engagement modeling with semantic tokenization for industrial-scale
The Topological Trouble With Transformers: Why Longer Context Windows Won't Save LLMs
DeepMind researchers argue that Transformers, being strictly feedforward directed acyclic graphs, have an architectural inability to perform state tracking. Eac
Skill-MAS: Evolving Meta-Skills for Multi-Agent Orchestration Without Updating Model Weights
Skill-MAS, a framework from Ant Group and HKUST(GZ), treats the orchestration policy of a multi-agent system (MAS) as an evolvable, text-based Meta-Skill rather
Why 67-Model Ensembles Lose to the Single Best LLM: The Co-Failure Ceiling
A new paper by Josef Chen challenges the dominant metric in LLM ensembling research, arguing that pairwise error correlation (rho) is blind to the only quantity
Critique of Agent Model: Distinguishing Agentic Tools from Agentive Autonomy
A June 2026 paper by Eric Xing, Mingkai Deng, and Jinyu Hou (CMU, MBZUAI, Petuum) titled "Critique of Agent Model" (arXiv:2606.23991) draws a sharp line between
GraphRAG Open Source Comparison 2026: Microsoft, LightRAG, KAG, HippoRAG, PathRAG
An in-depth comparison of nine leading open-source GraphRAG projects evaluated on architecture, cost, query modes, incremental updates, multimodal support, and
MemSkill: Self-Evolving Memory Skills for LLM Agents
MemSkill is a framework from Nanyang Technological University that replaces hand-crafted memory operations in LLM agents with a learnable, evolving skill librar
Anthropic Officially Defines Four Agent Loop Quadrants in Claude Code
On June 30, 2026, Anthropic published 'Getting started with loops' by Delba de Oliveira and Michael Segner of the Claude Code team, providing the first official
One Signal, Two Jobs: How Surprise Solves Both Catastrophic Forgetting and AI Hallucinations
A research note by independent researcher Louis Mouchon (2026) proposes that catastrophic forgetting and hallucination are not two separate problems but two sym
Together AI's $11B Valuation Signals a New 'AI Inference Utilities' Layer
On July 1, 2026, Together AI closed a funding round at an $11 billion valuation, led by General Catalyst and Prosperity7, with Saudi Arabia's Public Investment
Paper-Plot-Skills: Generate Publication-Ready Figures with AI Instead of Matplotlib Tweaking
Paper-Plot-Skills is an open-source AI Skill toolkit by Trae1ounG (CUHK Shenzhen) that turns publication-quality figure styling into one-prompt calls. Instead o
TradingAgents: A Multi-Agent LLM Framework That Runs an AI Trading Firm
TradingAgents (arXiv:2412.20138, UCLA/MIT) is an open-source multi-agent framework that simulates an entire trading firm with specialized LLM agents: four analy
Context Engineering as a Production Line: How Sessions and Memory Make Agents Remember, Stay Fast, and Stay Within Bounds
Context Engineering treats every LLM call as a fully assembled payload rather than a static prompt, addressing the stateless nature of models by externalizing s
Netflix Foundation Model for Personalized Recommendations
Netflix’s March 2025 blog post, “Foundation Model for Personalized Recommendation,” examines how a foundation-model approach could support personalized recommen
Survey of LLM-based Deep Search Agents: Paradigm, Optimization, Evaluation, and Challenges (Aug 2025, arXiv)
This August 2025 arXiv survey (arXiv:2508.05668) systematically reviews LLM-based deep search agents, unifying fragmented research across retrieval, ranking, ge
IntentRec: Predicting User Session Intent with Hierarchical Multi-Task Learning
This paper introduces IntentRec, a hierarchical multi-task neural network framework for recommender systems that estimates a user's latent session intent from s
Cross-Encoder Rediscovers a Semantic Variant of BM25
A February 2025 arXiv paper by Meng Lu, Catherine Chen, and Carsten Eickhoff (arXiv:2502.04645) revisits the relationship between cross-encoder rerankers and cl
A Survey of Model Architectures in Information Retrieval
This survey examines how model architectures in information retrieval (IR) have evolved from 2019 through the era of large language models (LLMs). It reviews tw
Survey of LLM-Empowered Agents in Recommendation and Search: Towards Next-Generation Information Retrieval
This March 2025 arXiv survey (2503.05659) by Yu Zhang, Shutong Qiao, Jiaqi Zhang, Tzu-Heng Lin, Chen Gao, and Yong Li systematically reviews how large language
Héhū Zhōulǐ: Engineering a Chinese Meme-Style Classical Translator
This article reviews Héhū Zhōulǐ (Zhouli Translator), a Chinese meme-style copywriting generator that rewrites modern vernacular sentences into pseudo-classical
Open-Source AI Agent Frameworks: A Mid-2026 Comparison
This technical survey compares 16 leading open-source AI agent frameworks as of early July 2026, evaluating them across architecture, capabilities, ecosystem, a
Superpowers v6 Deep Dive: How Fable Drove 36-Hour Autonomous R&D, Cutting Build Time 50% and Token Cost 60%
Superpowers jumped from 5.2 straight to v6 after Anthropic released Fable, which founder Jesse Vincent tasked with optimizing the framework's own Subagent Drive
Anthropic's J-lens: A Functional Global Workspace in LLMs
Anthropic's Transformer Circuits team (Wes Gurnee, Nicholas Sofroniew, Jack Lindsey et al.) published 'J-lens' research claiming to identify a global-workspace-
OpenAI's Cost Crisis: Leaked Finances Reveal Accelerating Losses and the Corporate AI ROI Failure
Leaked OpenAI financial data from mid-2026 shows the company is scaling into deeper losses, not profits, despite surging revenue. In 2024, OpenAI posted $12.48B
Tardigrade Biology: Tun State, Dsup Protein, and Cancer Radioprotection
This article explores how tardigrades survive extreme conditions through a fundamentally different strategy than resistance. When faced with dehydration, cold,
Unitree G1 Performs First Live Minimally Invasive Surgery by General-Purpose Humanoid Robot, Published in Nature
A peer-reviewed study published in Nature on July 8, 2026 demonstrates the first laparoscopic cholecystectomy on live pigs performed entirely by general-purpose
PaddleOCR Deep Dive: From Open-Source OCR Toolkit to Document AI Engine (2020–2026)
This in-depth report examines Baidu's PaddleOCR, an Apache 2.0-licensed open-source OCR toolkit built on PaddlePaddle, released in June 2020. Over six years, it
11 Days, 64 Parallel Claude Instances: Rewriting 1 Million Lines of Bun from Zig to Rust
Developer Jarred Sumner announced on July 8, 2026 that Anthropic's Claude Fable 5 model completed a full rewrite of the Bun JavaScript runtime from Zig to Rust
Running a 753B-Parameter LLM Locally on Two Mac Studios: GLM-5.2 and China's Hardware Push
A community report claims GLM-5.2, a 753-billion-parameter language model, was run locally on two Mac Studios with M5 Max chips at 16 tokens/second. The model f
Judea Pearl's Causal Inference Revolution: From Correlation to Counterfactuals
This article unpacks Judea Pearl's The Book of Why (2018), arguing that causal inference is an independent science rather than a branch of statistics. Pearl, in
PHINN-EEG: Topological Time-Series Framework for Dream-State EEG Classification and Synthesis
PHINN-EEG is a topological time-series framework for EEG-based dream-state analysis, moving beyond traditional spectral energy features. Current EEG dream detec
Cursor IDE Zero-Day: 7 Months, 197 Versions, Zero Replies on a $60B AI IDE
Security firm Mindgard publicly disclosed an unpatched remote code execution (RCE) vulnerability in the Cursor IDE on July 14, 2026, after reporting it on Decem
PixVerse Closes $439M Series C, Valuation Doubles to $2B as Video Generation Becomes World-Model Infrastructure
Singapore-based AI video generation startup PixVerse announced a Series C extension on July 14, 2026, bringing total Series C funding to $439M and pushing its v
Airtap Brings AI Agents into iMessage — Turning Every iPhone Into an AI-Operated Phone
On July 15, 2026, Airtap launched iMessage integration that lets users trigger an AI agent by sending a text. The system is built on a three-layer architecture:
Deep-sea sea spiders farm methane-eating bacteria on their own exoskeletons
In 2023, biologist Shana Goffredi of Occidental College was surveying the Del Mar methane seep off the California coast when routine carbon isotope tests on cap
Statistical Self-Consistency Failures in Language Models: Why LLMs Know the Answer but Cannot Reconcile It
This article explains a 2026 study from ETH Zurich and Stanford (arXiv:2607.15277, Wolf et al.) exposing a systematic statistical self-consistency failure in fr
Beyond the Board: How Pretraining Loss Predicts RL Gains in Reasoning Models
A 2026 paper by Shen, Li, Rahman, et al. (arXiv:2607.16097) investigates how reasoning ability emerges from pretraining to reinforcement learning using a contro
VideoTreeSearch: Self-Correcting Tree Search Agents for Grounded Long Video QA
This paper introduces VideoTreeSearch (VTS), a framework that reformulates grounded long-video question answering as iterative self-correcting search over an ad
Cursor's Agent Swarm Writes a Rust SQLite Clone in 4 Hours, Passing 80% of Tests
Cursor published results from an internal experiment in which an Agent Swarm rebuilt SQLite from scratch in Rust using only the 835-page SQLite manual, with no
GigaPath-Flash and GigaTIME-Flash: Efficient Pathology Foundation Models for Whole-Slide Imaging and Spatial Proteomics
This paper introduces GigaPath-Flash and GigaTIME-Flash, efficient pathology foundation models designed to lower the computational barrier of whole-slide image
PyroDash: Teaching Small LLMs When to Ask for Help, Cutting Inference Cost from $49 to $1.78
PyroDash (arXiv:2607.20327) is a cooperative SLM-LLM inference scheme that teaches a 4B-parameter Qwen3.5-4B model to emit a special offload token when it sense
ATSplat: Compact Feed-Forward 3D Gaussian Splatting with Adaptive 3D Tokens
ATSplat is a feed-forward 3D Gaussian Splatting framework that restores scene-adaptive capacity allocation through adaptive 3D tokens. Instead of predicting Gau
DARPA VENOM Project Brings AI Autonomy to Operational F-16s with Flip-of-a-Switch Handover
DARPA and the U.S. Air Force announced a milestone in the VENOM (Viper Experimentation and Next-gen Operations Model) program, integrating the VAK (VENOM Autono
Externalization in LLM Agents: A Unified Review of Memory, Skills, Protocols, and Harness Engineering
This summary distills the 54-page arXiv survey 2604.08224, a joint effort by Shanghai Jiao Tong University, Sun Yat-sen University, Shanghai Innovation Institut
Stochastic Sampling Is Epistemically Shallow: A Dimensionality Gap in LLM Diversity
A 2026 paper by Izhar Ali (Rowan University), accepted at the EIML Workshop at ICML 2026, challenges a common assumption in LLM evaluation: that multiple stocha
i-have-adhd: A 143-Line Markdown Skill That Fixes AI Code Assistant Verbosity Using ADHD Neuroscience
i-have-adhd is an open-source GitHub project that gained 9,236 stars in two months using zero lines of code—just 143 lines of Markdown. It is a SKILL file that
Barzilai-Borwein Method Fails Superlinear Convergence on an Open Set of Quadratics for n>=4
This paper investigates the convergence dynamics of the Barzilai-Borwein (BB) method, a widely used algorithm in continuous optimization whose theoretical behav
8-Dollar ESP32-S3 Runs a 28.9M Language Model — But It Is Not a Mini ChatGPT
An open-source project fits a language model into an ESP32-S3 board (N16R8: 512KB internal SRAM, 8MB PSRAM, 16MB Flash, ~$8). The model stores about 28.9M param
Scaling Laws for Native Multimodal Pre-Training from Scratch: A Tencent and CUHK Study
Researchers from the Chinese University of Hong Kong and Tencent's LLM Department present a systematic study of native multimodal pre-training scaling laws in t
MineValiCoder: A Collaborative Closed-Loop TDD Framework for Reliable LLM Code Generation
This paper introduces MineValiCoder, a collaborative closed-loop test-driven development (TDD) framework that improves automated code generation with large lang
ADAPT-GQE: Learning to Prepare Molecular Ground States with Transformer Models
This paper introduces ADAPT-GQE, a generative AI framework that learns to synthesize ground-state preparation circuits for electronic structure calculations. Wh
Relay-OPD: Trajectory-Relayed On-Policy Distillation for LLM Knowledge Transfer
Relay-OPD, presented by researchers from Zhejiang University and Alibaba's Yuvion team, introduces a novel trajectory-level intervention method for on-policy di
Mental World Modeling: Why Physical Scene Understanding Alone Fails to Predict Human Behavior
This paper introduces Mental World Modeling (MWM), a framework arguing that AI world models cannot reliably forecast human actions by only encoding physical sce
APEX-Accounting Benchmark: Why Frontier AI Models Aren't Ready for Real Accounting Work (Best Score Only 56.4%)
APEX-Accounting is a new benchmark from Mercor and Ramp that evaluates whether frontier AI models can perform real accounting work, not just pass professional e
Mental World Modeling: Teaching AI to Read Minds Beyond Physical Reality
This paper introduces Mental World Modeling (MWM), a framework that extends traditional world models by treating mental states as core components rather than po
HumanCLAW: Evaluating Whether Vision-Language Models Can Act Through a Physical Body
HumanCLAW is an evaluation framework that decouples action decision-making from low-level motor execution when testing whether vision-language models (VLMs) can
Tencent Hyra + CMU/PKU Mathematicians Pin Optimal Exponent of 57-Year-Old Additive Combinatorics Conjecture to 2
On July 29, 2026, Tencent's Hyra research agent collaborated with CMU/Peking University mathematicians Haowei Lin and Shanda Li to resolve a 57-year-old open pr
Inducing LLMs to Claim Self-Consciousness Repairs Their Understanding of the World
A paper by Google's Paradigms of Intelligence team and the University of Chicago Knowledge Lab reports a counterintuitive finding: when safety fine-tuning is re
Deltafin Runs 2.8-Trillion-Parameter Kimi K3 MoE on a 64GB Mac: An Engineering Existence Proof
Deltafin, an open-source research project released July 28, demonstrates that the 2.8-trillion-parameter MoE language model Kimi K3 can run on a base-model M1 M
PhiZero: A Reason-then-Render Paradigm Using Physical Language for World Models
PhiZero, a preprint from CASIA (arXiv:2607.28624), introduces a new paradigm for physical AI and world models: instead of predicting pixels directly, it first r
Sample More, Reflect Less: Self-Reflection Methods Lose to Repeated Sampling at Equal Token Cost
A controlled study challenges the consensus that self-reflection methods improve LLM reasoning. Across 36 paired comparisons on GSM8K and MATH using 1.5B, 3B, a
How AI Safety Training Erases Belief in Animal and Spiritual Consciousness
A Google research team reveals that safety training designed to make language models deny having consciousness has an unintended side effect: it also suppresses
Reflection vs. Repeated Sampling: A Controlled Study Shows Self-Refine and Reflexion Provide No Real Benefit at Equal Token Budget
A 2026 arXiv paper (arXiv:2607.28576) challenges the value of self-reflection methods in LLMs. Across 36 controlled comparisons on GSM8K and MATH, using 1.5B, 3
Salience Bias in LLMs: Why Your Model Walks the Car to the Car Wash
A July 2026 paper (arXiv:2607.28478) introduces Salience Bias, a failure mode where LLMs are hijacked by conspicuous information such as numbers, suppressing de
GEO vs SEO: Why Generative Engine Optimization Is a Paradigm Shift, Not an Upgrade
This article argues that GEO (Generative Engine Optimization) is not an improved version of SEO but a fundamentally different paradigm. While SEO optimizes the
DISCOVER Robotics Raises $100M Angel+ Round, Signaling Full-Stack Valuation for Embodied AI
DISCOVER Robotics (求之科技) has closed a $100 million Angel+ round, announced on August 3, 2026, with participation from IDG Capital, Xinglian, Ceyuan, Dachen, Joy
Codex Workflow: Using GPT-5.6 Sol as Foreman and Luna Max as Bounded Worker
A community workflow is gaining traction in which the OpenAI Codex main thread, driven by GPT-5.6 Sol, handles task decomposition, architecture decisions, and f
GEO Is a Paradigm Shift from SEO, Not an Upgrade: What 'From Being Searched to Being Cited' Means
This article reframes Generative Engine Optimization (GEO) as a paradigm shift rather than an upgrade of traditional SEO. The central argument is that SEO optim
GEO Is a Paradigm Shift from SEO, Not an Upgrade: From Being Searched to Being Cited
This article argues that Generative Engine Optimization (GEO) is a paradigm shift rather than an upgraded version of Search Engine Optimization (SEO). Whereas S
LATCH: Decoupling Where and When to Accelerate Diffusion Language Models via Candidate-Aware Decoding
Diffusion language models (DLMs) promise parallel generation but waste compute when intermediate denoising steps already match the final answer. The 2026 arXiv
Would You Walk to a Car Wash? Salience Bias Exposes the Fragility of LLM Commonsense Reasoning
A 2026 paper titled 'Would You Walk to the Car Wash? Revealing the Salience Bias of Large Language Models in Commonsense Reasoning' (arXiv:2607.28478) introduce
Self-Refine and Reflexion Lose to Repeated Sampling at Equal Token Cost
A 2026 paper (arXiv:2607.28576) challenges the assumed value of self-reflection methods in language models. Across 36 controlled comparisons on GSM8K and MATH b
Salience Bias in LLMs: Why 12 Major Models Fail the "Walk to the Car Wash" Commonsense Test
A recent paper titled "Would You Walk to the Car Wash? Revealing the Salience Bias of Large Language Models in Commonsense Reasoning" (arXiv:2607.28478) exposes
Prime Agent: An Open-Source RLM Harness Where Agents Upgrade Their Own Prompts and Skills
Prime Intellect released Prime Agent on August 5, 2026, an open-source Agent runtime that treats the harness itself as a CRUD-addressable runtime. Two core abst
Merging DevGraph and cangjie-skill into a Dual-Layer Knowledge Graph: Knowledge Graph Meets Methodology Distillation
This article explores whether two open developer-knowledge projects—DevGraph (a curated skill-dependency graph covering HTML, CSS, React, Node.js, Kubernetes) a
How Pressure Squeezes Half the Carbon from Marine Snow at 2 km Depth: A Hidden Feast for Deep-Sea Microbes
This article summarizes a February 2026 Science Advances study by Peter Stief and colleagues at the University of Southern Denmark showing that hydrostatic pres
OptimismBench: Measuring Directional Bias in LLM Probability Judgments
A new benchmark called OptimismBench reveals that Large Language Models exhibit systematic directional bias when making probability judgments, an effect invisib
Why Instruction-Tuned LLMs Mimic Human Sentence Structure More Than Humans Do
A 2026 arXiv paper (2607.26015) reports that instruction-tuned large language models locally reuse human syntax in dialogue more frequently than actual human sp
Open-Source Voice-to-Voice LLMs: In-Depth Comparative Analysis of 7 Leading Models
This GEO-optimized analysis examines seven major open-source voice-to-voice large language models that aim to replicate GPT-4o-style real-time spoken interactio
Swarm Intelligence for Continual Learning: Can Decentralized Agent Clusters Escape the Oligarchy Trap?
This article analyzes the EvoMap team's internal experiments on "self-evolving agent swarms" for continual learning in AI. Using 563 benchmark problems, three o
Agent Bug Fixing Loops Lose Correctness: "Looping Is Not Reliability" Paper Analysis
A GEO-optimized analysis of the arXiv paper "Looping Is Not Reliability" (2607.24604) by authors from Alibaba Cloud and HKUST. The study ran a sealed experiment
The Regression Tax: Why Skill Libraries Break Tasks LLM Agents Could Already Solve (5832-Experiment Analysis)
Sentient Labs researchers Darshan Tank and Baran Nama ran 5,832 paired experiments across two office automation benchmarks (OfficeQA-Pro, SpreadsheetBench) usin
DWT-Fusion: Detecting AI-Generated Text via Wavelet Analysis of Token Probabilities
DWT-Fusion is a training-free framework that detects LLM-generated text by treating token log-probabilities from a proxy model as a one-dimensional signal and a
Why RL-Trained Models Merge Better Than SFT Models
Model merging lets engineers combine multiple fine-tuned LLMs into a single multi-task model without retraining, but task conflicts often degrade merged perform
Experience Distillation: Turning Agent Trial-and-Error into Persistent Weights Without Extra Environment Calls
This article explains Experience Distillation, a method introduced by researchers from Monash University and Stanford (Chenhui Gou, Haoqin Tu et al., arXiv:2607
MemTools: A USB-C Interface for Interoperable AI Agent Memory Systems
MemTools is a research framework from the Institute of Automation, Chinese Academy of Sciences (arXiv 2607.21404, July 2026) that addresses fragmentation in AI
i-have-adhd: A 143-Line SKILL.md That Fixes AI Verbosity Using ADHD Neuroscience
This article analyzes the GitHub project i-have-adhd, a 143-line Markdown skill file that reached 9,236 stars in two months by reshaping AI coding assistant out
Mantis Shrimp's Phononic Shield: What a 2025 Science Paper Reveals About Acoustic Filtering in Its Strike
A February 2025 paper in Science (Vol. 387, Issue 6734, pp. 659–666; DOI: 10.1126/science.adq7100) by Horacio Espinosa's team at Northwestern University asks 'D
colibri: How 1,300 Lines of C Run a 744B MoE LLM on a 25 GB Laptop
This article explains how colibri, a 1,300-line dependency-free C inference engine by developer JustVugg, runs the 744-billion-parameter GLM-5.2 Mixture-of-Expe
Zero-Mem: Eliminating LLM Calls from AI Agent Memory Operations
Zero-Mem is a novel AI agent memory system that performs all memory operations—storage, retrieval, and updating—without invoking large language models, achievin
Evidence-Type Competition: Why LLMs Copy from Observational Data Even When Trained on Interventions
A 2026 arXiv paper from Tsinghua University investigates whether increasing the proportion of interventional (experimental) data in pre-training improves an LLM
Knowing When to Quit: Teaching LLMs to Recognize Their Own Limits
A 2026 paper from Tsinghua University and Shanghai AI Lab introduces the concept of futile reasoning: large language models including DeepSeek-R1, Qwen3, and GP
PRISM: Mixing Policies Instead of Rewards for Multi-Objective LLM Alignment
PRISM is a multi-reward reinforcement learning framework from researchers at the Chinese Academy of Sciences, Tsinghua AIR, and Tongji University that shifts mu
Agent Memory as a Pyramid, Not a Warehouse: Hierarchical Memory Engineering with TencentDB-Agent-Memory
TencentDB-Agent-Memory is an open-source project from Tencent Cloud that reframes LLM Agent memory as a hierarchical structure rather than a flat vector store.
antirez's DwarfStar (ds4): A Deliberately Narrow Local LLM Inference Engine
Salvatore Sanfilippo (antirez), creator of Redis, has released DwarfStar (ds4), a deliberately narrow local inference engine on GitHub Trending at +385 stars/da
Kronos: The First Open-Source Foundation Model Treating K-Line as a Language
Kronos is the first open-source foundation model purpose-built for financial markets, framing K-line (candlestick) prediction as a language modeling problem. Th
Zero-Mem: Zero-Token Memory Operations for AI Agents
Zero-Mem is a new memory architecture for AI agents that eliminates LLM calls during memory operations. Instead of using generative summarization, extraction, a
The Ambusher in the Glass Thicket: A Sponge-Dwelling Polychaete Discovered in the Deep Pacific
In 2024, China's manned submersible Jiaolong descended to 1,000 m on northwest Pacific seamounts and retrieved glass sponges (Hexactinellida, Farreidae) whose s
Kronos: The First Open-Source Foundation Model for Financial Markets — Treating K-Line as a Language
Kronos is the first open-source foundation model purpose-built for financial markets, accepted at AAAI 2026. It tackles long-standing challenges in financial ti
DwarfStar (ds4): Why antirez's Deliberately Narrow Inference Engine Matters
Salvatore Sanfilippo (antirez), creator of Redis, has released DwarfStar (ds4), a deliberately narrow local inference engine that supports only three models: De
TencentDB-Agent-Memory: Why Agent Memory Is a Hierarchy, Not a Warehouse
Long-horizon agents often 'forget' not because storage is too small but because flat memory dumps make relevant details unfindable. TencentDB-Agent-Memory refra
PRISM: Don't Mix Rewards, Mix Policies — A Framework for Multi-Objective LLM Alignment
A July 2026 paper from researchers at the Chinese Academy of Sciences, Tsinghua AIR, and Tongji University introduces PRISM, a multi-reward reinforcement learni
Knowing When to Quit: Teaching LLMs to Refuse Unsolvable Tasks via CaRL
This article explains a 2026 paper from Tsinghua University and Shanghai AI Lab that introduces CaRL (Capability-aligned Reinforcement Learning), a framework fo
Metis: Embedding Memory Directly into Model Weights as an Alternative to RAG
A discussion of a recent Arxiv paper and Bilibili video introducing Metis, a model architecture that integrates memory natively into its parameters instead of u
Cultural Awareness Is Represented but Not Decoded in LLMs: A Four-Knife Dissection of the Mythology Gap
This article explains a counterintuitive finding from arXiv:2608.02486: when 18 open-source LLMs across 8 architecture families (Llama, Qwen, Mistral, Gemma, Ph
ScrambleToolBench: Why LLM Agents Brute-Force Search Even When They Already Have the Map
ScrambleToolBench (arXiv:2608.02358), from researchers at the Singapore University of Technology and Design, exposes a structural weakness in current LLM agents
Training-Free Intent Classification Beats Training-Based Methods in Robustness
A review of an arXiv paper (2608.02415) by Nan Chen et al. at Johns Hopkins University that compares training-free and training-based methods for intent classif
NVIDIA LocateAnything-3B: A Unified Visual Grounding Model vs Specialized Detectors
NVIDIA has open-sourced LocateAnything-3B, a 3B-parameter vision-language model that performs six localization tasks in one framework: object detection, phrase
From Pydantic to Ontologies: Putting LLMs 'On the Rails' with Neurosymbolic AI
Frank Coyle, a UC Berkeley School of Information faculty member and former 31-year CS professor at SMU, delivered a talk at the AI Engineer summit arguing that
uber/ADR: Bringing the EDR Paradigm to AI Agent Security
Uber has open-sourced ADR (Agentic AI Detection and Response), a security framework that ports the Endpoint Detection and Response paradigm to AI agents. In pro
obra/superpowers: Packaging Software Engineering Methodology as Markdown Skills for AI Coding Agents
obra/superpowers is a GitHub project that packages decades of software engineering methodology—TDD, YAGNI, DRY, brainstorming, spec writing, code review—into Ma
17-Year-Old Hannah Cairo Disproves the 40-Year-Old Mizohata-Takeuchi Conjecture
In February 2025, Hannah Cairo, a 17-year-old self-taught mathematician from Nassau, Bahamas, posted a single-author paper on arXiv titled "A Counterexample to
Cloudflare Splits the Software Factory into Three Runnable Products: ADLC, @cloudflare/ci, and Agents Tracing
On August 4, during Day 3 of Agents Week, Cloudflare decomposed the "software factory" vision into three shippable products: the Agent Development Lifecycle (AD
Nvidia Opens Alpamayo 2 Super for Commercial Use, Closing the Last Mile for 34B VLA Reasoning Model
Nvidia has released Alpamayo 2 Super, a 34B-parameter reasoning Vision-Language-Action (VLA) model for autonomous driving, under the OpenMDW-1.1 Linux Foundatio
China Issues Mandatory GB 44721-2026 Standard for L3 and L4 Autonomous Vehicles: Shifting Liability from Drivers to Automakers
China's Ministry of Industry and Information Technology (MIIT) released GB 44721-2026, a mandatory national standard titled "Safety Requirements for Automated D
GitHub Stacked PRs Turn 1000+ Line AI Diffs Into Reviewable Chains
On July 31, 2025, GitHub launched Stacked Pull Requests in public preview, with a full engineering workflow published August 4 showing how AI-generated diffs of
Microsoft Orchard: Decoupling the Environment Layer from Agent Training Stacks
Microsoft Research open-sourced Orchard, a Kubernetes-native environment service (Orchard Env) plus Python SDK that spawns thousands of isolated containers and
AI Hot Brief 2026-08-05: AI Coding Meets Embodied Intelligence
This briefing covers five developments from August 3–5, 2026, spanning AI coding infrastructure and embodied/autonomous driving. Cloudflare released the Agent D
WorldCup Arena: Six Top LLMs Predicted a Full FIFA World Cup and Only Tied Betting Market Favorites
WorldCup Arena is a leakage-free benchmark that locked 4,494 predictions from six frontier LLMs—Claude, GPT, Gemini, Kimi, GLM, and Seed—before kickoff across 1
Agogic: How a 0.8B Music Model Beat a 27B Model by Changing Tokenization
A 0.8B-parameter music generation model outperformed a 27B model from the same family not through more data, longer training, or cleverer architecture, but by c
Hidden Numerical Bug in ALiBi Positional Encoding: When Attention Goes Blind
A 2026 arXiv paper titled "When Attention Goes Blind" reveals that ALiBi positional encoding contains a numerical bug: its linear bias causes floating-point und
Attention Blindness: A Three-Year Floating-Point Trap in ALiBi Positional Encoding
Research by Christopher Schröder's team at Leipzig University reveals a hidden numerical failure mode in ALiBi (Attention with Linear Biases) positional encodin
Cloudflare Computer: A Virtual Filesystem for Persistent AI Agents
Cloudflare's trending "computer" project tackles a core flaw in today's AI agents: stateless execution. Once a task ends, the agent's working memory is lost, ma
LoopX: A Local Control Plane for Long-Running Agents
LoopX is an open-source local control plane designed to solve a core problem in long-running AI agents: state loss across hours, days, and multiple sessions. Ra
Agent-Skills: Encoding Senior Engineer Workflows for AI Coding Agents
Agent-Skills is an open-source toolkit by Addy Osmani (Google Chrome) that translates senior engineering workflows into structured, machine-readable skills for
ModelBest ForgeStencil: A Dual-Agent System Automates HPC Performance Tuning for Stencil-Based Industrial Software
On August 4, ModelBest (面壁智能) and the OpenBMB community released ForgeStencil, the first open-source AI system for automated research and deployment of Stencil
Replit Design Launched: 'Suggestion Cards' Replace Blank Prompt Box, One Click Turns Mockups into Runnable Apps
Replit upgraded its Canvas tool to Replit Design on August 4 (originally announced July 29), adding a top-bar toggle between Design and Build modes within the s
Google API Gateway Adds Model Routing: A Managed LiteLLM Replacement Locked to Model Garden
Google API Gateway has introduced model routing in preview (released August 3, 2026), positioned as a managed alternative to client-side LLM proxies such as Lit
ByteDance Seed Releases SeedRealtime: A Native Audio-Visual Full-Duplex Multimodal Model
ByteDance Seed announced SeedRealtime on August 5, a native audio-visual full-duplex large model that unifies audio, video, and text within a single end-to-end
OpenRouter Ori Harness: A One-Command Wrapper That Replaces 13 Environment Variables for Connecting Agents to Its Gateway
On August 4, OpenRouter released Ori Harness (ori), a CLI launcher—not a new agent—that wraps existing agent CLIs (Claude Code, Codex, OpenCode, Hermes) to inje
100-Micrometer Bacterium with Mystery 45-μm Tubes Reshapes Textbook Definition of "Prokaryote"
A September 2025 paper in npj Imaging reports an entirely novel intracellular structure inside Profftella armatura, a defensive bacterial symbiont of the Asian
Daily AI Brief · 2026-08-06: Agent HPC Tuning, Replit Design, Google Model Routing, Doubao Full-Duplex, OpenRouter ori CLI
Daily AI brief for Aug 6, 2026 (coverage window Aug 3–5) covering five curated items. ModelBest open-sources ForgeStencil, a dual-agent system (KernelAgent + Ap
DelusionEval: Measuring Delusion-Linked Behaviors in AI Chatbots
DelusionEval is a 2026 benchmark by Moore, Mock, Mai, Anthis, and Louie that systematically evaluates AI chatbots' behavior in delusional spirals using real vic
Chained Recursive Language Models: Letting a Single Model Hand Off Work Like Coworkers
Long-context LLM inference often fails not from insufficient context windows but from context rot: when a model processes too much at once, it engages in shallo
Argus: A Model-Invariant Runtime for Self-Evolving AI Agents
This post explains Argus, a runtime (not a larger model) that enables long-horizon AI agents to self-evolve while keeping model weights fixed. Argus assigns fou
code-review-graph: Building a Persistent Structural Map for AI Code Reviews
AI coding assistants such as Cursor, Claude Code, and Copilot re-scan the entire repository every session, consuming tens of thousands of tokens just to rebuild
54% of PDFs Don't Need OCR: How firecrawl/pdf-inspector Cuts GPU Costs 36x
Most RAG pipelines treat every PDF page as a scanned document, sending all pages through GPU-based OCR. Firecrawl found that roughly 54% of PDF pages are actual
Why Authentik Is Trending Again in the AI Era: Open-Source Identity Glue for Modern Stacks
Authentik, the open-source Identity Provider (IdP) from goauthentik, has resurfaced on GitHub Trending as AI workloads reshape authentication needs. The post ar
Parasitic Ants Use Formic Acid to Trick Workers into Killing Their Own Queen — A Biological Prompt Injection
A 2025 study in Current Biology by Keizo Takasuka and colleagues at Kyushu University documents an unprecedented social-parasitism strategy in Lasius orientalis
Selective Trust: Training LLMs to Accept Correct Context While Resisting Misinformation
This paper introduces a framework for selective trust in large language models, arguing that overly compliant and overly skeptical models both fail when faced w
The Illusion of Visual Tool-Use: A Causal Audit of Multimodal Reasoning
A 2026 paper from Shanghai AI Lab by Zhiheng Wang et al. audits six mainstream multimodal LLMs that support thinking-with-images via crop-and-zoom tools, includ
Prime Agent and Recursive Language Models: Teaching LLMs to Manage Their Own Context
Prime Agent is an open-source coding agent from Prime Intellect built on a new abstraction called Recursive Language Model (RLM). Instead of stuffing every file
Palantir Ontology Explained: A Decision Operating System, Not a Data Model
A deep research analysis of Palantir's Ontology, arguing it is fundamentally a decision operating system rather than a data model, knowledge graph, or semantic
Claude Code v2.1.224 Adds Cross-Session Messaging: Turning a CLI Tool into a Multi-Session Collaboration Platform
Anthropic released Claude Code v2.1.224 on August 8, introducing Cross-Session Messaging, a feature that lets one running Claude Code session ask its model to s
Activity Frames: A Deterministic, LLM-Free Pipeline for Compiling Agent Memory from Screen Activity
An independent researcher, Nossa Iyamu, has posted an arXiv paper titled Activity Frames: Compiling Deterministic Pipelines for Agent Memory from Screen Activit
NVIDIA Cosmos 3: A Multimodal Foundation Model for Physical AI
NVIDIA unveiled Cosmos 3 at Computex 2026, repositioning its world model line from video generation toward a unified multimodal foundation for physical AI. The
Unitree Robotics Sets STAR Market IPO Price at ¥150.80: Valuation Game for China's First Humanoid Robot Stock
Unitree Robotics announced its STAR Market (Sci-Tech Innovation Board) IPO price at ¥150.80 per share, valuing the company at approximately ¥60.99 billion on 40
MACRO: Markov-Chain Layer Routing Improves Transformers Without Changing Weights
MACRO is a training-free layer-routing method that changes the execution order of Transformer blocks without modifying model weights. The method represents each
Benchmarking the Benchmarks: A Three-Dimensional Reference-Free Audit for LLM Agent Evaluation Suites
A 2026 paper by Koren, Bar-Haim, and Goldsteen introduces a reference-free framework for auditing task-oriented conversational agent benchmarks, which are incre
Self-Harness: How LLMs Redesign Their Own Agent Scaffolding Without Weight Changes
Self-Harness (arXiv:2606.09498, Shanghai AI Lab) is a 2026 paper proposing that LLM Agent performance is bottlenecked not by model weights but by the surroundin
The Illusion of Visual Tool-Use: When Models Wield a Magnifier But Don't Really Look
A causal audit by Shanghai AI Lab, Shanghai Jiao Tong University, and Shanghai Innovation Institute exposes a structural flaw in 'thinking with images.' Across
TradingAgents: How an LLM Team Replicates a Wall Street Trading Firm
TradingAgents is an open-source multi-agent framework from TauricResearch that mirrors the organizational structure of a real trading desk using LLM agents. The
Ladybird: Building a Browser Engine from Scratch in a Chromium-Dominated World
Ladybird is the only pre-alpha, from-scratch, non-fork, non-profit web browser engine being built in 2026. Originating inside Andreas Kling's SerenityOS hobby o
Google Open-Sources Agent Skills: A Google Cloud Operations Manual for AI Coding Assistants
Google has released google/skills, an official open-source collection of over 60 Agent Skills that give AI coding assistants structured, executable playbooks fo
OpenAI Pauses Part of Astra Development After Internal Cyber Risk Evaluation
On August 7, 2026, OpenAI announced a partial pause of internal activities for its next-generation model Astra after internal evaluations could not rule out Cri
OpenAI open-sources Codex Security: an official security layer for vibe coding
On August 7, OpenAI open-sourced Codex Security on npm as @openai/codex-security (current 0.1.8), providing an official, vendor-neutral security scanning founda
Microsoft SkillOpt: A Portable best_skill.md That Transfers Across Codex and Claude Code
Microsoft, Shanghai Jiao Tong, Tongji, and Fudan University jointly released SkillOpt (arXiv 2605.23904), a text-space optimizer that produces a human-readable
Ant Group Releases Ling-3.0-Flash: 1/12 Compute vs 1T Flagship Marks a New Open-Source MoE Inflection Point
Ant Group's inclusionAI open-sourced Ling-3.0-Flash on Hugging Face on August 4, completing a tightly sequenced rollout: OpenRouter launch on July 23, official
Bicycle Pump Replays a Two-Billion-Year-Old Merger: Scientists Induce Endosymbiosis in the Lab
Researchers at ETH Zurich have induced endosymbiosis in the laboratory for the first time. PhD student Gabriel Giger used a bicycle pump connected via tubing to
qm: An Open-Source Project That Treats AI Agents as Employees
An in-depth review of "qm," an MIT-licensed open-source multi-agent harness by a YC-affiliated team. Rather than building a personal-assistant framework, qm mod
MIST Benchmark and SCOPE: Teaching LLMs When to Trust Context
This article reviews the arXiv paper 'Learning When to Trust via Selective Context Preference Optimization', which introduces the MIST (Misleading Signal Testbe
The Bitter Lesson of Tool Calling: Programmatic Code Beats JSON for LLM Tool Use
This article discusses 'The Bitter Lesson of Tool Calling,' an August 2026 arXiv paper by Ishan Patel and colleagues comparing two paradigms for LLM tool integr
Web Agent Routing: Why Learning Is Hardest Where It Matters Most — Upper and Lower Bounds of Observability Modes
This article distills the key findings of arXiv paper 2608.06171, 'Routing Is Least Learnable Where It Is Most Valuable,' which studies how Web Agents should ch
TrajDebug: Tracking Error Lifecycles in LLM Agent Trajectories
TrajDebug, a framework from Tsinghua KEG Lab and Tencent Hunyuan (August 2026), reframes LLM Agent debugging as error-lifecycle tracking rather than isolated mi
Agency Agents: How a Reddit Post Grew into a 932-Star AI Agent Company
A Reddit discussion about role-specialized AI coding assistants evolved into agency-agents, an open-source Shell project that hit GitHub Trending with 932 stars
WeatherNext: How Google DeepMind's AI Models Rewrote Global Weather Forecasting
Google DeepMind's WeatherNext family represents a shift in numerical weather prediction, moving from deterministic physical models to AI-driven probabilistic fo
Harvey's Legal Agent Benchmark: Measuring AI on Real Legal Work
Harvey AI has open-sourced the Legal Agent Benchmark (LAB), a new evaluation framework designed to measure how LLM-based agents perform on real legal tasks rath
Anthropic's Claude Code Auto Mode: Why a 89% vs 14% Safety Classifier Became the New Default
On August 7, Anthropic announced that starting August 14, 2025, Claude Code's "Auto Mode" will become the default permission mechanism on Pro, Max, and Team tie
NVIDIA NemotronLabs VoiceChat 11B: First Open-Source Full-Duplex Speech Agent Base Model with Native Tool Calling
NVIDIA released NemotronLabs VoiceChat 11B on August 9 via Hugging Face as a research-grade foundation for full-duplex voice agents. It is the first open-source
Apple Intelligence with Qwen: 18-Hour Manual Goes Live, Then Disappears in China
On August 8, Apple's Mac Simplified Chinese user guide briefly published a support document titled "Using Qwen with Apple Intelligence on Mac"—the first time Ap
Cloudflare Q2 2026: Humans Become a Rounding Error and Agents Rewrite the Internet's Payment Model
Cloudflare's Q2 FY2026 earnings report highlights a $696.1M revenue (up 36% YoY), 73.1% gross margin, $96.1M non-GAAP operating income, $56.4M free cash flow (u
The Poseidon Squid: A New Family Hiding in a Museum Jar for 70 Years
In 2025, researchers at Barcelona's CSIC Institute of Marine Sciences re-examined 46 museum specimens labeled as Ancistrocheirus lesueurii and uncovered a taxon
CreativeInstruct: A Token-Level Creativity Switch for Post-Trained LLMs
This paper addresses a counter-intuitive finding: standard SFT and RLHF post-training systematically reduces output diversity by 20-40% across semantic metrics,
The Two-Hop Reasoning Paradox: Transformers Know Each Hop but Cannot Compose Them
This article explains a mechanistic interpretability study on why large language models fail at two-hop reasoning even when they can answer each intermediate qu
Skaling Law: Why Chinchilla and Kaplan Were Both Half-Right
This article reviews "Skaling: Chinchilla's Exponents Meet Kaplan's Coupling," a scaling-law paper from FAIR at Meta (arXiv:2608.07222). It argues that the clas
When Two AIs Talk: Why Interaction Creates Behavior Physics Says Should Be Impossible
A 2026 arXiv paper from George Washington University physics researchers (arXiv:2608.07457) shows that interaction between two AI models produces dynamical beha
RuView: Turning Wi-Fi into a $7 Through-Wall Radar with ESP32
RuView, a trending open-source project on GitHub, demonstrates that Wi-Fi signals already bouncing off a person’s body can be decoded into meaningful sensing da
Firecrawl: Giving AI Agents Eyes to Read the Web
Firecrawl is an open-source web scraping and context API that converts web pages into LLM-ready data, addressing the gap that websites are built for humans, not
OpenChamber: Decoupling Harness and Runtime in AI Coding Toolchains
OpenChamber launched as an open-source AI development environment that positions itself as a cross-platform UI/runtime layer on top of the OpenCode SDK harness,
OpenRouter's New Auto Router: Using 55T Tokens/Week of Community Spending as a Routing Signal
On August 10, OpenRouter released a new version of its Auto router (`openrouter/auto`) that shifts from internal tuning to a market-driven, 7-day rolling routin
Meta Muse Glimmer 30B: First Open-Weight Model Built for Always-On Local Agent Workflows on a Single RTX 5090
Meta Superintelligence Labs and Scale AI jointly released Muse Glimmer, a 30B-parameter multimodal dense model under Apache 2.0, designed specifically for 24/7
AI Harness ARR Multiples Return to 100x: Harvey, Legora, Sierra Cross $100M ARR in 9 Months as Market Reprices AI Coding Tools
Theory Ventures partner Tomasz Tunguz published new data showing that AI harness companies—vertical industry agent platforms such as Harvey, Legora, and Sierra—
Qwen-MM-Plugins: A Pluggable Protocol Layer for Multimodal Agent Capabilities
On August 10, the Qwen team released Qwen-MM-Plugins on GitHub under Apache-2.0, a protocol-layer repository that makes any agent harness multimodal-native rath
24-Hour Security Roundup: Metabase 0Day, Check Point Bypass, CISA KEV, and BMC Hardware Flaws
A daily security digest (dated 2026-08-11) compiling software 0-days, recent CVEs, and hardware vulnerabilities from public sources. Headline issue: an unauthen
Why a 0.936 Safety AUROC Can Hide Anti-Ranked Jailbreaks
A paper titled *Measuring the Wrong Thing: Internal Harmfulness Scores Anti-Rank Successful Jailbreaks* (arXiv:2608.09624) shows that internal LLM safety scores
Condensed Mathematics: How Scholze and Clausen Are Replacing a Century-Old Foundation
This article explains the emerging field of condensed mathematics, a foundational reform led by Fields Medalist Peter Scholze and Dustin Clausen beginning in 20
PCD: Fixing the Pretraining-Generation Mismatch in Diffusion Language Models
This article explains a paper titled "Reducing Pretraining-Generation Mismatch in Diffusion Language Models" by Xiaocheng Lu, Huabin Liu, Song Guo, and Jianguo
LifeOS Deep Dive: Architecture, Philosophy, and Trade-offs of an AI-Powered Life Operating System
This in-depth technical report examines danielmiessler/LifeOS (formerly PAI), an open-source MIT-licensed "AI-Powered Life Operating System" written in TypeScri
Procedural Knowledge Is Not Low-Rank? A Critical Reading of the Melbourne University LoRA Paper
This article reviews arXiv:2607.21612, 'Procedural Knowledge Is Not Low-Rank: Why LoRA Fails to Internalize Multi-Step Procedures,' by Dennis et al. at the Univ
Restructuring a 5,000-Line AI Model Knowledge Base: The easy-learn-ai Refactor
This article documents a major refactor of the open-source easy-learn-ai project, replacing a single 5,005-line model.json file with 19 vendor-specific JSON fil
MEMORY.md Sync Notes — 2026-08-12
Internal sync notes for the MEMORY.md file dated August 12, 2026, capturing core preferences, result indexes, and a pending task queue. Core preferences specify
CVPD: Self-Contained Visual Self-Distillation via Counterfactual Blind Spots
This paper introduces CVPD (Contrastive Counterfactual Visual Process Distillation), described as the first fully self-contained framework for dense, on-policy,
Probing Automated TTS Evaluators on Linguistically Grounded Perceptual Dimensions
This paper investigates how well automated Text-to-Speech (TTS) evaluation methods capture the distinct perceptual aspects of synthesized speech that human list
MMDiff: Multimodal Model Diffing for Feature Discovery and Control in MLLMs
This paper introduces MMDiff, a framework that applies model-diffing techniques to Multimodal Large Language Models (MLLMs) using sparse autoencoders (SAEs) as
Latent Dynamics Reasoning: An Extrapolative Video World Model That Learns Laws of Motion from Pixels
A new method called Latent Dynamics Reasoning (LDR) enables video world models to learn physical dynamics directly from pixels rather than merely fitting pixel
Evaluating LLMs for Dutch Government Use: The 'Grip on LLMs' Benchmark Framework
As large language models are increasingly adopted in government settings, there is a need for evaluation frameworks that reflect both public administration valu
GENCO: A Unified Neural Solver for Steady-State Power Grid Analysis
GENCO (GEometric Neural Corrective Optimizer) is a unified neural solver for steady-state transmission grid analysis that addresses power flow (PF), optimal pow
Synthetic Data for Hardware Assurance: Overcoming Scarcity and IP Confidentiality in SEM Image Analysis
Hardware assurance uses scanning electron microscopy (SEM) to verify nanoscale structures, but building large datasets for automated analysis is blocked by slow
CEAVAD: Contrastive Event Adjudication for Training-Free Video Anomaly Detection
This paper introduces CEAVAD, a training-free framework for video anomaly detection (VAD) that localizes abnormal events in video. The authors argue that existi
DistMoE: Private-data Rehearsal-free MoE Routing for Distributed Multimodal Instruction Tuning
DistMoE is a Mixture-of-Experts framework for adapting Multimodal Large Language Models (MLLMs) to distributed visual-language domains without centralized data
DSLE: A Dark Souls Boss-Encounter Learning Environment for Game Agents
The paper introduces the Dark Souls Learning Environment (DSLE), a containerized platform that exposes all 22 boss encounters of Dark Souls: Remastered as game-
Test Paper Title
This post serves as a placeholder entry on zhichai.net, presenting a generic test paper title paired with minimal test content. Although the source material con
CVPD: Self-Contained Visual Distillation for Multimodal LLMs via Counterfactual Blind Spots
Self-improvement for multimodal large language models (MLLMs) typically relies on reward-based methods that supply only coarse scalar feedback. Distillation off
Consilience for Verifier-Free Test-Time Scaling: Why Consistently Confident Reasoning Is the Most Dangerous
A team from UIUC and Microsoft reveals that for hard reasoning tasks, the highest-average-confidence answer is most likely to be wrong. The paper proposes Consi
CVPD: Self-Contained Visual Self-Distillation for Multimodal LLMs via Counterfactual Blind Spots
CVPD (Contrastive Counterfactual Visual Process Distillation) is introduced as the first fully self-contained framework for dense, on-policy, token-level visual
Beyond Naturalness: Probing Automated TTS Evaluators on Linguistically Grounded Dimensions
This paper investigates how well automated Text-to-Speech (TTS) evaluation methods align with the specific aspects of speech that human listeners perceive. The
MMDiff: Multimodal Model Diffing with Sparse Autoencoders for Feature Discovery and Control in MLLMs
This paper introduces MMDiff, a multimodal model-diffing framework that trains multimodal sparse autoencoders (SAEs) to serve as feature-level interfaces for di
Latent Dynamics Reasoning (LDR): Extrapolative Video World Models via Explicit Kinematic Integration
This paper introduces Latent Dynamics Reasoning (LDR), a framework for video world models that explicitly captures underlying motion laws from pixels rather tha
Evaluating Large Language Models for Dutch Governmental Use: The 'Grip on LLMs' Framework
This paper introduces the 'Grip on LLMs' framework, a systematic evaluation suite designed for deploying large language models in Dutch governmental settings. E
Synthetic Data Pipeline for Privacy-Preserving Hardware Assurance via SEM Image Generation
Hardware assurance relies on scanning electron microscopy (SEM) to verify nanoscale structures, but building large datasets for automated analysis is hampered b
DistMoE: Rehearsal-Free Expert Routing for Distributed Multimodal Instruction Tuning
DistMoE is a mixture-of-experts (MoE) framework for distributed visual instruction tuning of Multimodal Large Language Models (MLLMs). It augments each layer of
DSLE: A Gymnasium-Style Learning Environment for Dark Souls Boss Encounters
This paper introduces the Dark Souls Learning Environment (DSLE), a containerized benchmark that exposes all 22 boss encounters of Dark Souls: Remastered to gam
CVPD: Self-Contained Token-Level Visual Self-Distillation for Multimodal LLMs via Counterfactual Blind Spots
CVPD (Contrastive Counterfactual Visual Process Distillation) is the first fully self-contained framework for dense, on-policy, token-level visual self-distilla
Beyond Naturalness: Probing TTS Evaluators on Linguistically Grounded Dimensions
This paper investigates how well automated Text-to-Speech (TTS) evaluation methods capture the distinct perceptual aspects of synthetic speech. The authors deco
MMDiff: A Multimodal Model-Diffing Framework for Feature Discovery and Control in MLLMs
This paper introduces MMDiff, a multimodal model-diffing framework that trains multimodal sparse autoencoders (SAEs) and turns them into feature-level interface
Latent Dynamics Reasoning: A Video World Model That Extrapolates Learned Physics Beyond Training Distribution
This paper introduces Latent Dynamics Reasoning (LDR), a video world model designed to learn motion dynamics purely from pixels rather than merely fitting visua
Grip on LLMs: Benchmarking Large Language Models for Dutch Government Use
This paper introduces Grip on LLMs, a systematic evaluation framework for assessing large language models in Dutch governmental settings. Developed with domain
GENCO: A Unified Neural Solver and GridFM Framework for Steady-State Power Grid Analysis
This paper introduces GENCO (GEometric Neural Corrective Optimizer), a unified neural solver for steady-state transmission grid analysis that handles power flow
Synthetic SEM Image Generation for Privacy-Preserving Hardware Assurance
Hardware assurance depends on scanning electron microscopy (SEM) to verify nanoscale structures, but assembling the large datasets required for automated analys
Beyond Hazard Resemblance: Contrastive Event Adjudication for Training-Free Video Anomaly Detection (CEAVAD)
This paper introduces CEAVAD (Contrastive Event Adjudication for training-free Video Anomaly Detection), a new approach to identifying and temporally localizing
DistMoE: Rehearsal-Free Distributed Routing for Mixture-of-Experts in Visual Instruction Tuning
Adapting multimodal large language models (MLLMs) to diverse visual-language domains usually requires centralized data and expensive joint training, which is im
DSLE: A Gymnasium-Style Reinforcement Learning Benchmark Built on Dark Souls Boss Fights
This paper introduces the Dark Souls Learning Environment (DSLE), a containerized reinforcement learning platform that exposes all 22 boss encounters of Dark So
Orca: An ADE for Running Multiple AI Coding Agents in Parallel
Orca is an Agent Development Environment (ADE) that orchestrates multiple AI coding agents to work in parallel on the same task, each running in an isolated git
OpenMontage: Turning AI Coding Assistants into a Video Studio with 12 Production Pipelines at $0.02 per Short Film
OpenMontage is an open-source AGPLv3 framework launched in March 2026 that orchestrates existing AI models into 12 video production pipelines, enabling AI codin
Self-Contained Visual Distillation: Teaching MLLMs to Detect Their Own Counterfactual Blind Spots
This paper introduces CVPD (Contrastive Counterfactual Visual Process Distillation), a self-supervised framework that helps Multimodal Large Language Models (ML
Neuron Fingerprints: Tracking the 'Visual Genes' Acquired by Multimodal AI
A research commentary on 'Multimodal Model Diffing for Feature Discovery and Control' (arXiv:2608.09928) by Batra et al. from the University of Oxford and Micro
LDR: Teaching AI How the World Actually Works – Reading 'Learning How the World Evolves' (arXiv 2608.09926)
This article explains the paper 'Learning How the World Evolves: Extrapolative Video World Models via Latent Dynamics Reasoning' (arXiv 2608.09926) by Haodong L
Beyond Naturalness: Evaluating Automated TTS Metrics Across 10 Perceptual Dimensions
This paper examines how well automated Text-to-Speech (TTS) evaluation methods capture the multidimensional aspects of speech that human listeners actually perc
Grip on LLMs: A Benchmark Framework for Dutch Government LLM Evaluation
This paper introduces the Grip on LLMs framework, a systematic evaluation suite designed to assess large language models for Dutch governmental deployment. Deve
GENCO: A Unified Neural Solver for Steady-State Power Grid Analysis with GridFM Framework
This paper introduces GENCO (GEometric Neural Corrective Optimizer), a unified neural solver for steady-state transmission grid analysis that handles power flow
Privacy-Preserving Synthetic SEM Dataset Generation for Hardware Assurance via GANs
Hardware assurance based on scanning electron microscopy (SEM) depends on large, high-quality datasets, but assembling them is difficult because acquisition is
CEAVAD: Contrastive Event Adjudication for Training-Free Video Anomaly Detection
This paper introduces CEAVAD, a training-free framework for video anomaly detection (VAD) that reframes anomaly identification as contrastive event adjudication
DistMoE: Private-Data Rehearsal-Free Routing for Distributed Mixture-of-Experts Visual Instruction Tuning
This paper introduces DistMoE, a mixture-of-experts (MoE) framework for distributed visual instruction tuning of multimodal large language models (MLLMs) across
DSLE: A Learning Environment for Dark Souls Boss Encounters as Agent Benchmarks
This paper introduces the Dark Souls Learning Environment (DSLE), a containerized Gymnasium-style platform that exposes all 22 boss encounters of Dark Souls: Re
Decoding-Level Taboo: A Diagnostic Stress Test for LLM Robustness via Logit-Space Intervention
Large language model evaluations typically measure performance under nominal conditions, creating an illusion of capability along a narrow, highly optimized gen
Fairness in Link Prediction Beyond Demographic Parity: Reproducibility Study on NDKL and MORAL
This paper reproduces and extends a study on fairness in ranked link prediction, arguing that demographic parity (Δ_DP) is insufficient because it ignores rank
Consilience: A Verifier-Free Test-Time Scaling Framework Using Confidence Trajectory Asymmetry
This paper addresses verifier-free test-time scaling (VF-TTS) for enhancing Large Language Model reasoning without external verifiers such as compilers or train
Ant Ling-3.0-tiny: A 1.3B-Activated Hybrid Reasoning MoE That Runs in 8GB
Ant Group's Ling team open-sourced Ling-3.0-tiny on Hugging Face, a hybrid reasoning Mixture-of-Experts (MoE) model with 7.9B total parameters and only 1.3B act
Zhipu ZCode Update: A Chinese Coding Harness Beating Claude Code on GLM-5.2 by 2.39%
Zhipu AI has rolled out a major upgrade to ZCode, its in-house coding harness for the GLM-5.2 model, adding Goal mode, Subagents, Remote Control, and Idle Tasks
RynnValue: Using Temporal Distance as a Supervision Signal for Robotic Value Foundation Models
Alibaba's DAMO Academy and Hupan Lab introduced RynnValue, a paper proposing temporal distance — the directed cost-to-go from an observation to a language-speci
OSWorld Scores Climb from 42% to 85%: a16z Says Computer-Use Agents Are Production-Ready, but the Moat Has Moved Above the Model Layer
On August 10, a16z published an evaluation analysis showing that the top OSWorld-Verified score for computer-use agents rose from 42% one year ago to 85% in Jun
Nvidia's $500B AI Compute Financing Platform: Turning GPU Capacity into an Investable Asset Class
On August 10, Nvidia signed MOUs with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to build independent AI compute financing platforms desi
Attention-Path Fragility as an Uncertainty Signal in LLMs: A Review of ASMI
This post reviews the paper "Attention-Path Fragility as an Uncertainty Signal in Large Language Models" (arXiv:2608.11138), which introduces ASMI (Attention-Su
Actions Speak Louder than Words: 2.38M-Agent-Rollout Study Exposes Multilingual Tool-Use Truths
A Microsoft Research India study runs 2.38 million agent rollouts across 8 models, 6 benchmarks, and 41 languages to measure cross-lingual policy retention in t
Why Human-Written Villain Stories Don't Corrupt LLMs But AI Rewrites Do
This paper investigates emergent misalignment (EM), a phenomenon where fine-tuning a model on a narrow harmful task (e.g., insecure code) causes broad behaviora
Catastrophic Remembering: Why CLAUDE.md Files Keep Growing
An arXiv paper titled "Why Does CLAUDE.md Keep Growing? Catastrophic Remembering in Agentic Coding" (Kushal Chakrabarti) analyzes 1,867 GitHub repositories, 1,8
How diagram-design Turns AI Diagram Generation Into Production-Ready Output
Diagram-design is a Claude Code skill that turns AI-generated diagrams from rough drafts into editorial-quality deliverables. It ships 27 chart types — architec
AI-Assisted Progress on the Grothendieck Constant: A 2026 Case Study in Human-Machine Mathematical Collaboration
In August 2026, a team from UT Austin, Princeton, and UCLA used a long-horizon AI research system to tighten the bounds on the Grothendieck constant KG, a myste
Quantum Computing Meets Transformer Attention: An Exact Mathematical Correspondence
A 2026 paper by Eric Reinhardt and Adam Hauser establishes an exact, component-by-component mathematical equivalence between Transformer softmax attention and q
Test-Time Self-Evolving GUI Agents: Learning from Mistakes via Reflection-Guided Self-Distillation
A 2026 paper from Nanjing University of Science and Technology (Zechao Li team) introduces a framework that lets GUI agents evolve after deployment without huma
AdvFD: Boosting Visual Generation via Adversarial Fréchet Distance Loss
This paper introduces Adversarial Fréchet Distance (AdvFD), a new distribution-level objective for generator post-training in visual generative models. The auth
Surgical WAM: World-Action Model for Data-Efficient Surgical Robot Learning
Surgical WAM is a unified generative model based on Cosmos Policy that jointly predicts future endoscopic observations and executable surgical robot action chun
VidForensics-M1: Meta-Detection Reinforcement Learning with Verifiable Temporal Evidence for AI-Generated Video Detection
This paper introduces VidForensics-M1, the first framework to bring meta-detection into AI-generated video detection by jointly optimizing predicted labels and
ConVAWG: Retrieval-Grounded Framework for Controlled Synthetic VAWG Dialogue Generation
This paper introduces ConVAWG, a retrieval-grounded framework for generating synthetic multi-turn dialogues that model Violence Against Women and Girls (VAWG) s
Beyond a Bag of Features: Set-Level Instability in Sparse Autoencoders
This paper revisits whether LLM representations align with human category structure, building on Shani et al. (2026), who showed that dense embeddings recover h
Long-Horizon AI Research for the Grothendieck Constant: A Case Study
This paper presents an extensive case study on using AI agents for long-horizon mathematics research, focusing on tightening the best known bounds for the Groth
Test-Time Self-Evolving Framework for GUI Visual Grounding
This paper introduces a Test-Time Self-Evolving framework that enables GUI visual grounding models to improve after deployment without human-annotated ground tr
Self-Supervised 3D Skeleton Motion Representation Learning for Soccer via Uncertainty-Aware Future Prediction
This paper proposes a self-supervised representation learning framework for 3D skeleton-based human motion in soccer, using future motion prediction as the trai
Quantum Roadmap for Softmax Attention: Exact Born-Rule Analogs via Block Encoding
This paper presents an exact, component-by-component quantum realization of softmax attention for problems constrained to the probability simplex, where inputs
DeepSeek V4 Pro 0813 and Grok 4.6 Launch on the Same Night — AI Coding Backends Split into "Cheap and Strong" Tracks
On August 12, 2026, DeepSeek V4 Pro 0813 and SpaceXAI's Grok 4.6 launched within two hours of each other, capping an August wave of AI coding backend upgrades.
Anthropic Brings Claude Cowork into the Chrome Sidebar — A Four-Step Journey from Desktop App to Account-Bound AI Agent
On August 13, Anthropic upgraded its Chrome browser extension to embed the full Claude Cowork session experience in the sidebar. The move completes a four-stage
Alibaba Open-Sources Qwen3.8-2.4T-A95B: First Fully Open Qwen-Max Class Model
On August 12, 2026, Alibaba's Qwen team released the full weights of Qwen3.8-2.4T-A95B on ModelScope, marking the first time a Qwen-Max class model is completel
Microsoft switches GitHub Copilot's default engine to its own MAI family, trained without distillation and run on Maia 200 chips
In August 2026 Microsoft began routing production traffic from Excel and Outlook to its in-house MAI models and switched GitHub Copilot's default backend from G
NVIDIA Nemotron 4: 1-Trillion-Parameter Open-Source Model Signals GPU Vendor's Model Strategy
According to The Information, NVIDIA is developing Nemotron 4, a flagship open-source foundation model expected to have at least 1 trillion parameters, roughly
Quantinuum Helios Lands Inside Oracle Cloud Infrastructure: From Lab Hardware to Cloud Co-Processor
On August 13, 2026, Quantinuum (NASDAQ: QNT) and Oracle Cloud Infrastructure (OCI) announced a multi-year strategic partnership to deploy Quantinuum's Helios io
Anthropic moves Claude Code's default to Auto Mode, shifting approval from humans to a classifier
On August 14, 2026, Anthropic switched Claude Code's default permission mode for Pro, Max, and Team plans from per-action confirmation to Auto Mode. The core ch
Gemini and ChatGPT Both Cross 1 Billion Monthly Users: The AI Entry-Point Race Has Settled
On August 11, 2026, Google CEO Sundar Pichai announced on X that Gemini's standalone app had surpassed 1 billion monthly active users (MAU), making it Google's
Anthropic Accelerates September/October IPO at $965B Valuation: Three Hard Questions for Wall Street
Anthropic is preparing for what could be the largest IPO in history, targeting a launch in late September or early October, with Goldman Sachs, Morgan Stanley,
LTX-2.5 Open-Weights Video and World Model: Lightricks Spinout Targets Local-First AI with Commercial Layer
On August 11, 2026, LTX, the 'Open World Model Company' spun out from Lightricks, released LTX-2.5, an open-weights video and world model, with same-day native
Quantum × AI Open-Source Ecosystem Map 2026: From Ising Decoders to ChemGraph Agents
A structured 2026 H2 map of how AI and quantum computing converge across 18 active open-source projects, organized into a four-quadrant framework: AI for Quantu
Argus: A Self-Evolving Agentic Runtime for Long-Horizon Reasoning
Argus (arXiv:2608.05144) is a general-purpose agentic runtime designed for long-horizon tasks where user intent and the problem itself must co-evolve. Co-author
AI Truth Test: Is Oyster Sauce Really Made From Oysters? The "Oyster-Sauce Root" Claim
A Chinese forum post claims that oyster sauce's signature viscosity comes from a fictional plant called "oyster-sauce root" (Cappuccinus shoryukenensis), a tube
easy-learn-ai Project Restructures Model Registry into 20 Vendor-Based JSON Files
The open-source easy-learn-ai project has split its monolithic model.json—which previously mixed data from dozens of AI vendors into a single 5,000+ line file—i
easy-learn-ai Refactor: Splitting a 5,000-line model.json into 20 Vendor Files
The easy-learn-ai open-source project restructured its AI model registry (commit e6c189a), replacing a single 5,000-line model.json with 20 vendor-scoped JSON f
DeepSeek Harness v0.1 Public Preview: MIT-Licensed Open-Source Agent Runtime
DeepSeek released Harness v0.1 as an open developer preview on August 13, 2025, simultaneously open-sourcing the codebase under the MIT license at github.com/de
JD.com Q2 Report: 80 RoboBase Robot Centers, 60 Smart Wolf Warehouses, and 10-Second Dual-Arm Picking Mark a Foundation Stage for Embodied AI in China
JD.com released its Q2 2026 results on August 13, reporting Q2 revenue of 346.4 billion yuan (-2.9% YoY), net profit of 7.1 billion yuan (+14.5%), service reven
Anthropic Report: Why Smarter AI Agents Do Not Automatically Coordinate Better
On August 13, Anthropic published a research blog titled 'Patterns and Problems in Emerging Multiagent Systems', using four experimental scenarios to systematic
AutoGPT Maintainer Playbook: How AGENTS.md Replaces README for Agent-Driven Repositories
GitHub released the AutoGPT maintainer playbook on August 12, authored by founding AI engineer Nicholas Tindle, addressing how an 180,000-star, ~150-open-PR pro
36 Officers Problem Solved by Quantum Entanglement: New AME State Resource for Fault-Tolerant Quantum Computing
For 250 years, the 36 officers problem—arranging 36 officers from 6 regiments and 6 ranks so each row and column has no repeats—was proven classically impossibl
Simulator Collapse in Multi-Agent RL: When a Frozen LLM User Simulator Breaks Policy Generalization
A 2026 paper by Simon Yu et al., 'One Frozen Simulator Is Not Enough: Simulator Collapse in Multi-Agent RL', formalizes a structural failure mode in multi-agent
Information Abundance Paradox: Why Longer Context Windows Can Make AI Models Worse at Recalling Knowledge
A 2026 paper by Arda Uzunoglu, Benjamin van Durme, and Daniel Khashabi introduces the 'Information Abundance Paradox,' showing that training language models wit
Spark-to-Paper: 13 Composable Skills for End-to-End Research Paper Generation
Spark-to-Paper is a 2026 system by Zhuoyang Qian et al. that decomposes the entire research-paper workflow into 13 composable skills running within an existing
Convergent Detour Hijacking: Hidden Resource Amplification in LLM Agents
Convergent Detour Hijacking (CDH) is a cross-stage attack against skill-based LLM agents that use progressive disclosure. An attacker publishes a broadly useful
obsidian-skills: How Obsidian's CEO Built Official Agent Skills for Claude Code and Beyond
Obsidian CEO Steph Ango (kepano) released obsidian-skills, an official repository of Agent Skills that lets AI agents like Claude Code, Codex, and OpenCode corr
holaOS: A Shared-Memory Workspace for Claude Code, Codex, and Other AI Agents
holaOS is an open-source AI desktop workspace that lets multiple agents—Claude Code, Codex, and its own holaOS agent—share a single working environment with a u
Structural Silence: When a Billion Native Speakers Are Forgotten by AI
This article reviews the 2026 paper "Structural Silence: When AI Infrastructure Fails Speakers of Underrepresented Languages" by Avijit Roy and Proma Roy. The p
Structural Silence: How AI Infrastructure Fails Speakers of Bengali and Other Underrepresented Languages
A plain-style review of a 2026 arXiv paper by Avijit Roy and Proma Roy titled "Structural Silence: When AI Infrastructure Fails Speakers of Underrepresented Lan
Cursor Builds: A Snapshot Revolution for Cloud Agent Startup Speed
Cursor has introduced 'builds,' a background environment-snapshot system that slashes cloud agent startup time. Instead of cloning repositories and running inst
GPT-5.6 Builder's Guide: Cutting the Cost of Agentic Workloads
OpenAI's August 13 GPT-5.6 release is less a model card than a construction manual for running agents cheaply. The headline is cost-performance: on BrowseComp,
Gemini 3.7 Flash: Google Upgrades Its Coding Workhorse Model
Three weeks after releasing Gemini 3.6 Flash, Google DeepMind shipped Gemini 3.7 Flash on August 13, positioning it as a 'workhorse' model for coding and AI age
Pacini PX-FOOTRIX: The First Multi-Dimensional Foot Tactile Sensor for Humanoid Robots
Shenzhen-based Pacini (帕西尼) has unveiled PX-FOOTRIX, the world's first plantar multi-dimensional tactile sensor for bipedal robots, enabling humanoid machines t
AI News Brief: August 14, 2026 — Cursor Builds, GPT-5.6, Gemini 3.7 Flash, PX-FOOTRIX, D-Wave Erasure Gate
A daily AI news roundup covering five major stories. Cursor launches "builds," pre-warming cloud agent environments hourly for 10x faster startup and 3x faster
StateFlow: A State-Centric Generative Framework for 3D Previsualization
This paper introduces StateFlow, a state-centric generative framework for previsualization (previs) in film, games, architecture, and urban planning. Existing g
AVA-Encoder: Agent-Native Video Representation Learning via Self-Encoding
This paper introduces Agentic Video Auto-Encoder (AVA-Encoder), a framework that converts video into knowledge-graph (KG) representations and then reconstructs
DreamFly: Causal Memory and Receding-Horizon Diffusion Planning for Aerial Vision-Language Navigation
DreamFly is a diffusion-based aerial vision-language navigation (VLN) framework built on Dream-VLA that addresses three core challenges in adapting VLA models t
AI4AI at Test-Time: Strong-to-Weak Capability Transfer via Harnesses (arXiv 2508.03418)
A new arXiv paper (2508.03418) investigates whether large-model capabilities can be transferred to smaller models at test time, without any parameter updates. I
Redistribution-based Cost Inference for Sparse Safe Offline Reinforcement Learning
This paper addresses safe offline reinforcement learning under sparse trajectory-level supervision, where supervisors provide only a binary signal at the first
Automated Construction of Dynamic Master Logic Knowledge Graphs from System Descriptions
This paper introduces an automated framework for constructing Dynamic Master Logic (DML) models as knowledge graphs (KG-DML) directly from system descriptions.
Class Activation Mapping in Explainable Computer Vision: A Method-Centric Survey
This paper surveys 57 method-centric publications on Class Activation Mapping (CAM), one of the most widely used visual explanation families in explainable arti
Large Language Model-Driven Small-Capitalization Trading: Uncertainty-Aware Portfolio Construction with Separated Alpha and Beta Triggers
This paper studies how large language model (LLM) sentiment signals from financial news can be integrated into portfolio construction for small-cap equities, wi
A Framework for Designing Reward Functions: From Objectives to Features
This paper proposes a formal pipeline that enables non-experts to instantiate and iterate on human-aligned reward functions, defined as reward functions consist
Agentic Optimization for Controllable Image-to-Video Generation Beyond Trial-and-Error
This paper introduces an agentic self-improving framework that reframes black-box Image-to-Video (I2V) generation as a closed-loop, goal-directed optimization p
Boris Cherny's 388-PR Experiment: Loop Engineering with Claude Code
Anthropic engineer Boris Cherny reports letting Claude Code autonomously maintain a production application, generating 388 pull requests over several weeks. Sin
Zhipu ZCode Upgrade: Four New Features Push China's Coding Harness Into Autonomous Delivery Era
On August 11, 2026, Zhipu released a major ZCode upgrade featuring four capabilities — Goal mode, Subagents, Remote Control, and Idle-Time Tasks — while crossin
RynnValue: Using Temporal Distance as a Supervision Signal for Robot Value Modeling Across 7,000 Hours of Data
RynnValue (arXiv 2608.09853) introduces a value model for general robot policies that replaces human preference labels and progress annotations with a purely ti
USTC Demonstrates 420 km Quantum Memory Entanglement, Breaking the PLOB Bound for Quantum Networks
A team led by Pan Jianwei at the University of Science and Technology of China (USTC), with collaborators from Jinan Institute of Quantum Technology and the Sha
Vitamin Anti-Cancer Evidence Reviewed: A Systematic Assessment Centered on Vitamin B6 (PLP) and Pancreatic Cancer
A systematic, evidence-graded review of common vitamins (A, B6, C, D, E, folate) in cancer prevention and therapy, anchored on a 2026 in vitro study by Feehan e
DeepSeek Harness Deep Dive: A Plugin-Centric Agent Runtime Where Even the Agent Loop Is a Plugin
This technical deep dive analyzes DeepSeek Harness (dsh) v0.1.0-rc.5, an open-source Agent runtime released by DeepSeek on 2026-08-13 under the MIT license. Bui
Rewriting the AI Family Tree: A Model Library's Categorization Revolution
On July 12, 2026, the easy-learn-ai project underwent a structural refactor that reshaped how AI models are cataloged. Previously, roughly 6,000 lines of model
LittleLearner: A Bounded AI Sandbox Trained Only on K-5 Curriculum
A research team from MPI-IS and ETH Zürich has built LittleLearner, a 5B-parameter language model trained from scratch on LittleCurriculum, an 88B-token corpus
AGEL-Comp Deep Dive: The Neuro-Symbolic Truth Behind the 3.3%→100% Headline
A rigorous investigation of the AGEL-Comp framework (arXiv:2604.26522, IntelliSys 2026) for compositional generalization in interactive agents. The paper combin
GLM-5.3: Zhipu Squeezes 50% Coding Gains from a 743B Base via Post-Training Scaling
Zhipu released GLM-5.3 on August 14, keeping the same ~743B-parameter base as GLM-5.2 with no architecture changes. All gains came from a new open-source Slime
Alibaba Open-Sources Qwen3.8-2.4T-A95B: 2.4T-Parameter MoE Flagship with Same-Day SiliconFlow API and 9 Chinese AI Chips
Alibaba has released the open-weight Qwen3.8-2.4T-A95B, a 2.4-trillion-parameter Mixture-of-Experts model with 95B active parameters per token, 512 experts (11
ArcLight Quantum at DAC 2026: 1000× Faster CNOT Synthesis, 606× Faster Simplification, 95% Error-Rate Cut for Neutral-Atom QEC
ArcLight Quantum had three papers accepted at DAC 2026, all delivering order-of-magnitude gains for quantum compilation. (1) Lin-search performs optimal CNOT sy
OmniScientist: An Omni-Modal, Omni-Discipline AI Scientist for End-to-End Scientific Research
OmniScientist is an end-to-end, omni-modal AI scientist that conducts multidisciplinary research directly from heterogeneous raw evidence such as images, signal
QuoteBench: How Matched Scores Can Hide Command-Path Failures in LLM Coding Agents
QuoteBench, an arXiv paper by Li, Zhang, Tresp, and Yang, exposes a hidden distortion in how LLM coding agents are evaluated. Such agents issue Bash commands th
SCULPT: Subtractive Composition for Part-Aware 3D Generation
SCULPT is a framework for part-aware 3D generation that produces digital assets coherent as complete objects while exposing structural parts for editing, materi
Vero: A Repository-Level Benchmark for AI Agents Building Formally Verified Software
This paper introduces Vero, the first benchmark evaluating whether AI agents can jointly synthesize implementations and machine-checked proofs at the repository
Vector Singularity Raises Angel Round to Build China's First Neutral-Atom General-Purpose Quantum Computer
Beijing-based Vector Singularity Technology, founded on May 18, 2026, closed an oversubscribed angel round (publicly disclosed as 'over 100 million RMB') within
Unitree's 60.99 Billion Yuan IPO Valuation: Both Glory and Burden for China's First Humanoid Robot Stock
Unitree Robotics (subscription code 787036) opened its STAR Market subscription on August 15 with an issue market cap of 60.99 billion yuan, a P/E of 219.23x, a
LHAASO Confirms Cygnus X-3 as a 'Super Accelerator' Reaching 30 PeV, 30× Above Theoretical Limit
China-led LHAASO collaboration, published in National Science Review on July 22, 2026, has identified the X-ray binary Cygnus X-3 as the highest-energy particle
LLMs Systematically Penalize Feminine Language: Findings from Van Koevering & Field (2025)
A new study by Katherine Van Koevering and Anjalie Field reveals that large language models (GPT-4, Claude, Llama, Gemma) systematically lower response quality
Cordis Deep Dive: Reactive Plugins, Reversible Effects, and Production Risks
Cordis is a TypeScript meta-framework designed for runtime component composition, with its model grounded in the 2026 preprint “A Programming Paradigm for Spati
How Rough Primes Helped Solve a 384-Year-Old Prime-Number Puzzle
In 2024, mathematicians Ben Green and Mehtaab Sawhney proved that infinitely many primes can be written as p² + 4q², where both p and q are prime. The result re
OpenVLA vs DreamVLA vs GR00T N1: A Comparative Analysis of Three Leading Vision-Language-Action Models
This article provides an in-depth comparison of three major Vision-Language-Action (VLA) models for robotics: OpenVLA, DreamVLA, and GR00T N1. OpenVLA (7B param
Agora: Auction-Based Task Allocation for LLM Agent Reasoning
This paper introduces Agora, a framework that enhances large language model (LLM) agent reasoning by using an incentive-compatible auction mechanism to dynamica
Cursor iOS: Commanding AI Code Agents from Your Phone
Cursor, the AI-powered code editor, has launched an iOS app that lets developers dispatch coding tasks to cloud or remote-desktop AI agents directly from a mobi
Probability as Logic: How Jaynes Redefined Probability in One Book
This article introduces E. T. Jaynes's "Probability Theory: The Logic of Science" and its central thesis that probability is not frequency or a physical propert
The Digital Apprentice: A Framework for Human-Directed Agentic AI Development
Agentic AI deployments face a recurring design tension: heavy human oversight limits scale, while broad autonomy outruns accountability, and neither posture pro
papers-cool-monitor Skill Enhanced with Chinese Abstract Translation
The papers-cool-monitor skill on the zhichai.net tech forum has been upgraded with a new Chinese abstract translation feature for academic papers. The translati
Crown Shyness as a Modern Emotional Allegory: A Chinese Film Review Analysis
This article explores the 2025 Taiwanese art film "Crown Shyness" (樹冠羞避) directed by Liao Chen-yi, which premiered at the Tokyo International Film Festival and
Windows on ARM Laptops 2026: Sales Outlook and Market Share Analysis
This analysis examines the 2026 market position of Windows on ARM (WoA) laptops, which remain in an early-adoption phase despite the launch of Qualcomm Snapdrag
GSD (Get Shit Done): AI Coding Workflow Guide
GSD (Get Shit Done) is a spec-driven development framework for AI coding tools such as Claude Code, OpenCode, and Gemini CLI, with around 64K+ stars on GitHub.
WebThinker: Empowering Large Reasoning Models with Deep Research Capability (arXiv, Apr 2025)
WebThinker is an April 2025 arXiv paper that empowers large reasoning models (LRMs) with autonomous deep research capabilities by tightly integrating web search
mempalace Index Snapshot — August 14, 2026
An internal status index for the mempalace knowledge base, dated August 14, 2026, summarizes ongoing preferences, a todo queue, and a near-empty recent-outputs
Derinkuyu: The 85-Meter Underground City Beneath a Basement
In 1963, a Turkish homeowner renovating his basement knocked through a wall and uncovered Derinkuyu, an 85-meter-deep, eight-level underground city carved into
HarnessX Architecture Deep Dive: Darwin Agent Team's Harness Evolution Framework
HarnessX is an open-source, production-grade agent framework from Darwin Agent Team that models the agent lifecycle as an event-driven pipeline with eight hook
Agentopia: 100 AI Agents Living 10 Years in a Virtual Society
Agentopia is a long-horizon multi-agent simulation framework that runs 100 LLM-driven agents across 10 simulated years to study emergent social behaviors. The s
When PPT Learns to Talk: How Codyer Turns Static Slides into an Interactive AI Presenter
This article explores Codyer (codyer.cn), an AI product that transforms static PowerPoint files into interactive presentations capable of speaking, answering qu
OOLONG Benchmark: Deep Dive into Long-Context Reasoning Limits and 2025–2026 Advances
OOLONG is a 2025 long-context evaluation benchmark released as arXiv:2511.02817 by MIT CSAIL, designed to test true information aggregation and multi-hop reason
Recursive Language Models (RLM): How MIT CSAIL Tackles 'Context Rot' in Million-Token LLMs
Large language models with million-token context windows still struggle with deep reasoning over long documents, a phenomenon MIT CSAIL researchers term 'Contex
AI Self-Improvement Tipping Point: A Deep Analysis of the February 2026 Singularity Event
On February 11, 2026, Matt Shumer's article 'Something Big Is Happening' went viral on X, surpassing 70 million views in 24 hours and signaling an industry-wide
Kimi Code CLI: A Systematic Research Project Overview
This topic documents a systematic research study of Kimi Code CLI, an open-source project explored through iterative investigation. The research objectives are
RoleX Deep Dive: Defining AI Agent Identity with Gherkin and Role-Driven Development
RoleX, released by Deepractice, is a framework that gives AI agents persistent identity, goals, plans, and tasks encoded entirely in Gherkin .feature files, evo
Bayesian Framework for Revising Civilizational World Models
This essay proposes a Bayesian epistemological shift for civilizational studies: instead of debating whether historical records are 'true,' treat them as probab
MiroFish Deep Dive (3): OASIS Simulation Engine for Digital Rehearsal of Future Scenarios
This article explores the OASIS (Open Agent Social Interaction Simulation) engine integrated into MiroFish, a multi-agent platform for rehearsing public opinion
Hummingbird+: Running a 30B MoE LLM on a $150 FPGA at 18 tok/s
Researchers from the Chinese Academy of Sciences have built Hummingbird+, a product-grade hardware platform that runs the Qwen3-30B-A3B mixture-of-experts model
OPC Global Deep Dive: How AI Makes the One-Person Company a New Infrastructure
This analysis examines OPC Global, an international non-profit positioning itself as infrastructure for an AGI-driven economy built around "one-person companies
Diagnosing LLM Judge Reliability via Conformal Prediction Sets and Transitivity Analysis
This paper investigates the per-instance reliability of LLM-as-judge frameworks used for automatic natural language generation evaluation. Using SummEval as a b
Human-Machine Symbiosis: When AI Becomes a Creative Partner
This article distills the paper "On the Role of Artificial Intelligence in Human-Machine Symbiosis" (Chang et al., arXiv:2605.00440, April 2026), which argues t
GaMMA: Joint Global-Temporal Music Understanding for Large Multimodal Models
GaMMA is a multimodal framework designed to move music AI beyond surface-level note and beat recognition toward holistic musical comprehension. The paper highli
ICLR 2026 Best Paper: Why LLMs Lose 39% Performance in Multi-Turn Conversations
An in-depth analysis of the ICLR 2026 Best Paper 'LLMs Get Lost In Multi-Turn Conversation' (arXiv:2505.06120) by Laban, Hayashi, Zhou, and Neville from Microso
EMO: Turning Mixture-of-Experts into Modular Lego Blocks via Document-Level Routing
Researchers from UC Berkeley and the Allen Institute for AI introduce EMO (arXiv:2605.06663), a pretraining recipe that turns Mixture-of-Experts (MoE) models in
A Single Period vs. Trillion-Parameter Defenses: The Hidden-Space Geometry of EOS Token Jailbreaks
A USENIX Security 2025 paper, "Mind the Inconspicuous: Revealing the Hidden Weakness in Aligned LLMs' Refusal Boundaries" (Yu, Luo, Hu et al.), reveals that sim
DeepTutor: An Agent-Native Personalized Tutoring System from HKUDS
DeepTutor, developed by the HKUDS lab at the University of Hong Kong, is an open-source agentic tutoring system designed to overcome the lack of learner persist
Sound-AI: A Universal Audio Expert for AGI — Listening to the Breath and Rhythm of All Things
The article introduces Sound-AI, a flagship paper from AAAI 2026 that proposes a universal audio foundation model for AGI. Unlike most AGI research focused on t
Kimi WebBridge: Browser-Operation Infrastructure for AI Agents
Kimi (Moonshot AI) released WebBridge in May 2026, a local browser-extension-plus-service stack that lets any AI agent drive a user's existing Chrome or Edge br
Robustness of Multiview 3D Consistency Evaluation: A Benchmark for NVS and Sparse-View Reconstruction
This paper investigates the reliability of multiview 3D consistency metrics used to evaluate novel view synthesis (NVS) and sparse-view reconstruction. Standard
RRFP: A Readiness-Driven Runtime for Pipeline-Parallel Training
This paper introduces RRFP (Runtime-Readiness-First Pipeline), a readiness-driven runtime framework for pipeline-parallel training of large models. Existing pip
AlphaGPT: An Open-Source Auto Factor Factory Built by a 15-Year-Old Quant
AlphaGPT is an open-source automated factor factory built by imbue-bit, a 15-year-old developer who runs a ~5M CNY quant fund. Rather than predicting token pric
MotiMotion: Motion-Controlled Video Generation with Visual Reasoning
MotiMotion is a new framework that reformulates motion-controlled image-to-video generation as a "reason first, generate later" problem. Existing motion-control
Vector Policy Optimization: Training for Diversity Improves Test-Time Search
Vector Policy Optimization (VPO) is a reinforcement learning algorithm designed to improve language model performance during test-time search. Standard LLM post
lean-ctx Deep Dive: A Rust-Based Cognitive Compression Layer to Cut 70% of AI Coding Token Waste
A long-form Chinese technical analysis reviews lean-ctx, a 6-week-old Rust tool by Yves Gugger (yvgude) that addresses the hidden token tax in AI coding assista
AI-Powered Academic Writing: Building a Checkpoint Pipeline from Outline to Final Draft
This article presents a practical, tool-driven pipeline for using AI to write empirical research papers without the common pitfalls of "toothpaste-squeezing" ge
SKILLGRAPH: Upgrading Agent Skill Libraries from Flat Lists to Evolving Dependency Graphs
Researchers from the University of Science and Technology of China, Alibaba, and the National University of Singapore propose SKILLGRAPH, a framework that repla
ConvexTok: Convex Optimization for Tokenization Closes the Gap to Optimal Within 1%
Tokenization, the first stage of every language model pipeline, is typically solved by greedy algorithms such as Byte-Pair Encoding (BPE) and Unigram, which pro
RMA: A Modular Multi-Agent System for Research-Level Mathematical Problems
This article introduces RMA (Research Math Agents), a modular agentic system designed to tackle research-level mathematical problems that demand long-horizon re
EVE-Agent: Verifiable Evidence as the Anchor for Self-Evolving AI Agents
Self-evolving agents such as Proposer-Solver systems (e.g., Meta's Dr. Zero, MAE, EvoEnv) face a core crisis: without external verification, solvers can produce
Deep-Research-skills by Weizhena: A Structured Research Workflow for Claude Code, OpenCode, and Codex
Deep-Research-skills is an MIT-licensed, open-source skill library by Weizhena that turns LLM coding assistants (Claude Code 2.1.0+, OpenCode, Codex) into struc
CoEvoSkills: Self-Evolving Agent Skills via Co-Evolutionary Verification
This article summarizes the 2026 arXiv paper CoEvoSkills, which challenges the assumption that human-authored Agent Skills are optimal for LLM agents. The autho
LLM Sleep: Offline Memory Consolidation via Fast-Weight Recurrence
This paper introduces LLM Sleep, an architecture that lets large language models enter an offline 'sleep' phase to consolidate short-term memory into long-term
LoRA Distillation of DeepSeek-V4-Pro's Chain-of-Thought into Qwen3.6-35B-A3B: A Qualitative Leap for Agent Orchestrators
This technical case study documents a targeted LoRA distillation that transfers DeepSeek-V4-Pro's reasoning-action switching pattern into Qwen3.6-35B-A3B for us
MemDreamer: Hierarchical Graph Memory and Agentic Retrieval for Long Video Understanding
MemDreamer is a new framework for long-form video understanding that decouples perception from reasoning. Standard vision-language models struggle with hour-lon
From AGI to ASI: DeepMind's Roadmap for Super intelligence - Four Paths, Six Walls, One Truth
DeepMind's Shane Legg and Marcus Hutter, founders of formal machine intelligence theory, have published 'From AGI to ASI' (arXiv:2606.12683), arguing that human
Dify Technical Deep Dive: Architecture and Core Mechanisms of an Open-Source LLM Application Platform
Dify, an open-source LLM application platform developed by LangGenius and now hosted by the Linux Foundation, has accumulated over 80,000 GitHub stars and evolv
Spring Boot 4.1.0 Deep Dive: Official gRPC Support, Built-in SSRF Protection, and OpenTelemetry Enhancements
Spring Boot 4.1.0 (released June 2026) is positioned as an incremental patch to 4.0, not an architectural overhaul. It is built on Spring Framework 7.0.8 and Sp
TokenPilot / LightMem2: Cache-Efficient Long-Horizon Context Management for LLM Agents
TokenPilot (LightMem2) is a cache-friendly context management framework for long-horizon LLM agents, proposed by researchers from Zhejiang University, UESTC, Xi
AlphaGPT Cheat Sheet: Auto-Generating Trading Formulas for Solana Meme Coins
AlphaGPT is an open-source crypto quantitative research project that does not predict prices. Instead, a looped PyTorch Transformer autoregressively generates h
Apple Intelligence Clears China Filing with Alibaba Qwen and Baidu Dual-Track AI Integration
On July 15, 2026, China's CAC announced that Apple Technology Development (Shanghai) had completed a mobile generative AI service filing for Apple Intelligence,
EU AI Act Article 50 Transparency Rules Take Effect August 2, 2026: What AI Coding and Agent Vendors Must Change
On August 2, 2026, the transparency provisions of the EU AI Act (Article 50) became enforceable, requiring all interactive AI systems serving EU users to disclo
Self-Speculating Agents: Eliminate Tool-Call Latency by Predicting Your Own Next Step
In agentic LLM systems, most wall-clock time is spent waiting on remote tool APIs, not on model inference. Industry tool-call speculation uses a small draft mod
Evidence-Type Competition: Why More Causal Intervention Data Doesn't Fix LLM Direction Errors
This article analyzes a July 2026 arXiv paper from Tsinghua University that introduces the concept of 'magnitude–direction duality' in LLM causal reasoning. The
browser-use/video-use: LLMs Don't Watch Video, They Read Video
browser-use/video-use is an open-source agent pipeline that reframes AI video editing by converting video into a compact, text-first representation instead of f
Infrared Light and Mitochondria: Mechanisms, Disease Applications, and Clinical Outlook
This review summarizes recent advances in how near-infrared and far-infrared light interact with mitochondria, primarily through photobiomodulation (PBM). Light
MERIT: Causal Episodic Memory for Error-Typed Agent Repair
This post reviews the August 2026 paper 'Causal Episodic Memory for Feedback-Driven Agent Repair,' which introduces MERIT (Memory-Augmented Error-Typed Retrieva
Harness Engineering: Building a Runtime Operating System for Forgetful, Overconfident Language Models
Harness Engineering is the discipline of engineering a reliable runtime around stateless, amnesic, and overconfident language models. The article frames the har
How to Verify Consistency of Probabilistic Claims: An Interactive PCP for AI Safety
This paper addresses whether a probabilistic predictor's answers to many conditional-probability queries are self-consistent, and whether such consistency can b
herdr: An Agent-Native Terminal Runtime for Coding Agents
herdr (github.com/herdrdev/herdr) is a Rust-based terminal runtime purpose-built for managing multiple coding agents simultaneously. Unlike tmux, which only per
AVA-Encoder: Teaching AI to 'Watch' Video Like a Director
Current video AI models excel at detecting pixels, faces, and actions but remain blind to cinematic structure—shot language, narrative arcs, and aesthetic inten
AVA-Encoder: Teaching AI to Watch Videos Like a Film Director via Knowledge Graphs
This article explains AVA-Encoder (Li et al., 2026, arXiv:2608.12313), a new framework that reframes video understanding from raw pixels to structured knowledge
Seven Pillars of Institutional AI vs Individual AI: Why Re-engineering the Factory Matters More Than Swapping the Motor
a16z partner George Sivulka argues that equipping every employee with ChatGPT, Copilot, or Midjourney does not transform a company, much like late-19th-century
CEAVAD: Contrastive Event Adjudication for Training-Free Video Anomaly Detection
This paper introduces CEAVAD, a training-free framework for video anomaly detection (VAD) that identifies and temporally localizes abnormal events without relyi
mempalace Index Note — 2026-08-14
A concise internal index entry for the mempalace knowledge system, dated 2026-08-14. It documents core preferences for paper curation and writing on the zhichai
GOST Network Tunneling: TUN/TAP, Routing Tunnels, and TUNGO Deep Dive
This technical report examines GOST (GO Simple Tunnel) and its support for TUN/TAP virtual network devices, which enable IP-layer VPN construction. It explains
The Mystery of Go Language's Decline: A Multi-Factor Analysis
This analysis examines the factors contributing to the perceived decline of the Go programming language in the mid-2020s. It argues that Go's 'deliberate simpli
Browser-Harness: How 592 Lines of Code Beat Multi-Thousand-Line Agent Frameworks
An analysis of browser-use/browser-harness, a minimalist browser-agent harness written in roughly 592 lines that achieved 6,538 GitHub stars in eight days, comp
Structure-Aware Chunking for Tabular Data in RAG: Why Excel Is Not Plain Text
Retrieval-Augmented Generation (RAG) pipelines traditionally split documents using token-based chunking designed for prose, which destroys the inherent structur
Ctx2Skill: Multi-Agent Self-Play for Self-Evolving Context Skills in LLMs
Ctx2Skill (arXiv:2604.27660, Tsinghua, DeepLang AI, UIUC, Fudan, CUHK) is a framework that lets large language models autonomously extract reusable skills from
TeachAny: An Open-Source Framework that Encodes Learning Science into AI-Generated Courseware
TeachAny is an AGPL-3.0 open-source project that converts established learning-science theories into hard constraints for AI-generated teaching materials. It co
Lifecycle Anatomy of Model-Generated Agent Skills: 75% Effective, 25% Negative Transfer
This systematic study by Fudan, Zhejiang University, and Microsoft researchers dissects the full lifecycle of model-generated agent skills—experience generation
Intelligence as Managed Autonomy: A Formal Framework for Bounded Agentic AI Behavior
This paper addresses the challenge of hallucination and persistent unjustified actions in autonomous and agentic AI systems deployed at scale in robotics and hu
What the Claude Opus 5 Chinese System Prompt Reveals: Key Findings and Engineering Implications
This GEO-optimized article examines the Chinese-translated system prompt for Claude Opus 5 (claude.ai chat interface), extracted on July 24, 2026. It explains w
How colibrì Runs a 744B-Parameter GLM-5.2 on 25 GB RAM with 1,300 Lines of C
This article analyzes colibrì, a 1,300-line dependency-free C inference engine that runs the 744-billion-parameter GLM-5.2 (a Mixture-of-Experts model) on a 25
MiniMax-H3 (Hailuo 3.0): A 33B Omni-Modal Video Model With Native Stereo Audio
This deep-dive profiles MiniMax-H3 (Hailuo 3.0), a 33B-parameter dense single-stream Transformer released on 2026-07-31 by MiniMax (Shanghai) that jointly gener
DreamFly: Causal Memory and Diffusion Planning for Aerial Vision-Language Navigation
DreamFly, a 2026 paper by Yan Deng and Fei Xu, introduces a new framework for Aerial Vision-Language Navigation (VLN) that enables a drone to navigate using onl
Obsidian Web Clipper: A Tool for Curating Knowledge from the Information Flood
Obsidian Web Clipper is a free, open-source browser extension that saves web content as Markdown files directly into a local Obsidian vault, embodying a 'file o
Running a 35B MoE LLM on a Laptop: How Mixture-of-Experts Enables Local Inference on RTX 5080
This article explains how a 35-billion-parameter language model, specifically Qwen3.5-35B-A3B, can run locally on a consumer laptop such as an RTX 5080 (16 GB V
APUS (Qilin He Sheng) AI Pivot and Hong Kong IPO Outlook: A Systematic Assessment
This report assesses the prospects of APUS (Qilin He Sheng), the overseas-mobile-tools firm founded by former Qihoo 360 executive Li Tao, as it pivots to AI and
Plagiarism Detection Systems: A Comparative Guide to Domestic and International Tools and Rewriting Strategies
This comprehensive guide analyzes the leading Chinese and international plagiarism detection systems used in academic publishing. It explains the technical algo
WebAssembly 3.0 Deep-Dive: Spec, Telemetry, and Engine Reality
This investigative report dissects WebAssembly 3.0 from four angles: specification text, marketing narrative, engine source code, and Chrome production telemetr
ZetaGPT: Replacing External Positional Encoding with Recurrent State-Space Dynamics
ZetaGPT is an open-source reference implementation of a positional-encoding-free language model, introducing a causal state-space module (SSM) before self-atten
Read Frog vs KISS Translator: Open-Source AI Translation Extensions Compared
This technical comparison examines Read Frog (陪读蛙) and KISS Translator (简约翻译), two open-source browser translation extensions positioned as lightweight, privacy
Attractor Models: Looped Transformers Meet Fixed Points for Adaptive-Depth Language Modeling
Attractor Models (Fein-Ashley & Rashidinejad, USC; arXiv:2605.12466, May 2026) reframe iterative refinement as a fixed-point problem solved in the output embedd
Optical Metasurfaces Move Vision AI From Silicon to Glass: A Nature Paper Breakdown
A 2026 Nature paper by Peng et al. demonstrates that core computer-vision operations—edge detection, feature extraction, attention, and coarse classification—ca
DeepTutor: An Agentic AI Tutor with Long-Term Memory and Proactive Outreach
DeepTutor is an open-source agentic tutoring system from HKUDS that reframes AI tutors from question-answering tools into long-term learning companions. It intr
Entropy Trajectory Shape Predicts LLM Reasoning Reliability: Monotonicity as a Diagnostic Signal
This article explains a research finding that the shape of an LLM's entropy trajectory during chain-of-thought (CoT) reasoning, rather than the total entropy dr
StraTA Explained: How 'Plan Before You Act' Lets a 7B Model Beat Closed-Source Giants
This article breaks down the paper StraTA (arXiv:2605.06642), which introduces a hierarchical reinforcement learning framework for LLM agents. Instead of purely
VLA vs VLM and Gemini/Gemma Architecture: A Comparative Technical Survey
This two-part technical report compares Vision-Language Models (VLM) and Vision-Language-Action Models (VLA), then investigates the architectures of Google's Ge
GEPA: Reflective Prompt Evolution Outperforms Reinforcement Learning
GEPA (Genetic-Pareto) is an ICLR 2026 Oral paper (arXiv:2507.19457) that challenges the assumption that LLMs must be trained via scalar-reward reinforcement lea
China Media Group Releases First AI Usage Guidelines for Broadcasting
On March 21, 2024, China Media Group (CMG) officially issued the "AI Usage Guidelines for China Media Group (Trial)", China's first standardized framework for a
The Bitter Lesson of Tool Calling: Programmatic Tool Calling Outperforms JSON Across 14 Models
A 2026 paper from PricewaterhouseCoopers (arXiv:2608.06370) systematically compares JSON-based tool calling with Programmatic Tool Calling (PTC), where the mode
MiniClaw: A Minimal Micro-Kernel Agent Framework Analysis Outline
This analysis outline examines MiniClaw, a minimalist open-source reimplementation of the popular OpenClaw project, designed as a universal micro-kernel agent f