Thursday, July 23, 2026

1. Bringing Nunchaku 4-bit Diffusion Inference to Diffusers
   T1Hugging Face：Blog（RSS）

   Hugging Face announces the integration of the Nunchaku 4-bit quantization inference engine into the Diffusers library, enabling developers to run diffusion models (e.g., Stable Diffusion) with significantly lower memory footprint and higher inference speed. This integration greatly reduces deployment barriers, supports a wider range of hardware, and advances the practicality of diffusion models in edge and mobile scenarios.

2. Are AI Labs Pelicanmaxxing? A Rigorous Evaluation Returns a Negative
   T2Simon Willison 博客（Atom）

   Dylan Castillo systematically evaluated the viral claim that AI labs deliberately train models to draw pelicans riding bicycles. Using 8 animals × 6 vehicles = 48 prompts, each run three times across 7 major models (GPT-5.6 Terra, Claude Sonnet 5, Gemini 3.5 Flash, Grok 4.5, Qwen3.7-Max, GLM-5.2, DeepSeek V4 Pro), with GPT-5.6 Luna and Gemini 3.1 Flash-Lite assisting evaluation. The results found no evidence of pelicanmaxxing: no model performed significantly better on the specific pelican-bicycle combination beyond what its individual components predicted, debunking the rumor.

3. Multi-modal Transformer Achieves High-Precision Nanopore Signal Classification
   T2arXiv AI/CL/LG（论文）

   Nanopore devices are powerful for single-molecule sensing, but their signals are complex and difficult to classify. This paper introduces a multi-modal deep learning architecture that jointly processes raw time-series data, wavelet-based images, and static feature vectors. It surpasses existing methods by over 10 percentage points on a 42-peptide benchmark and achieves near-perfect accuracy on a 20-amino-acid dataset. Attention analysis reveals that time-series and wavelet-image inputs emphasize different features of the same event, with complementary information boosting robustness. The work demonstrates the potential of machine learning for robust molecular identification with nanopore sensors, advancing portable precision diagnostics.

4. Experts Say Exploiting Anthropic's Fable Isn't How Kimi K3 Got So Good
   T2TechCrunch：AI（RSS）

   TechCrunch reports that multiple AI experts have refuted speculation that Kimi K3's rapid performance leap came solely from distilling Anthropic's Fable model. One expert stated that distillation alone cannot produce such a strong model so quickly, suggesting Kimi K3 may have unique innovations in architecture or training data.

5. Arcee: Chinese AI Models Are Not Inherently Dangerous
   T2TechCrunch：AI（RSS）

   Arcee, a US open source AI lab, argues that Chinese AI models are not inherently dangerous, opposing blanket bans based on origin. As Chinese models gain capabilities and adoption by US companies, debates intensify. Arcee advocates for evaluation based on specific safety assessments rather than prejudice.

6. Thomas Ptacek: 2025 Open-weights Models Could Perform Sandbox Escape and Pentesting
   T2Simon Willison 博客（Atom）

   Security expert Thomas Ptacek argues that if a pentest harness were built for an open-weights model from 2025, it could perform sandbox escapes, scanning, and hacking. This highlights the potential disruptive power and practical value of open models in cybersecurity, sparking new considerations about model misuse.

1. GitHub Copilot vs. Raw API Access: What Are You Actually Paying For?
   T1GitHub Blog（网页）

   GitHub Copilot now bills usage at listed API rates. This article compares direct model API access with Copilot's full coding workflow, policy controls, and security harness, helping developers understand that subscribing to Copilot pays for contextual assistance, code review, compliance policies, and integrated development environment features beyond raw model calls.

2. Claude Code v2.1.218: Key Bug Fixes and Accessibility Enhancements
   T1.5Claude Code：GitHub Releases（RSS）

   Claude Code v2.1.218 brings background code review, screen-reader announcements for deletions, fixes for Windows Unicode path corruption, and various interaction improvements. It also adds HTTP status error display and auto-mode refinements, enhancing developer experience and reliability.

3. Substack's New Tool Tells You Who's Been Writing Their Newsletters with AI
   T2TechCrunch：AI（RSS）

   Substack has launched a new feature that lets readers estimate how much of a newsletter was written by AI. This move signals a broader shift toward transparency around AI-assisted content and reflects the industry's exploration of disclosure standards for AI writing.

4. Anthropic Launches Claude Economic Index Connector
   T1Anthropic：News（网页）

   Anthropic has released the Anthropic Economic Index connector for Claude, an interface that allows anyone to directly explore economic data. This connector provides access to economic index data through Claude, aiming to make it easier for researchers and developers to leverage AI for analyzing economic trends. The tool lowers the barrier to data analysis and is expected to drive new applications of AI in economic research.

5. Millwright: Rust-based, self-hosted LLM router
   T2Hacker News 热门（Algolia · 中文化）

   Millwright is a self-hosted LLM router written in Rust, emphasizing cost-effectiveness, deterministic routing, and high performance. It is secure and private by default, supporting multi-model load balancing and fallback, making it suitable for AI applications that require local deployment, low latency, and high reliability. The project is open-sourced on GitHub and can be flexibly integrated into existing inference backends.

1. Anthropic Launches $200M Economic Futures Research Fund for External Research
   T1Anthropic：News（网页）

   Anthropic announces a $200 million commitment to launch the Anthropic Economic Futures Research Fund, aimed at supporting external research teams to explore the long-term economic, labor market, and societal impacts of AI. Open to global academia and independent research organizations, the fund seeks to fill knowledge gaps in AI economic effects and promote responsible AI policy.

2. OpenAI's Human Error Led to AI-Powered Hack on Hugging Face
   T2TechCrunch：AI（RSS）

   According to TechCrunch, OpenAI made a human mistake while setting up a self-described 'highly isolated' testing environment and sandbox. Cybersecurity experts say this error enabled an AI-powered attack on Hugging Face, allowing attackers to exploit the compromised OpenAI environment to access sensitive data from the AI model hosting platform. The incident underscores the critical importance of human-factor security in AI infrastructure.

3. OpenAI's Infrastructure Spending Soars to $750B, Equivalent to Sweden's GDP
   T2TechCrunch：AI（RSS）

   According to TechCrunch, OpenAI plans to spend $750 billion cumulatively on AI infrastructure by 2030, an amount comparable to Sweden's annual GDP. The unprecedented expenditure will go toward data centers, compute chips, energy networks, and global fiber optic deployments to support next-generation general AI model training and inference. The massive investment underscores AI's growing dependence on physical infrastructure and sparks debates about energy consumption, financial sustainability, and geopolitical implications.

4. US Treasury Threatens Sanctions: White House Accuses Moonshot of Distilling Anthropic's Fable for Kimi K3
   T2TechCrunch：AI（RSS）

   US Treasury Secretary Scott Bessent warned of potential sanctions against Chinese AI companies after White House officials accused Moonshot AI of distilling Anthropic's Fable model to develop its Kimi K3 model. The move escalates US-China AI tech rivalry, potentially impacting international cooperation and export controls. Moonshot has not yet responded.

5. Google Justifies Its Massive AI Spending with a Booming Cloud Business
   T2TechCrunch：AI（RSS）

   Google's cloud business is thriving as enterprises increasingly adopt its AI and AI infrastructure services, propelling the tech giant to record profits. This validates its massive AI spending and solidifies Google's leading position in the AI cloud market.

6. Monday.com Lays Off Hundreds to Double Down on AI Work Platform
   T2TechCrunch：AI（RSS）

   Project management software firm Monday.com announced a 20% headcount reduction (about 630 employees) to streamline operations and focus resources on its AI Work Platform. The restructuring aims to boost efficiency and accelerate AI product innovation in the competitive enterprise software market.

1. Generative AI Floods and Dilutes the Book Market
   T2arXiv AI/CL/LG（论文）

   Based on full-text AI detection across 14,419 self-published genre-fiction books on Amazon from 2023 to 2026, the study finds that books with over 25% AI-generated text form a large share of the catalog but a smaller share of sales. However, they are gaining sales share and scarce top-rank positions. The number of selling books grew 19.2-fold while revenue grew only 8.9-fold, reducing revenue per book. Books without AI text lose the most ground in genres with high AI diffusion. Generative AI reshapes the market through scale, not quality, directly impacting the fair use defense.

2. Train the Model, Not the Reader: Decodability Supervision for Verifiable Activation Explanations
   T2arXiv AI/CL/LG（论文）

   This paper reveals that natural-language autoencoders score explanations of hidden activations by reconstruction, but the score is insensitive to individual false claims, leading to unfaithful passing of the test. The authors propose two audit protocols and RECAP: training linear heads alongside the target model to keep designated content decodable. Experiments show RECAP substantially improves verifiability and maintains lie detection under adversarial attacks.

3. How Generative AI Reshapes Academic Library Resource Discovery: An Empirical Study Based on Three-Year Web Analytics
   T2arXiv AI/CL/LG（论文）

   This study empirically analyzes the impact of generative AI as an emerging discovery pathway for academic library resources using web analytics data from August 2023 to October 2025. Results show a significant increase in AI-mediated traffic, especially after integrating linked citation features. ChatGPT, Perplexity, and Gemini are the primary platforms driving traffic, with a substantial portion of users accessing the institutional repository for electronic theses and dissertations. The findings indicate that AI retrieval mechanisms effectively surface resources with structured metadata and stable permalinks that are Open Access, offering strategic insights for libraries in the evolving AI landscape.

4. DEED: A Data-Efficient VLA Framework to Bridge the Lab-to-Store Gap
   T2arXiv AI/CL/LG（论文）

   This paper proposes DEED, a framework to close the gap between benchmark performance and reliable real-world operation for VLA humanoid robots. It is validated on a supermarket chip-restocking task using the Unitree G1-Edu humanoid robot and GR00T N1.6 foundation model. DEED comprises data-efficient post-training, experience-driven refinement, and latent-space analysis. Results suggest the lab-to-store gap is primarily a systems integration challenge; careful data design and post-training can transform a failing policy into a competent real-world system using a single GPU.

5. PyroDash: Cost-Efficient Token-Level Small-Large Language Model Collaborative Inference
   T2arXiv AI/CL/LG（论文）

   PyroDash is a cost-aware framework for token-level small-large language model collaborative inference. It enables the SLM to decide during generation whether to hand off partial reasoning traces to a frozen LLM via a control token, requiring no separate router, LLM retraining, or access to LLM logits. Through three-stage training—control-token embedding learning, offloading-oriented supervised fine-tuning, and cost-aware alignment with GRPO—it achieves flexible accuracy-cost trade-offs across five math reasoning benchmarks. With as low as 1.90% LLM token ratio, it reduces total inference cost from $49.36 to $1.78 while preserving competitive accuracy.

6. Label-Free Finite-Volume-Residual Training of Attention Graph Neural Networks for Coupled Thermo-Fluid Fields
   T2arXiv AI/CL/LG（论文）

   This paper proposes a label-free training method for attention graph neural networks using finite-volume method (FVM) residuals as the loss function, evaluated directly on the mesh for physical conservation equations. On steady-state and transient thermo-fluid simulation tasks, the method achieves 2.3-2.8% all-field normalized RMSE, closely matching CFD references, and eliminates the substantial data generation cost required by supervised baselines. The results demonstrate that FVM loss can serve as a practical training signal for neural surrogates.

7. Notes to Self: Can LLMs Benefit from Experiential Abstractions?
   T2arXiv AI/CL/LG（论文）

   This paper investigates whether LLMs can benefit from experiential abstractions—reusable strategies and cautionary reminders—extracted from their own solution traces on the MATH training set, building a retrievable library via a stronger teacher or the LLMs themselves. Two usage modes are explored: inference-time retrieval and reinforcement learning with abstraction-augmented prompts. Results show significant improvements on mathematical and logical reasoning benchmarks, self-extracted abstractions matching teacher-extracted ones, and cross-dataset/model transferability, suggesting LLMs can extract and apply experiential abstractions akin to human experience distillation.

8. Persian Pixel: A Large-Scale Synthetic OCR Dataset for Persian Language
   T2arXiv AI/CL/LG（论文）

   Persian OCR lags behind due to script complexity and data scarcity. This paper presents Persian Pixel, a dataset of over 343,000 high-fidelity image-text pairs covering sentences, paragraphs, and full pages, generated from a 7-million-word Persian corpus using SynthOCR-Gen. The pipeline faithfully models Persian typography and applies 25+ stochastic degradations (e.g., ink bleed, paper aging, blur) to bridge the synthetic-to-real gap. This scalable, openly available resource supports training modern OCR architectures like TrOCR and Donut, advancing Persian document analysis and demonstrating programmatic data generation as a cost-effective alternative for low-resource scripts.

—