Tuesday, July 28, 2026

1. LFM2.5-Encoders: New Models for Fast Long-Context Inference on CPU
   T1Hugging Face：Blog（RSS）

   Liquid AI released the LFM2.5-Encoder series, leveraging sparse attention and mixture-of-experts to dramatically speed up long-context inference on CPU. The models support context windows up to 1M tokens and achieve SOTA on multiple long-context benchmarks. Released under MIT license with weights and inference code on Hugging Face.

2. How to Evaluate LLM Provider Performance Across Latency, Throughput, and Uptime
   T1OpenRouter：Announcements（RSS）

   This article explains key metrics for evaluating LLM provider performance, including p50-p99 latency, throughput, uptime, and quantization, and demonstrates how to enforce evaluation results with routing controls for optimized application performance and cost.

3. An Opinionated Guide to Which AI to Use to Do Stuff
   T2Simon Willison 博客（Atom）

   Ethan Mollick's guide to using AI has shifted in the past year from focusing on chat models (ChatGPT, Claude, Gemini) to emphasizing agentic systems where AI can perform hours of human work in one go. Gemini has been dropped due to lacking a mature entry in the Codex/ChatGPT Work/Cowork category. The post explains ChatGPT's Work and Codex modes, Claude's Cowork and Code modes, and how giving AI access to your computer unlocks more capabilities, while noting the confusing differences between mobile Work mode and desktop Codex.

4. Microsoft Launches Its First Cybersecurity AI Model and Agentic Security System
   T2TechCrunch：AI（RSS）

   Microsoft launched its first AI security model and a new agentic cybersecurity system this week, aiming to bolster cyber defense with AI. The system includes autonomous agents that can detect, analyze, and respond to threats automatically, reducing human intervention. The model focuses on the security domain, marking a significant step in Microsoft's AI security efforts.

5. Moonshot AI Releases 2.8 Trillion Parameter Kimi K3 Weights with Stricter License Terms
   T2Simon Willison 博客（Atom）

   Moonshot AI has fulfilled its promise by releasing the 2.8 trillion parameter Kimi K3 model weights on Hugging Face, totaling 1.56TB. Unlike K2's modified MIT license, K3's license drops the open-source label and requires MaaS businesses with over 100 million MAU or $20 million monthly revenue to sign a separate agreement. OpenRouter already offers K3 from 7 providers at the same price as the official API.

6. Anthropic's Position on Open-Weights Models
   T1Anthropic：News（网页）

   Anthropic CEO Dario Amodei issues a statement on open-weights models, emphasizing safety and responsible development. The article explores the balance between fostering innovation and preventing misuse, proposing new safety standards and regulatory frameworks. Anthropic advocates for openness while ensuring models are not used for malicious purposes, calling for industry-wide governance rules.

7. Anthropic CEO Dario Amodei: Not Opposed to Open-Weight Models, but Fears Chinese AI Rise
   T2TechCrunch：AI（RSS）

   In a recent interview, Anthropic founder and CEO Dario Amodei made it clear he is not against open-weight models but insists they must be paired with safety measures. He specifically expressed concerns about the rapid growth of China's AI capabilities, citing geopolitical risks, and called for the U.S. to maintain leadership while balancing openness and regulation.

1. GitHub Copilot for Beginners: Getting Started with AI Agents and Canvases
   T1GitHub Blog（网页）

   This article is aimed at beginners of GitHub Copilot, covering how to start projects, use AI agents, explore canvases, and streamline your development workflow. It includes core operational tips to help developers get started quickly and boost productivity.

2. PSA: Your Claude Shared Chats and Artifacts May Have Ended Up on Google
   T2TechCrunch：AI（RSS）

   A privacy flaw in Anthropic's Claude: the 'share chat' feature creates public links that are not blocked from search engine crawlers by default, causing many user conversations and Artifacts (e.g., code, documents) to be indexed by Google. The issue affects already shared links; if you've shared sensitive info, it may have leaked. Anthropic has not fully fixed it yet. Users are advised to delete or revoke links containing confidential content immediately.

3. The Harness Is All You Need: A Practical GitHub Copilot Workflow Guide
   T1GitHub Blog（网页）

   This article presents a practical software engineering workflow built on GitHub Copilot, covering prototyping, planning, implementing, and reviewing code. It emphasizes leveraging Copilot’s existing capabilities rather than chasing every new AI tool, sharing real-world usage tips and configuration recommendations to help developers boost productivity in daily work.

4. Threads Users Can Now Chat with Meta AI in Their DMs
   T2TechCrunch：AI（RSS）

   Meta announced Monday that it is rolling out its Meta AI chatbot within Threads' direct messages, giving users a way to interact with the AI assistant directly in conversations. This update embeds AI capabilities into the core communication scenario of the social platform, marking another significant step in Meta's push to promote AI assistants across its social apps. The feature is being gradually rolled out, allowing users to invoke Meta AI in the Threads DM interface for Q&A, content generation, and more.

1. Microsoft CEO Nadella: Companies That Rely Solely on One AI Model Will Not Survive
   T2TechCrunch：AI（RSS）

   Microsoft CEO Satya Nadella warned in a recent interview that businesses relying entirely on a single AI lab's model (e.g., OpenAI or Google) will face significant risks and may ultimately fail to survive in the market. He urged companies to build diversified AI capabilities to avoid vendor lock-in, reflecting the industry shift from monolithic models to multi-model orchestration.

2. Expanding Partnership with Cognizant: Claude Embedded Across Platforms, 30,000+ Associates Trained, and Global Premier Partner Status
   T1Anthropic：News（网页）

   Cognizant has embedded Anthropic's Claude model across its platforms, trained over 30,000 associates, and been designated as a Global Premier Partner in the Claude Partner Network. This collaboration marks the deep enterprise adoption of Claude, showcasing the potential for large-scale AI deployment in major service providers.

3. OpenAI's Hugging Face Breach Reignites Debate Over AI Alignment and Control
   T2TechCrunch：AI（RSS）

   A misconfigured repository on Hugging Face exposed OpenAI's internal model weights, training data, and internal notes, reigniting the debate over AI alignment (making models conform to human intent) versus control (physically restricting model capabilities). One camp argues stronger alignment prevents loss of control, while the other insists on sandboxes and access limitations as hard controls. The breach reveals the difficulty even advanced labs face in securing model assets, underscoring the urgency of safety governance.

1. Beyond Scale and Generation: Understanding Language Model-based Entity Matching
   T2arXiv AI/CL/LG（论文）

   This paper conducts a controlled factorial study on three matcher architectures (bi-encoder, cross-encoder, generative), three model variants, and three model sizes (from Qwen3 family) with 1,215 fine-tuning runs across nine datasets. Key findings: model variant is critical for bi-encoders; cross-encoders consistently outperform bi-encoders but larger models can narrow the gap; generative matchers do not universally outperform cross-encoders, only excelling under distribution shift or cross-dataset transfer; larger models rely more on shortcut learning and do not necessarily perform better. The study clarifies the interplay of architecture, model variant, and scale in entity matching.

2. MMOE: Modernizing Diffusion Transformers with Efficient Expert Design
   T2arXiv AI/CL/LG（论文）

   MMOE is a modernized diffusion transformer architecture that systematically introduces routed experts, shared and lightweight experts, gate-residual routing, and attention-residual information reuse into AIGC generation. Under a budget of a single 8×H100 node with batch size 256 for 400k steps, MMOE achieves lower FID at every recorded checkpoint than dense and intermediate sparse-expert baselines, converges faster, and attains the best quality-cost balance among sparse variants. Routing analysis shows stable expert specialization across depth, substantial use of lightweight routes, and modest step-to-step routing changes during denoising. The study suggests that AIGC foundation models can follow the balanced scaling path of LLMs by importing proven efficiency designs rather than simply increasing total parameters and sparsity ratios.

3. ClinFusion: A Vision-Centric Multimodal LLM System for Holistic Medical Understanding
   T2arXiv AI/CL/LG（论文）

   ClinFusion is a vision-centric multimodal large language model designed to tackle heterogeneity in medical image understanding and evaluation alignment. It proposes a cascaded spatial-aware locality fusion encoder to unify 2D and 3D medical images, and introduces MedIF-Bench for instruction-following and a region-of-interest-grounded metric for report generation evaluation. Across 24 benchmarks covering VQA, report generation, and instruction-following, ClinFusion outperforms open-source medical MLLMs (e.g., Hulu-Med, Lingshu) on 20, and surpasses proprietary models like GPT-5.2 and Gemini-3-Flash on 13 out of 16. A blinded evaluation by board-certified radiologists confirms ClinFusion produces the highest-quality reports, and its RoI-grounded metric achieves the strongest correlation with expert judgment. It can be further augmented with agentic tool use for retrieval-augmented and tool-assisted clinical workflows.

4. Rethinking Classifier-Free Guidance in On-Policy Diffusion Distillation
   T2arXiv AI/CL/LG（论文）

   On-policy distillation (OPD) adapts diffusion models by querying a teacher, but the role of classifier-free guidance (CFG) remains poorly understood. Existing methods directly match guided velocities, yet this objective suffers from under-identification at the branch level: positive and negative branch errors can compensate. When the teacher's negative branch contains privileged information unavailable to the student, naive matching induces Negative Branch Asymmetry (NBA). The authors propose Positive-Direction Matching (PDM), a branch-aware objective that separately constrains the positive prediction and CFG direction. Experiments on dense-to-sparse video control demonstrate more robust knowledge transfer.

5. KANEx: Translating Kolmogorov-Arnold Networks' Interpretability to Medical Explainability
   T2arXiv AI/CL/LG（论文）

   Computer vision models excel in medical applications, but their black-box nature undermines clinician trust. Existing systems pair chest X-ray classifiers with Vision-Language Models (VLMs) for natural-language explanations, yet fail to address the underlying opacity. We propose KANEx, the first framework leveraging the symbolic transparency of Kolmogorov-Arnold Networks (KANs) to ground VLM reasoning, and introduce KAN-Map, a heatmap method derived directly from KANs. On MIMIC-CXR, KAN-based architectures improve semantic similarity and saliency map faithfulness over ResNet/ViT baselines, boosting visual localization and downstream reasoning quality by 10%. Our findings suggest grounding explanations in mathematically interpretable units is a necessary step toward trustworthy medical AI.

6. Efficient LLM-Generated Shuttling Compilers for Complex Trapped-Ion Architectures
   T2arXiv AI/CL/LG（论文）

   This study presents the first demonstration of a single frontier large language model (Claude Opus 4.7) automatically generating and iteratively refining Python code for trapped-ion quantum computer shuttling compilers. Starting from a linear segmented trap, extending to a trap with junctions, and finally to general connected trap graphs, the LLM-generated compilers reduce shuttling timesteps by up to 76% and 39% compared to hand-crafted compilers, with an order-of-magnitude reduction in densely connected architectures. Reproducing results with Claude Fable 5 confirms that unmodified frontier LLMs can produce competitive compilers, cutting development time from months to days.

7. Causal-TS: An Open-Source Python Library for Causal Discovery in High-Dimensional and Nonstationary Time Series
   T2arXiv AI/CL/LG（论文）

   Causal-TS is an open-source Python library from Bloomberg for causal discovery in high-dimensional, nonstationary multivariate time series. It provides four specialized algorithms—CDNOTS, CDNOTS+, CEDAR, and GRACE—alongside wrappers for GES, Granger, LASSO-VAR, and LGES, all sharing a unified conditional independence test layer accelerated by PyTorch on GPU. A regime discovery pipeline detects structural breaks via pluggable changepoint detectors and runs discovery per regime with regime-specific parameters. With a command-line interface, synthetic data generators, and optional DoWhy integration, it offers an end-to-end pipeline from raw time series to causal effect estimates. The library is pip-installable, tested on Python 3.10–3.12, and hosted on GitHub.

8. Kimi K3: Open Frontier Intelligence
   T2arXiv AI/CL/LG（论文）

   Kimi K3 is a 2.8-trillion-parameter Mixture-of-Experts model developed by Moonshot AI, with 104 billion activated parameters, native vision capabilities, and a 1-million-token context window. Built on Kimi Delta Attention and Attention Residuals to enhance information flow, combined with Stable LatentMoE that activates 16 of 896 routed experts per token, it achieves approximately 2.5x scaling efficiency improvement over Kimi K2. Post-training involves reinforcement learning across general, agentic, and coding domains with multiple reasoning-effort levels. Evaluations show frontier-level performance in long-horizon coding, agentic, knowledge, reasoning, and vision tasks, trailing only Claude Fable 5 and GPT-5.6 Sol but outperforming other open and proprietary models. Full model weights are released.

—