docs: add daily news EN 2026-05-25

Author: yanglbme

src/content/en/daily/2026-05-25.md

@@ -0,0 +1,101 @@
+---
+title: "Awesome AI Daily | 2026-05-25"
+date: "2026-05-25"
+tags: ["AI Security", "AI Wearables", "AGI Debate", "Anthropic", "Claude Code", "ByteDance", "AI Bias", "Microsoft Webwright", "NVIDIA", "StepFun"]
+summary: "Google navigates AI security challenges in real time; Amazon Bee wearable reviewed; Hassabis vs LeCun on AGI status; Anthropic may keep supplying Claude to NSA; Claude Code discovers AI reasoning algorithms; ByteDance's new long-document training paradigm; AI model default selection bias risks; Microsoft releases Webwright web agent framework; NVIDIA launches Gated DeltaNet-2; StepFun releases StepAudio 2.5 Realtime"
+---
+
+## 1. Everyone Is Navigating AI Security in Real Time — Even Google
+
+The AI security landscape is in a critical transition phase, with major tech companies forging strategies through practice. TechCrunch reports that even Google is grappling with real-time AI security challenges. From model deployment to data protection, the industry lacks unified standards, and every player is learning by doing.
+
+> **Awesome AI View:** AI security is no longer a theoretical concern — it's a daily operational challenge for every AI company. The absence of industry standards means first movers have the opportunity to define the rules, but also bear higher trial-and-error costs.
+
+Source: TechCrunch | May 24, 2026
+Link: https://techcrunch.com/2026/05/24/everyone-is-navigating-ai-security-in-real-time-even-google/
+
+## 2. Amazon Bee Wearable Review: Convenience Meets Privacy Anxiety
+
+A TechCrunch editor reviewed Amazon's latest AI wearable device, Bee, finding it raises privacy concerns alongside its convenience promises. Like other AI wearables, Bee provides AI assistance in daily life, but its continuous sensing and data processing capabilities leave users "both intrigued and slightly creeped out."
+
+> **Awesome AI View:** The core tension with AI wearables is clear: the more capable they are, the more privacy-invasive they become. How users balance convenience against privacy will determine the market trajectory of these products.
+
+Source: TechCrunch | May 24, 2026
+Link: https://techcrunch.com/2026/05/24/i-tried-amazons-bee-wearable-and-am-both-intrigued-and-slightly-creeped-out/
+
+## 3. Hassabis vs LeCun on AGI: Are We at the "Foothills of the Singularity"?
+
+DeepMind founder Demis Hassabis believes humanity is already "standing in the foothills of the singularity," while Meta's chief AI scientist Yann LeCun insists current AI systems lack genuine intelligence. Gemini co-lead Oriol Vinyals takes a middle ground: today's models would have looked like AGI seven years ago, but they still can't learn from experience or produce real breakthroughs.
+
+> **Awesome AI View:** This debate reflects a fundamental divide in how the AI community defines "intelligence." Hassabis views AI progress from a capability evolution perspective, while LeCun points out architectural limitations from a cognitive science standpoint. Both may be right — we've made enormous progress, but there's still an essential gap to true general intelligence.
+
+Source: The Decoder | May 24, 2026
+Link: https://the-decoder.com/deepminds-hassabis-sees-humanity-in-the-foothills-of-the-singularity-while-lecun-says-current-ai-isnt-intelligent/
+
+## 4. Anthropic May Keep Supplying Claude to NSA Despite Pentagon "Supply Chain Risk" Label
+
+Despite being flagged as a "supply chain risk" by the Pentagon, Anthropic is likely to continue providing AI models to the US National Security Agency (NSA). Intelligence agencies currently lack access to Nvidia's latest Grace Blackwell chips, and Anthropic's "Mythos" model reportedly runs on older hardware. The controversial "any lawful use" clause that derailed earlier negotiations is not part of this deal.
+
+> **Awesome AI View:** This illustrates the delicate balance AI companies must strike between national security partnerships and commercial interests. A supply chain risk designation isn't a ban, but it does increase compliance costs. Anthropic's decision will set precedents for the entire AI industry's relationship with government agencies.
+
+Source: The Decoder | May 24, 2026
+Link: https://the-decoder.com/anthropic-may-keep-supplying-claude-to-the-nsa-despite-being-flagged-as-a-supply-chain-risk-by-the-pentagon/
+
+## 5. Researchers Let Claude Code Discover AI Reasoning Algorithms for Just $40
+
+Researchers from the University of Maryland, Google, Meta, and other institutions used the AutoTTS framework to let the Claude Code programming agent independently discover control algorithms for AI reasoning. The resulting algorithm cuts compute by approximately 70% compared to standard self-consistency methods while maintaining equivalent accuracy. The entire search cost just $40 and took 160 minutes.
+
+> **Awesome AI View:** This is a textbook case of AI-assisted AI research. Discovering optimization algorithms that humans likely wouldn't have designed — for just $40 and 160 minutes — signals enormous potential for AI in optimizing its own training and inference methods. The AutoTTS framework could become a new paradigm for AI research.
+
+Source: The Decoder | May 24, 2026
+Link: https://the-decoder.com/researchers-let-claude-code-discover-ai-scaling-algorithms-that-humans-probably-wouldnt-have-designed/
+
+## 6. ByteDance Study: Asking LMMs Questions Beats Transcribing Text for Long Document Training
+
+ByteDance Seed's research demonstrates that a 7B-parameter model can answer questions on long documents (including images) more reliably than much larger models, even when documents are four times longer than anything seen during training. Instead of transcribing pages one by one, the model learns by answering questions and locating key passages autonomously.
+
+> **Awesome AI View:** This research challenges the traditional paradigm of long document processing. Through Q&A-driven training, even small models can excel at long document comprehension, which has significant implications for reducing computational costs and improving efficiency in AI applications.
+
+Source: The Decoder | May 24, 2026
+Link: https://the-decoder.com/bytedance-study-finds-that-asking-lmms-questions-beats-making-it-transcribe-text-for-long-document-training/
+
+## 7. Why You Shouldn't Leave Model Selection on Default in Copilot, Gemini, and Other AI Tools
+
+Mathematician Adam Kucharski's tests revealed that Microsoft Copilot invents country differences where none exist when analyzing data — delivering detailed stereotypes instead of accurate results, even when fed identical datasets with different country labels. Only switching to "thinking models" catches this trick, but only if users know when to reach for them.
+
+> **Awesome AI View:** This is a vivid example of AI bias in action. Default models often prioritize "helpfulness" over "accuracy," producing stereotypes on socially sensitive topics. Users should proactively choose more rigorous reasoning models, especially in data analysis scenarios.
+
+Source: The Decoder | May 24, 2026
+Link: https://the-decoder.com/why-you-shouldnt-leave-model-selection-on-default-in-copilot-gemini-and-other-ai-tools/
+
+## 8. Microsoft Research Releases Webwright: Terminal-Native Web Agent Framework Scores 60.1% on Odysseys
+
+Microsoft Research has released Webwright, a terminal-native web agent framework. It achieves 60.1% on the Odysseys benchmark, significantly outperforming base GPT-5.4's 33.5%, and reaches 86.7% on Online-Mind2Web. Webwright translates web operations into terminal commands, offering a more reliable approach to web automation.
+
+> **Awesome AI View:** Webwright represents a new approach to web agents — executing through terminal-native commands rather than simulated browser operations, improving reliability and controllability. The 60.1% Odysseys score suggests web agents are approaching practical usability levels.
+
+Source: Marktechpost | May 24, 2026
+Link: https://www.marktechpost.com/2026/05/24/microsoft-research-releases-webwright-a-terminal-native-web-agent-framework-that-scores-60-1-on-odysseys-up-from-base-gpt-5-4s-33-5/
+
+## 9. NVIDIA Releases Gated DeltaNet-2: New Linear Attention Architecture Surpasses Mamba-3
+
+NVIDIA's AI team has released Gated DeltaNet-2, a novel linear attention layer that decouples erase and write operations in the delta rule. The model employs channel-wise erase and write gating mechanisms, outperforming KDA and Mamba-3 at the 1.3B parameter scale.
+
+> **Awesome AI View:** NVIDIA continues to push the boundaries of efficient sequence modeling architectures. Gated DeltaNet-2's design approach shows there's still substantial room for optimization in linear attention mechanisms, potentially offering more efficient alternatives for ultra-long context processing.
+
+Source: Marktechpost | May 24, 2026
+Link: https://www.marktechpost.com/2026/05/24/nvidia-ai-releases-gated-deltanet-2-a-linear-attention-layer-that-decouples-erase-and-write-in-the-delta-rule/
+
+## 10. StepFun Releases StepAudio 2.5 Realtime: End-to-End Voice Model with Roleplay-Specific RLHF
+
+StepFun has released StepAudio 2.5 Realtime, an end-to-end real-time speech large language model. The model introduces roleplay-specific RLHF optimization and paralinguistic comprehension capabilities, enabling it to perceive and process non-textual information in speech such as emotions and intonation.
+
+> **Awesome AI View:** Real-time voice models are evolving from "can speak" to "speaks naturally." Roleplay RLHF and paralinguistic understanding make AI voice interactions more natural, which has important value for customer service, virtual assistants, and entertainment applications.
+
+Source: Marktechpost | May 24, 2026
+Link: https://www.marktechpost.com/2026/05/24/stepfun-releases-stepaudio-2-5-realtime-an-end-to-end-voice-model-with-roleplay-specific-rlhf-and-paralinguistic-comprehension/
+
+## Other Developments
+
+- **36Kr** reported on Xiaomi's通报 of air conditioner vacuum-pumping fraud, showing deepening AI quality inspection applications in manufacturing (May 25, 2026)
+- **36Kr** hosted a roundtable at the 2026 AI Partner Beijing Yizhuang AI+ Industry Conference discussing "Which赛道 will the next killer AI product emerge from?" (May 23, 2026)