Also available in 中文 — Read in 中文

Daily Tech Digest | 2026-08-18

Today's picks: 15 articles — 2026-08-18

Tech News Overview

Today’s tech news revolves around three main threads: the AI infrastructure sector is seeing major consolidation—Stripe plans to acquire OpenRouter for over $7 billion, and Nvidia has disclosed OpenAI’s 12GW computing deployment commitment; AI safety and ethics controversies continue to escalate—Copilot Autofix has been exposed for introducing vulnerabilities that led to the compromise of Snowflake Jira, Amazon’s destruction of books to train AI has been tracked and exposed, and Anthropic’s text watermarking has sparked a major debate on writing ethics; in addition, technical and scientific advances such as the DuckDB v2.0 preview and sound-wave-driven micro robots are also making headlines.

🤖 AI and Machine Learning

Build an MCP Server in Rust with rmcp: AWS Official Hands-On Tutorial

The AWS developer community published a detailed tutorial demonstrating how to build an MCP (Model Context Protocol) server in Rust using the official rmcp SDK. The tutorial covers the complete workflow: from tool definition, JSON Schema authoring, and AWS service invocation, to stdio transport and manual protocol testing, and finally shows how to integrate the server with Claude Code. MCP is becoming the de facto standard for AI agents interacting with external tools, and the Rust implementation offers a better choice for teams seeking performance and memory safety. Original link: https://dev.to/aws-builders/build-an-mcp-server-in-rust-with-rmcp-a-walk-through-41o3

GPT 5.6 Sol Released: Roboflow Calls It OpenAI’s Strongest Vision Model Yet

Roboflow published a review stating that GPT 5.6 Sol is OpenAI’s most impressive vision model to date. The review garnered 196 points and 101 comments on HN, sparking considerable discussion. The iteration speed of vision models is visibly accelerating, though the review’s conclusions still await validation through broader benchmark testing. Original link: https://blog.roboflow.com/openai-gpt-5-6/

Anthropic Claude Text Watermark Sparks Writing Ethics Controversy

Daring Fireball published an opinion piece sharply criticizing Anthropic’s built-in text watermarking in Claude as “adulteration” of writing. The article ignited fierce debate on HN, scoring 619 points and 548 comments, making it one of the most-discussed items today. The core of the controversy: Do AI platforms have the right to embed invisible markers in user-generated content? This touches on deep issues such as AI writing ethics, user autonomy, and content ownership. Supporters argue that watermarks help trace AI-generated content and prevent misuse; opponents argue that it infringes on writers’ creative autonomy and may have a disproportionate impact on non-English users. Original link: https://daringfireball.net/2026/08/anthropics_watermark_text_adulteration_in_claude_is_a_perversion_of_writing

AirTag Tracking Exposes Amazon Destroying Books to Train AI, Data Collection Ethics Back in Focus

404 Media published a tracking investigation: investigators hid Apple AirTags inside shipments of rare books and followed them all the way to an Amazon warehouse in Las Vegas. Inside the warehouse, the VGT3 team cuts off book spines to improve scanning efficiency, destroying the original books during digitization, with the scanned data used to train the Nova model. Anthropic was previously also revealed to have a similar “Project Panama.” These practices have sparked intense controversy: How should the copyrights of publishers and authors be protected? Does the destruction of physical books cross the line of business ethics? Physical books are especially valuable because much of their content has never appeared on the internet, and a large number of books published before 2022 contain no AI-generated content. Original link: https://www.ithome.com/0/990/845.htm

Jensen Huang: OpenAI Commits to Deploying Approximately 12GW of NVIDIA Compute by 2030

NVIDIA CEO Jensen Huang revealed via official blog that OpenAI has committed to large-scale deployment of NVIDIA AI infrastructure by 2030, with existing and planned deployments totaling approximately 12GW. If the PORTS-Pike partnership expands beyond the initial 4.25GW phase, the total scale could grow to about 16GW. At this scale, NVIDIA’s corresponding business opportunity is approximately $600 billion. NVIDIA also announced a partnership with SB Energy to secure land, power, and building shell capacity in Portsmouth, Ohio, with OpenAI as the tenant. This marks the AI infrastructure competition escalating from the chip level to a scramble for physical resources such as electricity and land, and the “AI factory” construction logic is reshaping the entire industry chain. Original link: https://www.ithome.com/0/990/834.htm

Amazon Destroys Rare Books to Train AI Models, Data Ethics Sparks Controversy

TechCrunch reports that Amazon is destroying rare books to obtain training data. These books are extremely valuable for training LLMs because models have already consumed publicly available online content, while rare books provide a unique text corpus. This practice has prompted deep reflection on the sources of AI training data and the preservation of cultural heritage. With the surging demand for AI training data, how to strike a balance between technological development and data ethics has become a critical issue for the industry. Original link: https://techcrunch.com/2026/08/17/amazon-once-an-online-bookseller-is-destroying-rare-books-to-train-ai-models/

Why AI Novels Always Fall Apart Around Chapter 30? A Detailed Engineering Solution

A trending article on Juejin analyzed the root causes of AI long-form text generation collapsing around chapter 30: context window limitations lead to forgotten character settings, unresolved foreshadowing, and contradictory plot points. The author proposes engineering solutions, including segmented memory management and persistence of key information. This article offers direct reference value for developers of AI writing tools and also reveals the current technical bottleneck of large models in maintaining consistency across extremely long texts. Original link: https://juejin.cn/post/7674372498315460658

ByteDance Signs AI Copyright Protection Agreement with the Motion Picture Association

ByteDance has reached an agreement with the Motion Picture Association (MPA) to strengthen copyright protection for AI video and image generation models. The agreement covers the Seedance video generation model and the Seedream image generation model, which are offered through apps such as TikTok, CapCut, and Dreamina. ByteDance stated that the new model versions have incorporated stronger intellectual property protection mechanisms. This collaboration marks a shift in copyright governance in AI content generation from unilateral efforts to industry-wide cooperation. Original link: https://www.ithome.com/0/990/847.htm

Demystifying AI Coding: Back to Engineering Practice

An article titled “AI Coding Without the Vibes” has sparked discussion in the developer community. The author argues that many teams currently have overly “mystical” expectations for AI coding tools, lacking systematic methods for use and evaluation. The article proposes treating AI-assisted coding as an ordinary engineering tool, improving efficiency by defining clear inputs and outputs and establishing validation processes, rather than relying on “feelings.” This pragmatic perspective offers reference value for development teams trying to adopt AI, and also reflects the industry’s shift from the novelty phase to rational implementation.

Original link

The Rise of AI Token Reselling Economy: Arbitrage Opportunities Emerge

As major AI service providers introduce usage-based APIs and subscription quotas, a new market for “AI token reselling” is taking shape. Some analyses indicate that certain users hoard low-cost tokens in bulk (such as quotas bundled with enterprise plans) and resell them at a discount to individuals or small teams that need temporary compute, profiting from the price difference. This phenomenon shows that there is arbitrage space in AI service pricing across regions and plans, and also signals that AI compute is gradually becoming a standardized tradable commodity. However, reselling may violate terms of service, and platforms may tighten controls in the future.

Original link

Judge Who Relied Entirely on AI for Ruling Granted Judicial Immunity by Court

A U.S. court ruled that a judge who relied entirely on AI-generated content in an order is still protected by judicial immunity. The case involved a party accusing the judge of directly adopting AI output as the basis for a ruling without independent review. The court held that judicial immunity is intended to protect judges from civil lawsuits arising from their official acts, and even serious errors or reliance on AI do not constitute an exception. This ruling has sparked controversy at the intersection of law and AI: When AI is deeply involved in judicial decision-making, how should accountability and procedural justice be balanced? More explicit rules may be needed in the future to define AI’s role in the judiciary.

Original link

Open-Source AI Job Search Assistant: Automatically Scans Job Postings and Evaluates Fit

An open-source project named santifer on GitHub has attracted attention. This tool can automatically scan job boards, evaluate job listings on a 1.0-5.0 scale using structured A-F scoring rules, customize resumes, and track application progress. It runs in a local AI coding command-line environment and supports tools such as Claude Code and Codex. For job seekers, this is essentially a customizable automated job search agent that can significantly reduce time spent filtering positions. However, its effectiveness depends on the accessibility of job platform data, and its practicality awaits validation by more users.

Original link

🔒 Security

Reticulum: A Decentralized Mesh Network for Censorship-Resistant Encrypted Communication

Reticulum is a decentralized mesh network protocol that relies on no centralized servers, enabling direct device-to-device interconnection through encrypted communication. The project received 199 points and 65 comments on HN, generating considerable discussion. Networks like this hold significant value for communication freedom in censorship environments and for emergency communications in disaster scenarios, representing a notable project in the wave of decentralizing internet infrastructure. Original link: https://reticulum.network/

AI-Generated Copilot Autofix Introduced a Vulnerability, Leading to the Breach of Snowflake Jira

The Wiz research team disclosed a highly cautionary security incident: AI-generated GitHub Copilot “Autofix” code introduced a security vulnerability that ultimately led to the breach of Snowflake’s Jira system. Autofix was intended to automatically patch vulnerabilities, but flawed code it generated became the entry point for the attack. This incident is a typical case of security risks introduced by AI-assisted coding, highlighting the importance of auditing AI-generated code—developers cannot blindly trust AI’s fix suggestions and still need human review and security testing. Original link: https://www.wiz.io/blog/red-agent-snowflake-copilot-cicd-bug

strix: An Open-Source AI Penetration Testing Tool That Automatically Discovers and Fixes Application Vulnerabilities

An open-source AI penetration testing tool called strix has appeared on GitHub Trending. Written in Python, it automatically discovers and helps fix application vulnerabilities. It combines AI capabilities with traditional security testing, representing a significant trend toward intelligent security tools. For security teams, such tools can significantly lower the barrier and cost of penetration testing. Original link: https://github.com/usestrix/strix

Today’s Focus: The stability and resource allocation of AI infrastructure have become core issues—from the GitHub outage, grid pressure, to chip investment, the “foundation” of the technology ecosystem is under strain.

Hot Article Pool

  1. hackernews Does anyone run Postgres without PgBouncer?
  2. hackernews Asynchronous I/O in DuckDB: Work, Thread, Work
  3. devto Running Gemma 4 on EC2 G5g: Graviton2 AMD with NVIDIA GPU
  4. hackernews A Preview of DuckDB v2.0
  5. devto Build an MCP server in Rust with rmcp: a walk-through 🦀
  6. hackernews Claude: System Prompts
  7. hackernews Firefox for iOS now has a native adblocker
  8. devto Installing Rust for vLLM on Graviton: a G5g walk-through 🦀
  9. juejin Enabling AI Coding Agents to Directly Access CAD Documents: A GitMCP Practical Guide
  10. juejin 👍 Beyond ESLint: Oxc Adopts TypeScript 7 First, Rust and Go Team Up!
  11. hackernews Qwen 3.8 27B is excellent, but it defaults to overthinking things
  12. hackernews Cross-Validation From Scratch and a Surprise at n=100
  13. hackernews Reticulum – Decentralized Mesh Network
  14. devto Shipping a vision-model verdict on Bedrock and Lightsail
  15. ithome Without Motors or Gears, Scientists Create Miniature Boats and Aircraft Powered Entirely by Sound Waves
  16. hackernews AI-Generated GitHub Copilot “Autofix” Allowed Compromise of Snowflake's Jira
  17. github_trending akitaonrails/ai-memory
  18. hackernews Show HN: A public AI whose memory is shared across all users
  19. devto One-Shot UI Side Effects in BlocSignal: Snackbars, Dialogs, and Navigation Without State Pollution
  20. techcrunch Anthropic shares more details about how Claude’s new watermarks will work
  21. hackernews GPT 5.6 Sol is the best "vision" model OpenAI ever released
  22. hackernews Anthropic's ‘watermark’ text adulteration in Claude is a perversion of writing
  23. github_trending usestrix/strix
  24. github_trending immich-app/immich
  25. techcrunch Stripe will reportedly acquire AI gateway startup OpenRouter for $7B+
  26. juejin Stop Using nohup java -jar: How to Safeguard Spring Boot in Production?
  27. ithome AirTag Exposes Amazon's AI Training 'Book Destruction' Tactic: Cutting Off Spines for Faster Scanning
  28. ithome Jensen Huang: OpenAI Commits to Deploying ~12GW of NVIDIA Compute by 2030
  29. hackernews Kimi Work attaches raw agent sessions to feedback reports
  30. hackernews Superconducting monolayer cuprate with a single CuO2 plane
  31. hackernews A True Telnet BBS on a Casio Calculator
  32. hackernews A SAT Attack on Tarski's High School Algebra Problem
  33. devto I Can Finally Delete My max-height: 9999px Animation Hack
  34. juejin DeepSeek Can't Recognize Images? Install an Open-Source Skill to Let Text-Only Models 'Open Their Eyes'
  35. hackernews How I developed an Am29000 C compiler and web browser
  36. hackernews Apple's App Tracking Transparency treated its own apps better than rivals
  37. hackernews How to ship a database every day
  38. hackernews How Go detects struct copies with sync.noCopy
  39. github_trending nautechsystems/nautilus_trader
  40. github_trending jundot/omlx
  41. devto Functional languages are heavily imperative.
  42. devto Your Dog Can't Tell You Where It Hurts. MATCH_RECOGNIZE in Snowflake Can.
  43. techcrunch Amazon, once an online bookseller, is destroying rare books to train AI models
  44. techcrunch Crypto hardware wallet owners face fresh security risks after recent spate of personal data thefts
  45. theverge Anthropic explains how Claude’s invisible text watermarks will work
  46. juejin Why AI Long-Form Novels Always Fall Apart at Chapter 30
  47. ithome Florida Man Describes 'Kill Ex-Girlfriend' Plan to ChatGPT; OpenAI Reports to Police
  48. ithome German Regulator Requires Apple to Adjust App Ad Data Consent Rules: Pop-Ups Must Be Neutral, Remove Dissuasive Slogans
  49. hackernews St Lucie Nuclear Reactor Unit 1 manually shutdown, 3 control rods drop into core
  50. devto Opus 5: How to Fix Verbose Output
  51. theverge ChatGPT’s Computer History tracks your clicks and keystrokes
  52. juejin The Potential of SVG Drawing in the AI Era
  53. hackernews Launch HN: Speko (YC S26) – OpenRouter for Voice AI
  54. hackernews Linear algebra done right
  55. github_trending harry0703/MoneyPrinterTurbo
  56. devto Learn how to build a disclosure component using the native HTML details tag
  57. devto Why the "AI" Badge Doesn't Matter and How to Restore Trust in Our Code
  58. techcrunch Groq raises $350M to fuel its pivot from AI chips to neocloud
  59. theverge Apple ordered to stop scaring iPhone and iPad users away from third-party apps
  60. theverge Whisker’s AI-powered litter robot thinks my cats swapped bodies
  61. juejin Don't Be Fooled by AI Anymore: I Turned Tencent Cloud's 4 Skills into an 'AI Anti-Fraud Detective' — One Small Config Pitfall Screwed Me
  62. juejin Dart 3.13 Major Update: It Feels More Exciting Than Flutter
  63. juejin SQL Server to KingbaseES Migration: Performance Test of a Complex BI Query
  64. ithome ByteDance Signs Agreement with the Motion Picture Association to Strengthen AI Model Copyright Protection
  65. ithome GitHub Platform Experiences Sudden Outage; Official Investigation Underway
  66. ithome PJM, Largest US Grid Operator, Plans Tough Action on Power Shortage: Data Centers Must Self-Generate
  67. hackernews A U.S. Strategy to Prevent the Creation of Mirror Life
  68. hackernews Chestnut – eGPU dock with open-source firmware
  69. hackernews AI Coding Without the Vibes
  70. ars_technica Wildfire smoke now bigger prenatal threat than human sources of air pollution
  71. hackernews The AI Credit Resale Economy
  72. hackernews Research papers using "kidney disappointment" instead of "kidney failure"
  73. hackernews What I Learned Securing Sniffnet with the GitHub Secure Open Source Fund
  74. devto I built a security scanner that checks if you are a dog
  75. devto Claude Impact Lab LA: Community Changed the Code
  76. juejin One Line of Code to Integrate DWG/DXF Drawing Viewing: Measurement & Annotation, Data Stays On-Site
  77. juejin I Built a 'Jushi' Mini Program; the Hardest Part Wasn't the Business, but the Fonts
  78. juejin Recommending an Awesome Enterprise Intelligent Recruitment System
  79. hackernews How to put 170 atoms in an atom
  80. hackernews Show HN: Sokoban AI Solver