CloudyWing's Note

English

繁體中文

English

繁體中文

Theme

Original note

A Brief Discussion on Mainstream AI Service Ecosystems and Related Tools ​

TLDR ​

  • Developer Ecosystem (API): Data is not used for model training by default; Consumer Ecosystem (Web): Data may be used for training by default and must be manually disabled.
  • Claude is suitable for coding and logical reasoning, but note that the "entire conversation history" counts toward the Context Window; the longer the conversation, the faster the quota is consumed.
  • Google AI ecosystem is complex; it is recommended to utilize different carriers (App, NotebookLM, Gemini CLI) to distribute quota usage.
  • NotebookLM: If you encounter the 15-page limit when generating PPTs, you can bypass it by "generating in segments and merging files"; watermarks and image-based text can be handled via third-party tools (e.g., DeckEdit).
  • Gemini models, when handling long-form tasks, tend to miss details during iterative cycles if the scope is not explicitly defined; it is recommended to clearly limit the scope of modifications.
  • Agent Editors (e.g., Antigravity, Copilot): When handling complex tasks, avoid over-reliance on automated scripts and maintain basic logical judgment.
  • WSL Environment: When using the Copilot CLI, it relies on PowerShell; it is recommended to prioritize VS Code to avoid disrupting automated workflows.

Mainstream AI Service Ecosystems and Billing Architectures ​

Currently, mainstream AI services are divided into three categories, each with different billing and privacy policies:

  • Developer Ecosystem: Primarily based on API Keys, billed by usage; data is not used for training purposes by default.
  • Enterprise and Team Ecosystem: Provides SLA guarantees, with contracts explicitly prohibiting the use of data for training.
  • Consumer Ecosystem: Divided into free and subscription tiers; Web conversation data may be used for training by default and must be disabled in settings.

OpenAI Ecosystem ​

Centered around ChatGPT, featuring a powerful ecosystem.

  • ChatGPT (Web/App): Built-in Canvas editing window and custom GPTs.
  • Billing Mechanism: API usage and web subscription billing are completely separate.

Anthropic Ecosystem ​

Focused on coding and logical reasoning, with no image generation capabilities.

  • Token Limits: Claude counts the "entire conversation history" toward the Context Window during responses; the longer the conversation, the faster the quota is consumed.
  • Claude Code: Designed specifically for developers, supporting local codebase reading and MCP integration.
  • Artifacts: Solves the issues of formatting errors and unnecessary filler text when copying code from traditional chat interfaces.

Google Ecosystem ​

Features are complex; Gemini Advanced is available via Google One subscription.

  • Google AI Studio: Provides APIs for developers; data in the free tier is used for model training by default and is only disabled after upgrading to the paid tier.
  • Google One and Gemini Advanced: Google One is the name of the subscription plan, while Gemini Advanced is the unlocked advanced service.
  • Family Sharing Limits: Cloud storage and AI credits are shared by the entire family; individual member usage cannot be restricted separately.

Google Tools: Practice and Limitations ​

NotebookLM Usage Tips ​

When to use: When you need to convert long documents into presentations or summaries.

  • Bypassing the 15-page limit: First, ask the AI to generate a complete outline, split the transcript into multiple data sources, generate in batches, and then manually merge the .pptx files.
  • Removing watermarks and image-based text: If the exported file contains non-editable image text, you can use DeckEdit to restore it to an editable format.
  • Source synchronization: Google Drive files do not sync automatically; you must manually click the source to trigger an update.

Observations on Gemini Model Behavior ​

When to use: When performing complex tasks or proofreading.

  • Output Integrity: During multiple iterations, Gemini tends to miss details from earlier segments. Recommendation: Explicitly restate key conditions before regenerating, or limit the scope of modifications.
  • Proofreading Style: Gemini tends to over-embellish or over-simplify; it is recommended to explicitly request that the original meaning and tone be preserved.
  • Web Search: If a website has not been indexed (e.g., a newly launched GitHub Pages site), Gemini may hallucinate. Recommendation: Ensure sitemap.xml has been submitted to Google Search Console.

AI Agent Editors and Development Workflows ​

Tool Positioning and Selection ​

  • Antigravity: Suitable for tasks requiring back-and-forth discussion and architectural planning.
  • GitHub Copilot: Suitable for assigned, well-defined implementation tasks; billed by usage, offering higher efficiency.
  • Gemini CLI: Suitable for simple, repetitive background tasks.

Gemini Model Script Batch Processing Issues

Gemini tends to use scripts to handle tasks; if requirements are not clear, it can easily lead to project corruption. Even if the model guarantees the script is rigorous, human review of the logic is still required.

WSL Environment Limitations

When using the Copilot CLI in a WSL environment, it relies on PowerShell. If the environment does not have it installed, the Agent will be unable to execute scripts automatically. It is recommended to prioritize the VS Code extension to maintain the continuity of automated workflows.

Context Management Recommendations ​

  • Create Rule Files: Always create rule files such as AGENTS.md to prevent the Agent from repeating mistakes.
  • Workspace Isolation: Most Agent editors now restrict cross-workspace context reading; ensure that necessary files are loaded in the current window.
  • Conversation Length: Conversations that are too long will cause the model to hang; you should start a new conversation at appropriate intervals and keep the transmitted information concise.

Change Log ​

  • 2026-03-07 Initial document creation.
  • 2026-03-13
    • Added information regarding Google's adjustment of the Antigravity quota mechanism on 2026-03-12.
    • Added notes on Gemini's unstable output integrity and recommendations.
    • Added notes on Copilot CLI limitations in WSL environments and rule file reading issues.