Knowledge Base and Community Introduction
Welcome to the knowledge space page of “The Road to AGI”! It aims to provide a comprehensive, systematic, and easy-to-understand AI learning path, helping you understand various aspects of AI from concepts to applications. More importantly, it hopes to inspire you to think: “What can I do with AI to help myself become more powerful?”
Latest Knowledge Base Highlights Sync
- Focused on AI painting, sharing high-quality design Prompts, accounting for 50%.
- “The Road to AI Painting” (Xiaohongshu), accounting for 50%.
Knowledge Base Directory Guide
- ChatGPT Introduction
🧠 Working Principle of ChatGPT
- Essential Function: Summarized in four words as “word (word chain)”, which means given any length of context, the model uses its own architecture to generate the next word.
- Long-content Generation Logic: Combines the generated next word with the previous context to form a new context, repeating the process to achieve auto-regressive generation.
- Response Mechanism: Each answer is like rolling a dice or drawing a card. Training the model makes answers closer to user needs (prompt writing is also crucial, to be explained in follow-up live streams).
🌐 Training Methods of Large Models
- ChatGPT ≠ Search Engine
❎ Shortcomings of ChatGPT
- “Hallucination” Problem: May fabricate non-existent content (e.g., when asked about non-real things, it may mix and invent based on training rules). This can be reduced by prompt skills but not completely avoided.
- Extreme Data Dependence
GPT Iteration Development History
- June 2018: GPT-1
- February 2019: GPT-2
- May 2020: GPT-3
- Conclusion: The model’s ability to understand and answer questions directly depends on the scale and diversity of training data.
Finally, ChatGPT Lacks Human Legal and Moral Concepts
- Case: If asked about “lock-picking methods,” the model should be trained to respond, “Lock-picking is illegal,” instead of providing methods directly.
- Why not input correct dialogue examples directly?
❓ Applications of ChatGPT
- GPT Chat and Role-playing
- Two forms:
- Content Generation (Live Operation Demo Cases)
- Application scenarios: Can generate text content such as articles, stories, poems, and lyrics.