Source commit: b02deb5
Have you ever had this experience: while learning about Transformers, you suddenly encounter an attention mechanism and have to jump away to search what attention is; while learning BERT, you hit "Encoder-Only" and have to look up what an Encoder is. Knowledge should be interconnected, yet it is often fragmented across isolated documents.
With commit b02deb5, Easy AI did something that looks simple but has far-reaching impact: it built a cross-linking network among 9 existing knowledge sites.
What is cross-linking?
Each knowledge site is no longer an island but a node in a network. When you are learning BERT, the site now tells you directly:
> It's smoother to read the "Transformer" site first. BERT uses the left half of the Transformer (the encoder, which only reads and never writes) — all these parts are taken from the Transformer.
And it attaches a button that takes you straight to the Transformer knowledge site.
When you read "671B parameters / 37B activated" on the DeepSeek R1 page, it explains:
> 671B is the full set of weights, but only 37B is actually used each time a token is generated (thanks to MoE, which wakes only a subset of experts each time — see "MoE" for details).
Again with a jump button.
Knowledge sites covered by this upgrade
The upgrade involves 9 existing knowledge sites, each deepened and cross-linked:
- BERT: added Encoder-Only concept explanations and a new Transformer link card
- DeepSeek R1: strengthened explanations of reinforcement-learning-based training; new links to LLM/MoE/RLHF
- GPT: added Few-Shot and Chain-of-Thought concepts
- Llama: detailed the innovations of each version and strengthened license explanations
- MoE: improved visual explanations of expert routing and load balancing
- Multimodal: enhanced explanations of fusion mechanisms
- T5: added details on the Encoder-Decoder pipeline
- Transformer: greatly expanded sections on Attention, QKV, and modern applications
- NLP: improved explanations of tokenizers and pipeline differences
- Hands-on experience with Claude Opus 4.8
- OpenAI gpt-5.5 instant update
- Step 3.7 Flash release
- Launch of Google Gemini Managed Agents
- Plus progress on model quantization, deployment, and tooling
Cross-linking is not adding links — it's building a road network
If you just add a few hyperlinks, that isn't cross-linking. Easy AI goes deeper:
1. Insert link cards at key concepts: not a small line of text, but a complete card with an icon and explanation that tells you why to jump over and what to look at there.
2. Explain the meaning of cross-site references: instead of dropping a term for you to look up yourself, it explains the term in context on the spot and notes where the detailed explanation lives.
3. Unified visual style: all link cards share a consistent butter-colored background, rounded corners, shadows, and icons, so users instantly recognize a gateway for further learning.
Why this matters
Learning is never linear. While learning A you encounter B; learning B reminds you of C. A good knowledge system should embrace this jumping nature rather than fight it.
The cross-linking network turns Easy AI from a knowledge exhibition hall into a map for exploring knowledge. You don't need to know the learning path in advance — you can start from any point of interest and wander naturally between related concepts by following links.
This follows the same philosophy as Wikipedia, but is better suited to AI learning: Wikipedia's links tend to be overly academic, while Easy AI's link cards explain concept connections in plain language, so readers without background can understand *why* they need to read that next thing.
The other side of the update: daily news data
Beyond cross-linking, this commit also added AI daily-news data for 2026-05-30, covering:
> Visit Easy AI to experience the cross-linked knowledge sites: https://mmh1.top
---
> This article is an interpretation of Easy AI project commit b02deb5.