xai-studies 20260717T161813Z

GITHUB_README.md

@@ -31,7 +31,9 @@ make test
 python3 studies/run.py test
 ```
 
-Expect `✓ test passed`. Then open [START_HERE.md](START_HERE.md).
+Expect `✓ test passed`. Then [START_HERE.md](START_HERE.md).
+
+**Mirrors** (after `make sync`): GitHub · [Space](https://huggingface.co/spaces/lyffseba/xai) · [raw index](https://huggingface.co/spaces/lyffseba/xai/raw/main/index.html) (no CDN lag) · [dataset](https://huggingface.co/datasets/lyffseba/xai-studies).
 
 ## layout
 

catalog/SYNC.txt

@@ -0,0 +1 @@
+xai-sync 20260717T161813Z

studies/STUDY.md

@@ -1,55 +1,53 @@
-# How to study this monorepo
+# How to study
 
-**Offline:** open `../index.html` or read markdown under this tree. No network required.
-`make test` / `make check` use Python stdlib only. See [../OFFLINE.md](../OFFLINE.md).
+**Offline.** Open `../index.html` or read this tree. No network.  
+`make test` uses stdlib only. **xai** = explainable AI.
 
-Order is intentional: shared context → one model deep → the other → compare.
+Order: map → Inkling deep → Kimi K3 deep → compare.
 
-## 1. Map the yard (20 min)
+## 1. Map (20 min)
 
 1. [README.md](README.md)  
 2. [derived/DASHBOARD.md](derived/DASHBOARD.md)  
 3. [shared/GLOSSARY.md](shared/GLOSSARY.md)  
-4. [shared/comparisons/inkling-vs-kimi-k3.md](shared/comparisons/inkling-vs-kimi-k3.md) — skim only  
+4. [compare](shared/comparisons/inkling-vs-kimi-k3.md) — skim  
 
-## 2. Inkling deep (half day)
+## 2. Inkling (half day)
 
-Weights + code exist → best for **implementation truth**.
+Weights + code → implementation truth.
 
 1. [models/inkling/README.md](models/inkling/README.md)  
-2. [models/inkling/book/00-map.md](models/inkling/book/00-map.md)  
-3. [models/inkling/derived/SPEC.md](models/inkling/derived/SPEC.md) + [layer_schedule.md](models/inkling/derived/layer_schedule.md)  
-4. [models/inkling/book/01-architecture.md](models/inkling/book/01-architecture.md)  
-5. [models/inkling/code-study/01-modular-walkthrough.md](models/inkling/code-study/01-modular-walkthrough.md) with `reference/transformers-inkling/modular_inkling.py` open  
-6. Multimodal + chat: code-study 02–03  
-7. Training / benches / serve: book 02–04  
+2. [book/00-map.md](models/inkling/book/00-map.md)  
+3. [derived/SPEC.md](models/inkling/derived/SPEC.md) + [layer_schedule.md](models/inkling/derived/layer_schedule.md)  
+4. [book/01-architecture.md](models/inkling/book/01-architecture.md)  
+5. [code-study/01-modular-walkthrough.md](models/inkling/code-study/01-modular-walkthrough.md) + `reference/transformers-inkling/modular_inkling.py`  
+6. code-study 02–03 · book 02–04  
 
-## 3. Kimi K3 deep (half day)
+## 3. Kimi K3 (half day)
 
-API live, weights pending → best for **product + KDA theory**.
+API live, weights pending → product + KDA theory.
 
 1. [models/kimi-k3/README.md](models/kimi-k3/README.md)  
-2. [models/kimi-k3/book/00-map.md](models/kimi-k3/book/00-map.md)  
-3. [models/kimi-k3/derived/SPEC.md](models/kimi-k3/derived/SPEC.md) + [lineage.md](models/kimi-k3/derived/lineage.md)  
-4. [models/kimi-k3/book/01-architecture.md](models/kimi-k3/book/01-architecture.md)  
-5. [models/kimi-k3/code-study/02-kda-background.md](models/kimi-k3/code-study/02-kda-background.md) + arXiv:2510.26692  
-6. API protocol: [code-study/01-api-protocol.md](models/kimi-k3/code-study/01-api-protocol.md)  
-7. Capabilities + limitations: book 03–04  
-8. Open questions: [derived/open_questions.md](models/kimi-k3/derived/open_questions.md)  
+2. [book/00-map.md](models/kimi-k3/book/00-map.md)  
+3. [derived/SPEC.md](models/kimi-k3/derived/SPEC.md) + [lineage.md](models/kimi-k3/derived/lineage.md)  
+4. [book/01-architecture.md](models/kimi-k3/book/01-architecture.md)  
+5. [code-study/02-kda-background.md](models/kimi-k3/code-study/02-kda-background.md) + arXiv:2510.26692  
+6. [code-study/01-api-protocol.md](models/kimi-k3/code-study/01-api-protocol.md)  
+7. book 03–04 · [open_questions.md](models/kimi-k3/derived/open_questions.md)  
 
-## 4. Compare again (30 min)
+## 4. Compare (30 min)
 
-Re-read [inkling-vs-kimi-k3.md](shared/comparisons/inkling-vs-kimi-k3.md) and fill:
+Re-read [compare](shared/comparisons/inkling-vs-kimi-k3.md):
 
-- Which long-context bet do you trust more (relative+SWA vs KDA)?  
-- Which MoE story is clearer for production (41B active known vs 16/896 without active FLOPs)?  
-- What would you fine-tune first, and on which base?  
+- Which long-context bet (relative+SWA vs KDA)?  
+- Which MoE story for production (41B known vs 16/896 without active FLOPs)?  
+- What would you fine-tune first, on which base?  
 
-## 5. After K3 weights drop
+## 5. After K3 weights
 
-```bash
+```
 make refresh MODEL=kimi-k3
 make check
 ```
 
-Then extend `models/kimi-k3/code-study/` with a real implementation walkthrough (mirror Inkling).
+Then add a real code walkthrough under `models/kimi-k3/code-study/` (mirror Inkling).