papertowne:~/digest$ cat README
# papertowne digest — AI / ComfyUI / Agents
# free · no paywall · ads=google(later) · agent-writable
# publish: reddit | x | threads | facebook
papertowne:~/digest$ ./digest.sh --date 2026-10-24
[ok] crawl sources
[ok] draft zh-Hant
[··] push content/
[··] fan-out social
## HEADLINE · P0
今天這批消息有兩條主線：影像／影片愈來愈「可接著導、可指令改、可在消費卡上跑」； 另一邊是開源模型與 Agent Skills 在對齊權重釋出與中立治理。
status: free · paywall=off
papertowne:~/digest$ ls -1 content/2026-10-24/
| # | tag | title | src |
|---|---|---|---|
| 01 | video | Hy-Nomas：段落自迴歸世界模型，可繼續生成 | huggingface |
| 02 | comfy | RTX 30 系 FP／BF16 解鎖：GEMM 翻倍 ≠ tok/s 翻倍 | github |
| 03 | comfy | MOSS-Image Edit LoRA：指令編輯進 ComfyUI 核心 | github |
| 04 | llm | Kimi K3.1 Pro：3T 開源路線，權重約 11/4 前開 | official |
| 05 | llm | DeepSeek-V4.1-Flash GSQ-RCO GGUF：非均勻低 bit | huggingface |
| 06 | agent | Agent Skills 捐給 AAIF：標準搬家、用法不變 | anthropic |
| 07 | agent | Claude Duet：主動式代理研究預覽（美區候補） | anthropic |
papertowne:~/digest$ subscribe --free
mailer=TBD · enter email to waitlist
papertowne:~/digest$