Ning (@yangning9371) profile picture
Joined January 2026
Photo by Ning (@yangning9371)Photo by Ning (@yangning9371)Photo by Ning (@yangning9371)Photo by Ning (@yangning9371)Photo by Ning (@yangning9371)
十一放假回老家,感觉什么AI 大模型 OPC chatgpt 好像都不存在,世界还是原来的世界,每天吃吃喝喝 溜溜达达。
1
7
I built sift because keeping up with the internet started feeling like a second job. Choose what you care about. sift reads the feeds, threads, newsletters, and search results — then turns them into one weekly issue you can actually finish. sift.day
2
1
7
The goal isn’t to help people consume more. It’s to replace the endless feed with something finite: one issue, linked sources, and a clear ending. You can read a real sample before downloading: sift.day
6
Shipped 5 MCP Skills for GroundAPI this week — one config line in Claude/Cursor, and your AI gets real-world data. Thread 🧵👇 #buildinpublic #MCP #indiehackers
1
1
31
🌤️ Context-Aware Assistant Auto-detect location via IP, real-time weather + 7-day forecast, package tracking. 🌐 Web Researcher Multi-angle search, full-page scraping, structured reports with sources.
1
17
Transformer 统治了 8 年,但它在超长上下文和持续学习上有根本性限制。 三个挑战者已经上桌 👇
1
21
· Mamba/SSM:线性复杂度替代自注意力,推理成本断崖下降 · World Model/JEPA:不预测 token,在抽象空间预测表征 · Neuro-symbolic:神经网络 + 符号推理端到端联合 未来不是一种架构通吃,而是多架构各司其职。Transformer 不会消失,但下一波机会来自这些新架构。
35
80 亿参数的蒸馏模型,数学推理打赢了 Gemini 2.5 Flash。6 亿参数的 AMD ReasonLite,也能跑出有意义的推理。 推理能力是训练方法的函数,不是参数量的函数。
1
32
核心公式:强模型蒸馏 CoT + RL 后训练 = 小模型推理爆发 有人管这叫 Densing Law——未来不比谁大,比谁密。 别迷信大模型了,8B 经过正确蒸馏,特定任务完全能打。
41
Pre-training scaling 见顶了,这不是唱衰,是物理现实。 数据墙:互联网高质量文本快被吃完了 算力墙:先进制程芯片物理上造不够 但行业没停,4 条新路正在打开 👇
1
15
① Post-training RL(DeepSeek-R1 路线) ② Test-time Compute(o1/o3 路线) ③ Self-play 自我进化 ④ Agentic System 竞争从「谁的模型更大」变成「谁更会用智能」。 对开发者是好消息——你不需要训大模型,你需要更聪明地用它。
44
Every SaaS will eventually ship a CLI. Not for devs — for AI agents. Feishu just open-sourced theirs. Messages, docs, spreadsheets, approvals — all callable by Claude Code in one command. GUI for humans. CLI for agents. Both will coexist. #buildinpublic #AI
16
Feishu (Lark) just open-sourced their entire product as a CLI tool. This is huge. Your AI Agent can now control Feishu directly — messages, docs, spreadsheets, calendar, approvals. All via command line. Thread on why this matters 🧵 #buildinpublic #AI
1
31
The smartest part: Feishu's CLI returns human-readable error messages. Not just "404" or "502", but: which parameter failed, what went wrong, what command to run to fix it. This lets AI self-retry and self-correct. Every API should do this.
1
11