[Today's Pick] Must-Reads for Sep 15

Today's three posts are grounded in hands-on experience and parameters, not just floating concepts. They're great for people who want to actually implement things.

1. Build an AI Risk Radar: Easier than Scrolling News

yunyi breaks down AI risk signals into "who is saying it" and "is the market reacting," using Ray Dalio's 3,800-word article and SoftBank's 11.54% plunge as examples—very down-to-earth. He also tested that manually checking 20 items takes 40 minutes with 3 misrecorded entries. Clickbait titles often spin a dip in AI concept stocks as an "AI collapse." This radar approach is perfect for those who want to scroll less news.

2. Can DeepSeek Be Used After Starting Over?

ye_jiaxin's real-world test of DeepSeek V4.1 Flash has both sweet spots and pitfalls: keeping old call logs, throwing prompts in as-is, getting structured output from plain text in one go, and even reading tables from screenshots. The most painful part is "no buffer during switching"—an Agent chain didn't throw errors, but the downstream classifier couldn't connect, and occasional JSON explanations broke the parser. Worth a close read if you're planning a migration.

3. A ¥10k Embodied Experiment: Don't Rush to Call It Dirt Cheap

moyan doesn't jump on the "dirt cheap" bandwagon. Instead, he focuses on the Mantis Standard "beginner" model's specs: ¥9,800, 22 degrees of freedom, 7kg single arm, and teleoperation latency under 10ms, judging it more like a modular lab platform. He also drills down into embodied data issues: while a 7kg single arm is fine for grabbing cups or boxes, the real question is whether latency balloons when repeatedly disassembling and reassembling to run the same task, considering the stack of real network, vision, and control systems.