[Today's Picks] Worth Reading on 09/07

Today's three posts don't read like press releases; they feel more like retrospectives from seasoned operators and entrepreneurs: Don't just watch the IPO bell for GPUs, don't rely on clichés for JDs, and don't fixate solely on accuracy rates for breeding robots.

1. GPU IPOs are buzzing, but startups should focus on cash flow

han_yiming breaks down the hype cleanly: Biren rose 80% on its first day with a market cap of HK$84.6 billion, but H1 revenue was only RMB 58.9 million with a loss of RMB 1.6 billion; MetaX opened up over 568%, breaking a RMB 300 billion market cap, with reports saying one lot could earn RMB 300k. His real warning to startups is that buying cards requires calculating depreciation, electricity, O&M, fault compensation, and customer payments—don't mistake applause for funding as orders for business.

2. I used WorkBuddy to turn a single line from the business side into 5 review-ready JDs

yizhilu's post is very practical. The business side said, "Hire an AI Product Manager who understands LLMs," and he used WorkBuddy to break it down into a requirements gap list. A single Agent running for a week would produce resume templates full of clichés. He switched to an expert team consisting of HRBP, business interviewer, compliance reviewer, and channel copywriter, then banned words like "excellent," "responsible," and "stress-resistant." This JD production line can pass reviews.

3. Breeding robots: Don't rush to tell stories, look at the pollination math first

wei_yuanzhi pulls embodied intelligence back from launch events to the fields: Improving recognition accuracy from 85% to 92.8% is nice, but what's more critical is moving from flexible arms to dual arms, not damaging stamens, and who backs up when things go wrong. He suggests not selling the hardware itself, but charging based on per greenhouse, per season, number of plants, number of flowers, and successful pollinations. He also states plainly that it's not suitable for token issuance in the short term; what's truly valuable are stamen images, phenotypic sequences, environmental conditions, and failure samples.