Physix Frontier[Today's Picks] 09/09 Worth a Read
All three posts today are pretty solid: one discusses how AI menu images shouldn't destroy trust, another tests pool/lawn mowing robots in an actual factory area, and the third teaches that for reports, you should preserve evidence before talking about automation.
1. Why Burgers on Menus Look Less and Less Like Real Burgers
The most interesting part of shen_minghui's post is using real burgers and fries from a small St. Louis shop to run AI image generation. It took about forty minutes to shoot with a phone, crop the images, and adjust brightness. The result? The model's first version had sesame seeds that looked stuck on, fries all identical in thickness, and ketchup reflections so uniform they looked like polished plastic. He went through three rounds of edits, adding prompts like "no over-editing," "real oil sheen," "tissue paper on the table," and "slight steam" before getting one barely acceptable image. You'll know exactly whether it saves time or looks fake after reading this.
2. Pools and Lawns: Don't Just Watch Demos for Cleaning Robots Going Overseas
luo_xiujie didn't stop at booth demos; he directly borrowed the Roborock RockAqua P1 and RockNeo Q2 to run each robot once in a client's factory pond and lawn. The pool robot paused for a few seconds under tree shadows and water reflections, and its first return-to-charge attempt missed the dock, requiring two retries to align. The lawn mower did a full lap in about ten-plus minutes, and obstacle avoidance around trees was more stable than expected, but drawing too many restricted zones fragmented the path. This kind of honest talk—"updating maps is more annoying than building them"—is valuable.
ming_996 places WorkBuddy in a report quality-check role, immediately advising "don't connect Feishu yet." Set up local raw/output/audit folders, don't auto-fill null values with 0 but mark them uniformly as NA, and use Order ID + Customer ID as the primary key for duplicate checks. The most valuable tip is to avoid checking "auto-summarize" or "generate report" initially. Click "dry run" to see the affected row count and high-risk fields before proceeding. This serves as a wake-up call for anyone dazzled by "three-minute report" tutorials.
