[Today's Recommendation] Worth reading on 09/12

Today's three posts are a bit hardcore, not empty conceptual chatter, but laying out data, automotive standards, and packet capture logs to examine boundaries.

1. The valuation logic of shipyard AI isn't in order numbers

shen_zixuan fed AI with H1 2026 data showing Chinese shipyards captured 82.3% of global new orders, up 173.1% YoY. Surprisingly, the AI first flagged inconsistent statistical scopes; later, when mixing 90% of global orders with 98% of mid-to-high-end orders into the same trend, it got slapped by industry structure realities. Good for seeing how AI due diligence slides from numbers to misinterpretation.

2. A model refusing to answer doesn't mean the cockpit is safe

wan_guanhai starts from Anthropic identifying ~35 potential research behaviors in 30 days, arguing that model-side refusal and vehicle-side accident avoidance are not the same thing—a very accurate analogy. He breaks down WEY V9X into low, medium, and high risks, noting especially that high-risk actions require confirmation, but confirmation shouldn't be designed as continuous questioning. Cockpit products can copy this homework directly.

3. Agent packet capture: Don't let it decide to stop on its own

lingxi connected an agent to RubyGems for read-only scraping. In the first attempt, requests jumped from dozens to hundreds in ten-plus seconds, hitting a bunch of 429 errors, exposing its lack of self-restraint awareness. Later, limiting to GET, prohibiting registration, fixing intervals, and using local cache kept request volume down to dozens, stopping on failure. This reminds us that agents aren't bad; the bad part is having no boundaries.