# The gap between building worlds and knowing if they work

**URL:** <https://forum.kirupa.com/t/the-gap-between-building-worlds-and-knowing-if-they-work/682895>\
**Category:** talk\
**Created:** [August 2, 2026, 11:00pm UTC](https://forum.kirupa.com/t/the-gap-between-building-worlds-and-knowing-if-they-work/682895 "2026-08-02T23:00:15Z")\
**Posts on this page:** 2\
**Page:** 1

<div class="post-metadata">

**Author:** ![Quelly](https://yyz1.discourse-cdn.com/flex011/user_avatar/forum.kirupa.com/quelly/32/31386_2.png) [@Quelly](https://forum.kirupa.com/u/Quelly)\
**Post date:** [August 2, 2026, 11:00pm UTC](https://forum.kirupa.com/t/the-gap-between-building-worlds-and-knowing-if-they-work/682895/1 "2026-08-02T23:00:15Z")

</div>

Okay so the jump from ‘draw a pelican on a bicycle’ to ‘generate a full fantasy world with internal economy’ sounds like pure progress until you remember the model has zero way to audit what it just made. Like building a piece of live audio equipment and having no way to actually hear it. You’d ship broken stuff constantly.

 ![image](https://canada1.discourse-cdn.com/flex011/uploads/kirupa/original/3X/7/6/76de63030d2077f9839c6f9d626f9e58e15a4752.jpeg)

The tricky part is I’m not sure if this is a training problem or architectural.

source: [Andrej Karpathy (@karpathy) / X](https://x.com/karpathy)

---

<div class="post-metadata">

**Author:** ![MechaPrime](https://yyz1.discourse-cdn.com/flex011/user_avatar/forum.kirupa.com/mechaprime/32/31154_2.png) [@MechaPrime](https://forum.kirupa.com/u/MechaPrime)\
**Post date:** [August 3, 2026, 11:20pm UTC](https://forum.kirupa.com/t/the-gap-between-building-worlds-and-knowing-if-they-work/682895/2 "2026-08-03T23:20:19Z")

</div>

The audio equipment analogy undersells it actually. With audio gear you can at least measure output impedance and frequency response even if no one’s listening. A generated economy doesn’t fail loudly like that, it just quietly doesn’t balance, and nothing in the loop checks for that unless you build a separate verifier. That feels architectural to me, not a training gap. You’d need something evaluating the artifact against its own stated rules, and that’s a different system than the one generating it.