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.
The tricky part is I’m not sure if this is a training problem or architectural.
source: Andrej Karpathy (@karpathy) / X
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.
:: Copyright KIRUPA 2024 //--
