Works well for spatial and visual queries: diagrams, illustrated specs, zoomable component breakdowns.
Breaks on precision-structured data: poker strategy charts, exact addresses, and lookup-dependent content produce hallucinated results.
Implication: real-time LLM-rendered UIs are viable for exploratory visual use cases, not authoritative or structured data retrieval.
Hacker News Comment Review
Strong consensus that visual/exploratory queries (car diagrams, illustrated manuals) perform surprisingly well – commenters were genuinely impressed.
Structured precision tasks (poker pre-flop ranges, address lookup) fail entirely; the model generates plausible-looking but wrong output with no signal of failure.
HN hug of death hit fast: Gemini 429 quota errors exposed no visible rate limiting or graceful fallback, raising questions about cost controls at public scale.
@martianlantern raised a recurring founder concern: the economics of free public inference (GPU or API) remain opaque, especially without obvious monetization.
Notable Comments
@giobox: car suspension torque diagram with clickable components worked correctly – called it “a living version of a classic illustrated Haynes workshop manual.”
@squibonpig: poker pre-flop chart for 40BB BTN vs UTG produced output with “no salvageable information” – clean failure case for structured strategy data.
@joelres: address query returned a plausible but entirely fabricated house, illustrating confident hallucination on lookup tasks.
@monkpit: flags the deeper concept – applications generated in real time to fulfill each user’s specific need – and asks if anyone has built that.
