ChatGPT 4o gave wrong answers about Indian visa requirements. Not some answers. Most of them. The author found this out after spending an hour fighting a buggy government visa portal, then asking the model the same questions he had just answered himself.
This is not a hallucination story. It is a product design story. LLMs are probabilistic, not deterministic. They are not databases. They generate text that looks like a correct answer, which is useful in some contexts and dangerous in others. The real question the author is building toward is not whether models will improve, it is how you build mass-market products around systems that will keep getting things wrong at an unpredictable rate. RAG and multi-agent architectures are mentioned and then dismissed as incomplete solutions. The AGI framing gets the same treatment.
The piece lands on a 50-year lesson from consumer computing: you do not drive adoption by teaching users prompt engineering, any more than WordPerfect drove adoption with cardboard keyboard overlays. The original is worth reading for how it maps specific failure modes to specific product decisions, not just the conclusion that AI products need better design.
[READ ORIGINAL →]