Built in a few evenings out of mild frustration. Kept around because the frustration didn't go away.
I got tired of explaining the same thing in every other thread: large language models are not random number generators. They are not calculators either. They are not a database. They are not a search engine, even when wrapped in one.
“Pick a random number from 1 to 100” is the cleanest possible demonstration. The prompt is trivial, the failure is loud, and the bias is visible to anyone who can read a histogram. You don't need a benchmark suite or a research grant to see it — just a hundred API calls and a bar chart.
So instead of writing the same paragraph for the hundredth time, I built the site. Now I can just send the link.
The point isn't that AI picks 47. The point is that the same shape of mistake — asking a plausibility engine to do work that needs entropy, or arithmetic, or ground truth — is shipping in production right now. Token generators built on LLMs. A/B splits decided by “the model.” Password suggestions. Security-adjacent flows where “random” matters.
The model still generates output, so the bug doesn't show up in tests. It just produces the wrong distribution — quietly, every time. Pick the right tool for the right job; an LLM is the wrong tool for any job whose correctness depends on entropy.
I write about security, cloud, and the gap between documentation and production reality on gohack.in. The blog goes deep when I have something worth saying; LinkedIn is where the smaller, more frequent security & privacy notes land.
If this site changed how you think about LLMs, or if it didn't and you want to argue, the easiest channel is LinkedIn.
Next.js on Vercel, Postgres for the sample bag, OpenRouter for the model calls. The randomness score is chi-square against a uniform 1–100. The methodology page has the long version, including every caveat I could think of.
Disagree with a number, a method, or a take? The LinkedIn DM is open. The goal is to be wrong where I'm wrong and right where I'm right, in public.