this repo has no description
hauleth.dev
1+++
2title = "Jeopardy! world"
3date = 2025-07-28
4
5[taxonomies]
6tags = ["ai", "culture"]
7+++
8
9Some time ago, there was an anime available on Netflix — *Godzilla Singular
10Point*. It wasn't a spectacular success, but it featured a plot device that I
11think reflects something increasingly common today: you need to know the answer
12to your question before you can ask it.
13
14This is something I see all the time in the current wave of AI hype. You need to
15know what the answer *should* be before you can write a useful prompt.
16
17<!-- more -->
18
19The issue I have with many AI use cases is this: unless you have specialized
20knowledge about the topic you're asking about, you can't reliably tell the
21difference between a solid AI answer and complete nonsense.
22
23I've had a few discussions about this on various Discord servers. The example I
24often use is this simple question posed to an AI:
25
26> Does 6 character long identification number, that contains digits and upper
27> case letters (with exception to 0, O, 1, I, and L) is enough to randomly
28> assign unique identification numbers for 10 million records?
29
30You can see for your self answer from ChatGPT [there][chatgpt].
31
32At first glance, the answer looks valid and sensible. The math checks out. It
33calculates the number of available combinations correctly. Everything seems
34*fine*.
35
36**BUT…**
37
38There is huge issue there, and probably most of the people who have been working
39with basic statistic or cryptography will notice it. ChatGPT (and any other AI
40that I have tested out) fail to notice very important word there
41
42> \[…] randomly \[…]
43
44This single word invalidates the entire reasoning, despite the correct
45calculations. Because of the [birthday problem][], the answer isn't feasible.
46While it's technically possible to assign a unique ID to every record, doing so
47randomly introduces a high probability of collisions.
48
49- At around 35,000 generated IDs, there's already a 50% chance of a collision
50- At around 90,000, the chance of at least one duplicate reaches 99%
51
52So even though the math is correct, the logic fails under the randomness constraint.
53
54## *Jeopardy!* world
55
56This is my main issue with AI tools: if you already have knowledge about the
57subject, you don’t really need to ask the AI. But if you don’t have that
58knowledge, you have no reliable way of knowing whether the answer makes sense or
59not. It’s like playing *Jeopardy!* — you need to know the answer before you can
60phrase the right question.
61
62In my view, AI is most useful in areas where the results can be quickly reviewed
63and discarded if needed. That’s why the whole “vibe coding” (aka slop
64generation) approach falls short. If you don’t have a good sense of what the
65output should look like, you probably don’t have the expertise to verify it.
66
67[And gods forbid you from allowing AI to do anything on production][replit-fuckup].
68
69[chatgpt]: https://chatgpt.com/share/68879fe7-d4e0-8007-9a30-3a9e2ace791d
70[birthday problem]: https://en.wikipedia.org/wiki/Birthday_problem
71[replit-fuckup]: https://www.businessinsider.com/replit-ceo-apologizes-ai-coding-tool-delete-company-database-2025-7?op=1