Hacker News Reader

Comment by splonk | Hacker News Reader

Related: there are various bots for HQ Trivia that do similar things that have come up on HN before. Here's one with a bit of a writeup: https://medium.com/@tobymellor/hq-trivia-using-bots-to-win-m...

My general impression is that being able to screenshot and use real computing power is a fair bit easier than using OCR on a Raspberry Pi, but question parsing and querying Google is a lot harder than using a static corpus of questions and answers.

View on HN

Even with a good bot (displays info instead of likely answer), HQ stops you from withdrawing after you win too many games, no fun :(

You can directly connect to their websocket and get the text of the answers faster than they are displayed on the app.

I've built a similar proof-of-concept HQ Trivia algorithm before[0], and from my simple tests, in most cases isn't any additional work needed, just naively send the OCRed question to Google. Of course, some questions don't work using this naive approach, which is why a similar HQ Trivia bot used riffed on this concept with unique (and clever) heuristics[1].

And yes, it is easier to work with a screenshot than a grainy image of text produced by a camera.

[0]: https://applecrazy.github.io/blog/posts/protect-trivia-from-... [1]: https://hackernoon.com/i-hacked-hq-trivia-but-heres-how-they...

Edit: my HQ Trivia system was a simple 1-hour weekend project, so it is by no means perfect. Also, to prevent my Github from being taken down, none of the code is available.

I'd previously done OCR work with Twitch streams, but the reliability of the Tesseract libraries for custom fonts had stopped me from going any further with it. Since HQ uses a standardized font library, I'd imagine character extraction would be more successful when recording from screenshots because it would likely be closer to the library's training datasets.

Counting donations and subs? I’d thought about doing the same.

You can train Tesseract on other fonts.

Mux wrote a blog post about this too: https://mux.com/blog/how-we-used-machine-learning-to-predict...

Using machine learning to parse text helps, but it's still not a perfect solution.

Reply on news.ycombinator.com