X.com links are currently broken on HN so I posted a screenshot instead

https://x.com/janleike/status/1791498187313963308

The funny thing is I got better OCR throwing this at ChatGPT than at the actual OCR engine I use (Windows default.)

Makes sense since Windows is probably still just using the “latest” Convolutional Neural Network type algorithms for their OCR from 10+ years ago :-)

Posted a thread as a PNG! This is novel, but yeah, necessary unfortunately.

No. It will get offered to you once you exceed the monthly GPT-4o limit (currently 5 to avoid breaking the bank).

GPT-4o outperforms all local models so I figured using it as a fallback was the right approach :)

I pasted https://x.com/amanrsanger/status/1790947733899203027 as the URL but it appears to get rewritten to https://twitter.com/x/migrate?tok=eyJlIjoiL2FtYW5yc2FuZ2VyL3... (presumably due to twitter.com migrating to x.com), and the resulting URL is broken?

Thank you!

Thanks for the feedback! That feature is there, just not discoverable enough!

Try clicking on the relative time (e.g. "over 1 year ago"). Should take you immediately to that comment on HN, and you can then hit back in your browser to return to Hacker Search.

Agreed on the extension. That's coming next!

I’m so glad you like it! Thank you for taking the time to write this kind note.

Hey! This leverages the official HN API (https://github.com/HackerNews/API), no scraping involved. I don't think it's my place to opine on "who owns the posts".

Thanks, the site is blackholed here at the office.

Yeah, that'd be a really nice thing to have! My current (and naive) approach is to just limit to the last 3 years of data. That said, it should be easy to add a date filter as a first step, and potentially have a more ingenious approach for letting the LLM reconcile contradictory statements based on the time at which they were posted.

I feel that it performs really well on queries like those on the landing page, that generally have to do with understanding HN's sentiment about something or finding resources to learn about a topic.

As to your second question, someone looked up “What are some famous Google office pranks?” (https://hackersearch.net/ask?q=What%20are%20some%20famous%20...?) earlier today and I found some gold in there. Some of the hot takes on Gary Marcus have cracked me up too (https://hackersearch.net/ask?q=what%20do%20you%20think%20of%...?).

I got some feedback that GPT-3.5 was letting people down so improved my caching strategy and defaulted everything to GPT-4o!

Great to hear! Would you be willing to share a bit on what changes you made to your caching strategy to optimize costs?

I made the caching window substantially longer (2 days instead of a few minutes) upon noticing most queries weren't about current events, and fixed a bug in the cache resolution logic that'd lead to a substantial number of cache misses.

Cool, thanks!