conv.

All stories
AIQuiet 9d · day 9

Developer uses Gemini to train cheaper replacement model for $9

Engineer labels 4,290 Reddit comments with Gemini to train GLiNER, replacing expensive API calls with a locally-run model.

What to know

  • Developer used Gemini API to label 4,290 comments in bulk for $9, then trained open-source GLiNER model on those labels.
  • The trained model reaches cost parity after ~4,291 comments and runs locally on a GPU, eliminating recurring API expenses.
  • Comments raise concerns about model validation (trained against Gemini, not ground truth), bot-generated training data, and whether the task could be solved with simpler approaches.

The dispute Whether the post's technical approach is sound and properly validated, given that the model is graded against Gemini's labels rather than ground truth and trained on potentially bot-corrupted Reddit data. · positions read across 9 posts and comments

many voices

This demonstrates the right way to use expensive LLMs—as one-time tools to bootstrap cheaper, task-specific systems.

  • “I think this is the way. An LLM is an expensive general purpose tool and for repeatable tasks, after it's clarified the process flow, it builds cheaper special purpose tools for each step”

    quirkot · Hacker News ↗
many voices

The approach has validation and data quality problems: model is graded against Gemini rather than ground truth, and Reddit training data may be polluted by bot comments.

  • “Unfortunately, advertisers are getting smarter and using bots to praise their own products on Reddit… this method will, at some point, not identify the company with the best knife, but the one with the most ad spend on bot comments.”

    faidit · Hacker News ↗
some voices

This is overthinking a simple problem; direct research methods are more reliable than scraping Reddit.

  • “I found it much more useful to go to a knife shop and handle a whole bunch of knives for myself. They're all pretty similar besides material, so not much signal you're going to be able to glean from people arguing on reddit.”

    foltik · Hacker News ↗

Peter Vijeh Developer/engineer

How it unfolded 1 development · click the chart to see its coverage articlespostscomments

Peak 11 pieces in 3h at Sep 17, 8 AM; 12 pieces over 9 days (1 article · 2 posts · 9 comments) Sep 17, 8 AM — 11 pieces · 1 article · 2 posts · 8 comments — Hacker News 9, Newswires 1, Mastodon 1Sep 17, 11 AM — 1 piece · 1 comment — Hacker News 1Sep 17, 2 PM — quietSep 17, 5 PM — quietSep 17, 8 PM — quietSep 17, 11 PM — quietSep 18, 2 AM — quietSep 18, 5 AM — quietSep 18, 8 AM — quietSep 18, 11 AM — quietSep 18, 2 PM — quietSep 18, 5 PM — quietSep 18, 8 PM — quietSep 18, 11 PM — quietSep 19, 2 AM — quietSep 19, 5 AM — quietSep 19, 8 AM — quietSep 19, 11 AM — quietSep 19, 2 PM — quietSep 19, 5 PM — quietSep 19, 8 PM — quietSep 19, 11 PM — quietSep 20, 2 AM — quietSep 20, 5 AM — quietSep 20, 8 AM — quietSep 20, 11 AM — quietSep 20, 2 PM — quietSep 20, 5 PM — quietSep 20, 8 PM — quietSep 20, 11 PM — quietSep 21, 2 AM — quietSep 21, 5 AM — quietSep 21, 8 AM — quietSep 21, 11 AM — quietSep 21, 2 PM — quietSep 21, 5 PM — quietSep 21, 8 PM — quietSep 21, 11 PM — quietSep 22, 2 AM — quietSep 22, 5 AM — quietSep 22, 8 AM — quietSep 22, 11 AM — quietSep 22, 2 PM — quietSep 22, 5 PM — quietSep 22, 8 PM — quietSep 22, 11 PM — quietSep 23, 2 AM — quietSep 23, 5 AM — quietSep 23, 8 AM — quietSep 23, 11 AM — quietSep 23, 2 PM — quietSep 23, 5 PM — quietSep 23, 8 PM — quietSep 23, 11 PM — quietSep 24, 2 AM — quietSep 24, 5 AM — quietSep 24, 8 AM — quietSep 24, 11 AM — quietSep 24, 2 PM — quietSep 24, 5 PM — quietSep 24, 8 PM — quietSep 24, 11 PM — quietYesterday, 2 AM — quietYesterday, 5 AM — quietYesterday, 8 AM — quietYesterday, 11 AM — quietYesterday, 2 PM — quietYesterday, 5 PM — quietYesterday, 8 PM — quietYesterday, 11 PM — quietToday, 2 AM — quietToday, 5 AM — quietToday, 8 AM — quietToday, 11 AM — quietToday, 2 PM — quiet 1
Sep 18Sep 19Sep 20Sep 21Sep 22Sep 23Sep 24yesterdaynow · 3:36 PM ET
  1. 1
    “Gemini labeled 4,290 comments for $9, or $0.0021 a comment. That means the trained model pays for itself at roughly comment 4,291, as long as later comments are about the same length and it runs on a GPU I already own.”
    — Peter Vijeh
    1. first by HN Frontpage, 9d ago

    • Unfortunately, advertisers are getting smarter and using bots to praise their own products on Reddit. Thanks to training on genuine comments, some models are very good at sounding like a human commenter, and can easily generate a comment history with diverse interests to appear human, making them basically undetectable. So it seems like this…

      faiditHacker News9d agoview on Hacker News ↗
    2 more of the top 3 · 9 posts in this stretch
    • The thinking people who would find this interesting and read this are probably more than capable of understanding this and critical enough to expect that. Conversation over the title is distraction of what's important. Just stick to keeping original source title and let people vote and down vote if they don't like it. That's what votes are for.

      LurkandCommentHacker News9d agoview on Hacker News ↗
    • Isn't it just learning to map specific words, from the "knife world", to the correct class? If so, a simple dictionary would fit. What I think is a better way to validate is to split train/validation by words used presented in NER classes (like, it should be able to find new brands never seen before). It is a interesting problem.

      bguberfainHacker News9d agoview on Hacker News ↗
    all of them →

What people are saying 6 voices from 1 site · best of 9 · verbatim

Still unanswered
  • How does the trained model perform on truly novel brands or steels it never saw in training?
  • What percentage of the training data actually contains bot-generated or sponsored content?