The percentage of Show HN posts is increasing, but their scores are decreasing | snubiAstro
snubi
작업 글 링크 모음 목록으로 산돌고딕네오의 line-height 문제입니다. 헤더에도 똑같은 문제가 있어요.
그렇잖아요. align-items: center로는 해결이 안 돼요. 어쩌겠어요? 한 폰트
때문에 다른 폰트를 사용하는 사람들을 전부 불편하게 만들 수는 없잖아요(물론
산돌고딕네오만 그런 건 아니겠지만)? 물론 아이폰이나 아이패드의 기본 폰트인
건 알아요. 프리텐다드로 고정하는 게 가장 현실적인 해결 방안이겠지만 1.
고작 목록으로 돌아가기 버튼 하나 때문에 웹폰트를 불러오긴 싫고 2. 제가
뭐라고 사용자의 브라우저의 폰트 설정을 무시하겠어요. 참으세요. The percentage of Show HN posts is increasing, but their scores are decreasing 2026년 1월 13일 Last update: 2026-01-14
Recently, I felt like I was seeing more “Show HN” stories, and many of which were generated with LLMs. So I analyzed the data to see if that was true. Also I included the average score per month to see if people enjoy seeing them (because I don’t :P).
Charts
Stories in 2026 was omitted. 1) It’s only 13 days, 2) Scores are not stable yet.
Left axis: show_hn_ratio(show_hn / story * 100)
Right axis: average_show_hn_score and average_story_score

With LLM timeline

Analysis

Disclaimer: I am neither a data scientist nor a statistician. Some nuances may have been lost in translation.

Percentage
For about ten years (2012~2022), the percentage of Show HN stories was around 2-3%. Then, with the appearance of LLMs that can code, it’s been increasing. Claude Code and Cursor 1.0 accelerated it even more. As of December 2025, over 12% of all stories are Show HNs. It’s safe to say that there is a correlation between the increase in Show HN posts and LLM. People can create great things even if they don’t know how to code at all.
Scores
Show HN stories used to receive similar scores (around 15-18) to those of all stories until 2023~2024. However, it’s been declining while percentage of them are going up. As of December 2025, the average Show HN score is 10 points lower (9.04 vs 19.53).
Does it mean LLM-generated Show HNs are lower quality? I’m not sure. Maybe people are tired of seeing too many Show HNs.
Also I have no idea why the average score was increased in 2022. A lot of new users?
Data and codes
You can find python code and csv in https://github.com/plastic041/hackernews.
I exported BigQuery hacker news data to csv using this query:
SELECT
`time`,
`title`,
`type`,
`score`,
`id`
FROM
`bigquery-public-data.hacker_news.full`
WHERE
(`type` IN ('story')) and title IS NOT NULL;
The type field in BigQuery does not have a show_hn attribute like the Algolia API, so I lowercased titles and filtered using starts_with("show_hn: ") to determine if a post is a Show HN story.
I didn’t commit to the repo the original CSV because it was too big (~400 MB) but you can download it from BigQuery for free (I didn’t set billing account). I ran SQL above, exported it to google drive, and downloaded it.
I would like to analyze the percentage of Show HN stories generated with LLMs but I couldn’t find the way to do this, because many Show HN stories don’t mention that they’ve used LLMs in their text.
I’ll try to update this article every few months.

The document, authored by snubiAstro, presents an analysis of trends within Hacker News submissions, specifically focusing on "Show HN" posts, and their associated scores. The core investigation revolves around a noticeable upward trend in the percentage of Show HN posts produced over the past decade, correlating with the emergence of Large Language Models (LLMs) like Claude Code and Cursor 1.0. As of December 2025, approximately 12% of all Hacker News stories were identified as Show HN posts, suggesting a strong connection between the rise of LLMs and the proliferation of these submissions.

A key observation is the concurrent decline in the average score of Show HN posts. Initially, these posts received scores comparable to general Hacker News stories, ranging between 15 and 18. However, by December 2025, the average Show HN score had decreased to 9.04, representing a reduction of 9.53 points. This decline contrasts with the increasing volume of Show HN content.

The author, acknowledging a lack of formal data science or statistical expertise, clarifies that the analysis is a preliminary exploration. The data was derived from a query executed against BigQuery public data for Hacker News, which involved filtering titles based on the prefix "show_hn: " to identify Show HN posts. The creation of a sizable CSV file (approximately 400 MB) was facilitated using SQL queries and subsequent export to Google Drive. Acknowledging the difficulty in directly determining the quality of LLM-generated Show HN posts due to a lack of explicit mentions of LLM usage within the content itself, the author indicates an intention to periodically update the analysis. The investigation highlights a potential shift in the Hacker News landscape, driven by LLMs, with consequences for the engagement and perceived quality of Show HN submissions. The author’s approach—utilizing publicly available data and descriptive analysis—provides a foundational understanding of this evolving trend.