How the AI Problems ranking works

Top AI Problems People Want Solved is a weekly list of real questions and pain points people post about AI. This page explains where the items come from, how they're picked and scored, and where the method falls short.

Sources

Hacker News, via the public Algolia HN Search API (hn.algolia.com). We pull stories, Ask HN posts and comment threads on AI-related posts. Points and comment counts come straight from the API.
X, via public post search and lookups by post ID. Likes and reply counts come straight from X.
Readers of this site can also submit a problem or upvote one through the form on the ranking page (see "Reader submissions and upvotes" below). At launch (October 24, 2026) no reader submissions or votes had been recorded, so every row at that point came from public HN and X posts.

Not covered right now: Reddit (our requests are blocked with HTTP 403), Product Hunt, Stack Exchange, Discord, GitHub issues and non-English communities.

How items get on the list

Daily scan. Once a day we scan the previous 24 hours, 08:00 to 08:00 Beijing time (the first scan, on October 21, 2026, covered October 19 to 21). On Hacker News we search AI-related keywords (e.g. AI, LLM, agent, Claude, ChatGPT, Codex, Cursor, Gemini, vibe coding) and also pull every Ask HN post and every story with 30+ points in the window. On X we run topic searches.

Screening. An AI research assistant reads the candidates and decides which ones to keep. It's a judgment call, not a fixed classifier. It keeps only posts where someone describes a problem they have with AI or want AI to solve. It drops:

news and reporting
product launches and Show HN posts
tool recommendations and promotional posts (e.g. tutorials with discount codes)
opinion pieces and general business discussion
job-hunting and hiring posts
posts outside the 24-hour window
repeats of a problem that's already on the list

One main post per problem. Each problem is tied to one main post, and only that post's numbers count. Other posts on the same theme are listed under "Related" on the row but don't add to its score, so the same complaint posted in many places isn't counted several times.

Plain-English summary. Each problem title is our one-line summary of the post, not a quote. Click through to read the original.

One more filter matters: the scan is looking for problems that don't already have a good solution. If a quick web search turns up mature tools for a pain point, the item is usually left out, even when it's popular. So the list leans toward unsolved problems, not toward everything people complain about.

Scoring

For each problem:

Heat = engagement on the main post. Hacker News: points + comments (for a single HN comment, which has no public points, the number of direct replies). X: likes + replies. Missing counts are treated as 0. Views, reposts and bookmarks are not scored.
Normalised heat (0–100): this week's hottest scraped problem gets 100 and every other problem is scaled against it, then rounded.
Score = normalised heat + 5 for each reader upvote on this site.

Every row shows both numbers separately, and the list is sorted by score. Engagement counts are captured once a day during the scan. If a post can't be fetched again (for example, it was deleted or made private), its row keeps the numbers from its last successful capture, and the row shows that older date. The top of the page shows when the counts were last captured.

One HN point is not the same as one X like. The two audiences are different sizes and behave differently, so read the order as a rough signal, not a precise measurement.

Weekly reset and updates

Weeks run Monday to Sunday, Beijing time, and the ranking starts fresh every Monday.
A problem shows up in the week it was first spotted.
New problems and fresh engagement counts are added once a day by the scan.
The site rebuilds automatically every 6 hours. A rebuild doesn't fetch new counts itself; it publishes whatever the latest scan recorded, so a scan's results can take up to 6 hours to appear.

Reader submissions and upvotes

The form on the ranking page lets readers submit a problem or upvote one. Here's how it works:

You need your full name (at least two words) and an email address. Your name is shown next to a problem you submit; your email never is.
Each email address gets one upvote per problem per week.
Each IP address can submit up to 5 problems and cast up to 30 votes per hour, and a hidden field filters out simple bots.
A submitted problem starts at heat 0, so it can only climb the list through upvotes.
Submissions and votes are stored on Cloudflare and count only for the week they were made in.

Every item links to its source

Every scraped row links to the original HN or X post it's based on, along with the numbers we scored, so you can check each number yourself. They'll have moved since we captured them. Reader submissions have no external link; they show the submitter's name instead.

Known limitations

Small sample. A week holds about a dozen problems, and the cutoff is a judgment call.
Skewed toward a few platforms. Only HN and X are covered, so the list reflects developers and tech-Twitter, not AI users in general.
Mostly English. The keywords we search with are English.
X search is incomplete. Results come back sorted by relevance, with a limited number per search (25 in recent scans), and we can't filter by likes. Popular posts can be missed.
Replies count as much as likes. Heated arguments can rank above posts that are widely liked but have few replies.
Engagement can be inflated. Bots, viral reposts and pile-ons inflate numbers, and we don't correct for that.
Supply checks are shallow. "Already solved" is decided from a limited web search, so a problem we think has no good solution may in fact have one.
AI-assisted screening. The AI that picks and summarises items can get one wrong. If a summary misrepresents a post or a link is broken, please tell us.

The current week's data is also available as machine-readable JSON. Live reader votes and submissions are at /api/problems.