Startup Opportunities from Hacker News

HN produced 3,991,507 comments between 28 September 2025 and 28 September 2026.

A cheap classifier, TypeSafe Jev, read a probability sample of 605,125 of them. It asked one question first. Does the writer describe a problem they hit themselves? About 6.4% do.

That projects to roughly 235,000 firsthand problems across the year. We then sorted each problem into 149 need cards. 58% fit a card cleanly. The other 42% are real problems too specific for any card.

The monthly rate barely moves. It stays between 5.5% and 7.1% of comments. So HN does not get angrier. What changes is what people complain about.

Thirteen groups of needs cover the map. No single need dominates.

The biggest single need is broken, slow websites. It holds only 2.0% of problems.

This is a long tail. The top 20 needs hold about 22% of all problems. The words people use show each group's texture.

Who complains also differs sharply by group. The bars below show who wrote each group’s complaints.

Talk and pain are different things.

Some topics are all talk. We compared each topic's share of all comments with its share of complaints.

Consumer tech is 6.9% of all talk but 17.1% of complaints. That is 2.5 times. Product design runs 1.9 times. Infrastructure and ops run 1.7 times.

Legal and government topics are the opposite. They are 12.7% of talk, 1.9% of complaints. AI is 14.7% of talk but only 12.2% of complaints. People discuss it more than suffer it.

Some needs live inside one domain. Others appear everywhere.

Hostile discussions, unreachable support, and arbitrary moderation span many domains. Version control pain and database tuning stay inside software.

Six months is enough to see the ground move. AI rose from 9.0% of problems in the first half to 13.0% in the second.

AI chatbot complaints kept climbing all year. From the first quarter to the last, developer tools fell.

A complaint is cheap. Paying, hacking, or quitting is a stronger signal.

Paid products getting worse leads. 60% of those writers already pay, switch, or quit.

Other needs get hacked instead of bought. Smart home cloud lock-in and switching AI tools both hit 66% workarounds. People rebuild small tools themselves in 59% of those complaints.

We checked three things. Replies, named tools, and product launches.

Replies rarely close the loop. In the 40 biggest needs, 68 of 100 problems got no reply naming a fix. Authors almost never come back to say it worked. That happens under 1% of the time.

The same tools appear on both sides. The popular ones get blamed and recommended. A few small tools are recommended unusually often. uBlock Origin, uv, and Jujutsu stand out.

Then we asked where builders actually launch. We assigned a sample of 8,001 Show HN launches to the same 149 needs.

Builders crowd some needs. Rebuilding small tools gets 6.1% of launches but 0.8% of complaints. That is eight times more launches than complaints.

Other needs get almost no launches. Early hardware failure with no repair got none. Broken websites and updates that break devices got almost none too.

Now we combine all six tests into one score.

You choose the weights. The ranking updates live.

Switch to the underbuilt preset. Bad management and buggy car software rise to the top.

This whole study cost $22.51 of classifier credit, over 501,082 calls.

Our 50% confidence rule also needs probabilities. Many LLM APIs do not return them. A monthly refresh with Jev costs about $1.60.

The median call took 260 milliseconds. The main screen took 2 hours 19 minutes. We checked the classifier four ways.

Our choices shape the ranking. So we reran it under five other choices.

The top needs mostly hold their place. A few are fragile and marked. Some needs rest on a few threads or a few loud people.

Those are flagged. Treat them as leads, not findings. Other limits matter too.