The Hardest Part of My AI News Site Was One Number
Keeping up with AI news had become a part-time job.
Every launch gets covered ten times. The same story arrives from a lab blog, four tech sites and a hundred posts on X, each with a slightly different headline. I was reading the same news ten ways and still missing things.
So I built Newsference, a daily index of AI news. About fifty sources feed it, scanned every hour. Articles about the same event are grouped into one story, so ten articles become one row. Stories confirmed by three or more outlets rise to the top. Single-source stories sit lower, in their own lane.
And at the top of the page is one number: the NF Index, from 0 to 100, telling you whether today is quiet, steady or heated.
That number took more design than everything else combined.
A number has to feel true
The index looks simple. It weighs the day's stories, gives more weight to ones confirmed by several outlets and to news from the major labs, then compares today to the last two weeks.
Every early version was technically correct and felt wrong.
My first baseline was a fixed number. The news was busier than I'd guessed, so every single day came out as FRENZY. An index that's always at maximum isn't an index. It's a siren.
So I compared each day to itself. That pinned the score to exactly 50, every day, forever. Correct maths, useless product.
Then I compared today to the recent past on a straight line. It worked until a busy morning, when the index jumped to 100 out of 100 and had nowhere left to go. The next big day would look the same as this one.
The version that stuck uses a curve that eases off at both ends. Ordinary days move the number, big days push it up, and a genuinely wild day still has room above it. I widened that curve once more so there'd always be headroom for the day something truly huge ships.
None of those changes came from a spreadsheet. Each one came from looking at the number, reading the day's headlines, and thinking: that's not what today felt like. That's a design loop, not a statistics exercise. The number has to match what a reader would say if you asked them how the day went.
What I left out of it
X is noisy, and I wanted its signal on the page. Each story shows how much attention it's getting there.
But I kept X out of the index itself. In one early version, mentions were counted more than once, and the index jumped forty points on chatter alone. An index of how loud people are isn't the same as an index of what happened. The page shows both. The number only measures one.
The same thinking changed how stories are ranked. At first, a post from a trusted publisher counted as a top story on its own. But how much you trust a source and how big a story is are different things. Now importance comes from how many independent outlets confirm it.
The model made up a story about itself
The grouping is done by a language model, and one day it produced a top story called "Non-AI articles to be dropped." The summary read, in full: "Placeholder."
The model had turned a section of its own instructions into a news story, and my ranking put it at the top of the page, above a real story about Sam Altman's testimony.
It's funny, and it's a lesson. When a model is part of your product, its output is data, not truth. Newsference now checks what comes back and rejects clusters that don't look like news before anything reaches the page.
The cheapest fix was noticing
In the first five days, the model cost US$13.52 to run. That doesn't sound like much until you check where it went. Of 988 articles in one day, 826 were research papers, and every paper is its own story. There was nothing to group.
So papers skip the model entirely. The daily cost dropped to about fifty cents, and it now runs on a free model. The saving didn't come from picking a cheaper model first. It came from noticing where the model's answer was predictable.
Why it looks like a terminal
Newsference is dense on purpose. The brief I wrote for it said "Bloomberg Terminal meets brutalist minimalism", marked the design as locked, and put it plainly: design quality is the moat.
News sites are built to keep you scrolling. A terminal is built to be read fast and left. That's the whole goal: catch the day in sixty seconds, not sixty minutes.
Where it is
Newsference is at newsference.com. It's free. No paid placement in the rankings, and anything sponsored is labelled.
It now gets over 10,000 unique visitors a month, and it's still growing. Most of them, I suspect, just want to know whether today is heated.
The scraping was easy. The hard part was a single number that doesn't lie about your day.
---
