Speaking of data games, Dataguessr by David Bauer is a sorting game that follows in the steps of NYT’s Flashback quizzes. Each day is a new dataset, such as total population by country. Then you sort a selection of seven countries, one at a time. The goal is to place as many countries as you can in the right spot on the list.

This one is pretty tricky when the estimates, from Our World in Data, are close to each other.

This week, we highlight more literal representations of data, because in the end, it’s the only thing that separates us from the artificial.

Chartle, by Erwan Rivault and Adnaan Jiwa, is a game to test and/or improve your knowledge about world demographics. In the spirit of Wordle, the game presents a line chart each day that shows a time series for different countries. The goal is to guess the country highlighted by a red line. You get five guesses.

For the Washington Post, Taylor Telford, Jaclyn Peiser, and Federica Cocco report on job numbers, listings, and unemployment, which have not looked favorable for many over the past year.

Hardly any corner of the economy is untouched by jobs cuts and slowdown: Employment in all goods-producing industries slumped in August, with the deepest losses coming from manufacturing and mining. The service sector was racked by steep layoffs in business and professional services and IT.

My general feeling is that data folks have seen better days, which seems to be part of a broader trend. Hoping things start ticking in the other direction.

For Reuters, Mariano Zafra and Sudev Kiyada highlight the true size of Africa and use the opportunity to describe map projections with handy illustrations.

You would think by now, after many maps, illustrations, and interactive graphics, we would have a better intuition for the pros and cons of different map projections. But then the African Union wouldn’t still need to campaign for anything other than the Mercator projection.
Read More

OpenAI released a study of how people are using their chatbot.

Patterns of use can also be thought of in terms of Asking, Doing, and Expressing. About half of messages (49%) are “Asking,” a growing and highly rated category that shows people value ChatGPT most as an advisor rather than only for task completion. Doing (40% of usage, including about one third of use for work) encompasses task-oriented interactions such as drafting text, planning, or programming, where the model is enlisted to generate outputs or complete practical work. Expressing (11% of usage) captures uses that are neither asking nor doing, usually involving personal reflection, exploration, and play.

A relatively small percentage is for programming and even less for data analysis. Writing and how-to queries take the majority, which I can only assume is mostly for LinkedIn posts.

This summer was supposed to be a return to pre-pandemic levels of movie-going, but that was not the case, as shown by Christine Zhang and Brooks Barnes for the New York Times.

I’m into the movie poster as stacked bar geometry. Now do it with streamgraphs.

There is a growing trend among investors to put money in places with high dividends. They prefer money now over decades from now. However, the higher dividends come at a cost in the longer term. For Bloomberg, Denitsa Tsekova and Vildana Hajric, with graphics by Armand Emamdjomeh, discuss the trade-offs that make the money-now approach seem less favorable.

The current wave of interest is new enough — and many of the followers young enough — that it has been easy to ignore how the most popular funds have often lagged basic stock indexes and threaten to eat away at long-term returns. Samuel Hartzmark, a professor of finance at Boston College, has researched the issue for more than a decade and has found that investors tend to fall for the “free dividends fallacy,” treating them and capital gains as separate. A 2015 paper of his finds that investors prone to that bias have a preference for funds that report boosted dividends even if they don’t improve overall returns.

There is a scrolling chart at the beginning that compares returns with different strategies over time. At first, I wasn’t fond of a non-zero baseline on an area chart, but technically it’s a difference chart with a baseline that indicates the starting investment. I guess I’ll let it slide.

Bloomberg gained access to an email cache from Jeffrey Epstein’s Yahoo Mail inbox, spanning two decades between 2002 and 2022. They highlight the relationship between Epstein and Ghislaine Maxwell and how the timing and nature of the emails reflect a closer timeline than Maxwell has suggested.

The cache also included a spreadsheet from Epstein’s accountant that lists gifts and payments, such as watches and massage lessons. Bloomberg charted three years of transactions with a bubble chart over time.

It’s only a fraction of Epstein emails, as he used multiple addresses over decades, but it’s quite the digital trace. Find more on how Bloomberg verified and analyzed the cache here.

This sight is growing too common in the United States. The New York Times shows an aerial view of where Kirk was and where the shooter is believed to have been. A Utah locator map appears in the top left to provide geography, and a north arrow on the bottom right sets orientation.

In case you’re trying to navigate from one star to another in the Star Wars galaxy, there is an official detailed map of the fictional space. It is not comprehensive, as apparently the galaxy “contains billions of stars and is home to trillions of beings” but it seems like it should at least be good enough to figure out which direction to go. (via kottke)

I refreshed an old project with new data and interactions. This week we walk through the steps.

For Quanta Magazine, Shalma Wegsman provides a history of Jean-Baptiste Joseph Fourier’s transform. James Cooley and John Tukey get a nod:

Any 8-by-8 image, for example, can be built from some combination of the 64 building blocks [above]. A compression algorithm can then remove high-frequency information, which corresponds to small details, without drastically changing how the image looks to the human eye. This is how JPEGs compress complex images into much smaller amounts of data.

In the 1960s, the mathematicians James Cooley and John Tukey came up with an algorithm that could perform a Fourier transform much more quickly — aptly called the fast Fourier transform. Since then, the Fourier transform has been implemented practically every time there is a signal to process. “It’s now a part of everyday life,” Greengard said.

I’ve said this before, but this would’ve been useful for me in college. The two-hour lectures on Fourier transforms, after lunch and in the dark, were brutal and I might’ve missed a slide or twenty.

Synthetic sampling uses models to “survey” fake respondents. G. Elliott Morris and Verasight compared real polling data against the synthetic variety to find that the latter is error-prone.

We find that the AIs cannot successfully replicate real-world data. Across models, the LLMs missed real population proportions for Trump approval and the generic ballot by between 4 and 23 percentage points. Even the best model we tested overstated disapproval of Trump, and almost never produced “don’t know” responses despite ~3% of humans choosing it.

For core demographic subgroups, the average absolute subgroup error was ~8 points; errors for some key groups (e.g., Black respondents) were as large as 15 points on Trump disapproval, and smaller groups had larger errors still (30 percentage points for Pacific Islanders). This is unusable for serious analysis.

Find the white paper here.

The point of polling is to estimate reality, so the premise of synthetic sampling through mathematical models instead of through people does not make sense to me.

Bots have crawled the web for a long time, but the past couple years has been something different as companies release their AI crawlers to scrape as much as possible. Cloudflare broke it down by type of bot and source.

Not all crawlers are the same. Bots, automated scripts that perform tasks across the Internet, come in many forms: those considered non-threatening or “good” (such as API clients, search indexing bots like Googlebot, or health checkers) and those considered malicious or “bad” (like those used for credential stuffing, spam, or scraping content without permission). In fact, around 30% of global web traffic today, according to Cloudflare Radar data, comes from bots, and even exceeds human Internet traffic in some locations.

A new category, AI crawlers, has emerged in recent years. These bots collect data from across the web to train AI models, improving tools and experiences, but also raising issues around content rights, unauthorized use, and infrastructure overload. We aimed to confirm the growth of both search and AI crawlers, examine specific AI crawlers, and understand broader crawler usage.

Every now and then I glance at traffic sources, and AI bots seem increasingly common. I wonder if or when bot traffic outnumbers human visits.

A big promise behind AI coding tools is that they will make you more productive. You will be able to code like a half-human-half-machine, leaving all full humans behind to dream of your skills. Developer Mike Judge was skeptical, so he measured his own productivity and found “that the data isn’t statistically significant at any meaningful level.”

The Bureau of Labor Statistics revised job counts down by almost a million:

The preliminary estimate of the Current Employment Statistics (CES) national benchmark revision to total nonfarm employment for March 2025 is -911,000 (-0.6 percent), the U.S. Bureau of Labor Statistics reported today. The annual benchmark revisions over the last 10 years have an absolute average of 0.2 percent of total nonfarm employment. In accordance with usual practice, the final benchmark revision will be issued in February 2026 with the publication of the January 2026 Employment Situation news release.

Each year, CES employment estimates are benchmarked to comprehensive counts of employment from the Quarterly Census of Employment and Wages (QCEW). These counts are derived primarily from state unemployment insurance (UI) tax records that nearly all employers are required to file with state workforce agencies.

I am sure all interested parties will consider uncertainty through estimation, as is common in statistics that require timeliness and accuracy.

For NYT’s the Upshot, Aatish Bhatia, Francesca Paris, and Rumsey Taylor show how zodiac signs were determined by the position of constellations relative to Earth and the Sun thousands of years ago. That probably seemed like a good idea at the time, but more accurate measurements and records show that if we continued to go by the relative positions, our zodiac signs would be different.

The interactive elements and animations illustrate the shifts well. Enter your birth date to see how your sign would change, and a night sky moves to show how our perspective changes because of Earth’s wobble.

NPR enlisted the band Bettis And 3rd Degree to sonify rising temperatures in New Orleans. As the temperature rises from 1980 to present, listen as the music tempo speeds up.

Between 1980 and 2000, the average annual temperature in New Orleans goes up by more than a quarter of a degree, and it may not seem like much if you’re just looking at the data in a spreadsheet, but it is significant.

And between 2000 and 2015, it jumps up again, almost a full degree warmer than the first 75 years of data. This is where things really start to pick up.

Over the last decade, the temperature increase has accelerated even more, almost two degrees Fahrenheit more than the 1980s. And many climate scientists say temperatures are increasing faster.

Read More