Beef prices keep going up a noticeable amount in grocery stores. For Bloomberg, Ilena Peng, Denise Lu, and Stephanie Davidson charted how increasing costs in the supply chain feed into the dollar amount that we see at the end.

To demonstrate with relatable units, they follow the timeline of a single calf as it moves from ranch, to stocker, feedyard, meatpacker, grocer, and consumer. Illustrations provide a visual anchor through the process.

Chicken overtook beef in 2004. It doesn’t seem like that’s going to let up any time soon.

John Nelson and Peter Atwood review maps that appeared in movies, such as Indiana Jones, Harry Potter, and Goonies. They discuss implementation, practical aspects, and accuracy. It’s about as nerdy as you imagine it to be.

Researchers analyzed newly published websites from 2022 through mid-2025 to estimate what percentage used generated text and how this might affect future information online.

The proliferation of AI-generated and AI-assisted text on the internet is feared to contribute to a degradation in semantic and stylistic diversity, factual accuracy, and other negative developments. We find that by mid-2025, roughly 35% of newly published websites were classified as AI-generated or AI-assisted, up from zero before ChatGPT’s launch in late 2022. We also find evidence suggesting that increases in AI-generated text on the internet bring about a decrease in semantic diversity and an increase in positive sentiment. We do not, however, find statistically significant evidence supporting the hypothesis that an increased rate of AI-generated text on the internet decreases factual accuracy or stylistic diversity. Notably, our findings diverge from public perception of AI’s impact on the internet.

So it has grown to about a third of new sites that use AI-generated or AI-assisted text. That seems like a lot?

I’m more surprised that there didn’t appear to be a significant change in fake information or a convergence in style.

My theory is that most people putting up these generated sites are either experimenting or trying to make a quick buck. Either way, they just take whatever information is given to them via a probabilisitic model and forget about it. They don’t care what the words say or how it is said, just as long as it fills space. So the output defaults to mostly correct statements.

It seems every day the chances that AI transforms work trends towards certain. The less certain part is the how. Some jobs could go away but new ones might appear. Employment needs could rise, even for jobs with high AI exposure. For Financial Times, John Burn-Murdoch looks at jobs over the past few decades to see how new technologies changed work in unexpected ways.

The cost of an electrical vehicle used to increase quickly as you shopped for more range, but EVs that go farther have been getting less expensive in recent years. For NYT’s the Upshot, Francesca Paris shows the current trends.

I like the side-by-side trend line comparison between the steeper slow for 2016 to 2019 models versus the 2024 to 2026 models. The shift is clear and obvious.

Once the trend line flattens and Toyota makes a fully electric Corolla, I’m in.

Every month I collect tools, datasets, and resources to help you make more useful data things. Here is what happened in April.

Hank Green dissects a video that argues against climate change. The video in question cherrypicks, makes up data, and lies about many things, all wrapped up with a calm narrator to make it seem reasonable. Green less calmly explains the manipulation.

It’s always a good time to strengthen your defenses against dishonest charts.

Apparently, the words we use and how we structure our sentences in writing is nearly as unique as our fingerprints. Kelsey Piper has been using this to benchmark new LLMs by entering text and asking who wrote it. Anthropic’s Opus 4.7 model was the first to return all the correct answers.

For WaPo opinion, Megan McArdle tested the search with her own unpublished text.

Would Claude do better or worse with something more modern? I fed Claude a different opening chapter from an unpublished science fiction novel I started right before the pandemic — I contain multitudes — and this time Claude needed only 1,132 words. The eulogy I gave for my mother, lightly edited to remove some too-specific biographical details, was even faster: Depending on the passage, Claude was able to peg me as the author in as few as 124 words.

I’m too scared to try this on myself, but I’ll assume it works. Lucky for me, I’ve always written and made things with the assumption that my mother would see it.

However, if you publish words or share thoughts on social media, I hope you don’t value online anonymity too much.

For Rest of World, Rina Chandran reports on the big difference in excitement:

As AI adoption increases globally, anxiety about AI is rising — but so is optimism about its benefits, according to a recent study from Stanford University’s Human-Centered Artificial Intelligence center. Not in the U.S. To the prompt, “products and services using AI make me excited,” only 38% of respondents in the U.S. said yes, in comparison to 84% in China. Southeast Asians are among the most optimistic about AI, with 80% of Indonesians, 77% of Malaysians, and 79% of Thais agreeing.

The difference in sentiment appears to be related to each country’s trust in government regulation. From the Stanford study, here are the percentages for those who said they trust their government:

Singapore is over 80 percent trusting. Meanwhile, the United States is the lowest at 31 percent.

This isn’t all that surprising, but I wonder why there is such a big difference. Is there an overall distrust in government and AI companies in the United States? With the largest companies in the United States, do we get a closer look and therefore more skepticism?

For NYT Opinion, Paul Ford on the challenges for AI companies to build ethical systems:

All the while, money keeps gushing in. You start out transparent, sharing your journey, but then before an initial public offering of shares, you must honor the S.E.C.-mandated quiet period and restrict promotional communications. After that, the transparency never quite returns. The market demands a rising stock price. Your company still makes a lot of software, but a huge amount of time goes to tax strategy and compliance.

At that scale, people start to blur together, and human users can become aggregate pools of statistics and growth vectors that go up and down — a mulch into which you plant your products.

Cue the Harvey Dent scene about living long enough to become the villain.

The Economist shows probabilities that a person votes for each party, given a set of demographics.

But the electorate is not monolithic. The Economist has built a statistical model of it based on a survey of voting intentions by More In Common, a pollster. Our model estimates the probability that any individual will vote for one of Britain’s main political parties based on the eight characteristics that most influence voters’ choices: sex, age, ethnicity, region, education, employment status, type of housing and whether it is in a rural or urban area. In different combinations these characteristics yield 275,000 different voter profiles. Each week we get new polling data and update our calculations.

Select the demographics, such as sex, age, race, and education, and see how each factor swings the probability for each party. The overall prediction shows at the bottom.

The 2008 decision tree by Cox comes to mind.

The Kyoto Aquarium in Kyoto and the Sumida Aquarium in Tokyo each have detailed relationship diagrams for their penguins. The above is for Kyoto.

The networks are framed as reality shows with weddings, divorce, and cheating, along with likes and dislikes of each penguin. Watch out for the penguin named Pon:

Kuruma and Tako live next door to each other, and Pon has been visiting each of them in turn for snuggle sessions. Both boys are obsessed with Pon, but it seems neither of them can fully satisfy her. What’s the fate of this neighborhood love triangle!?

Oh my.

I don’t know why these exist, but it’s nice that they do. The aquariums have updated the networks each year since 2024.

[Thanks, Charlotte]

Even if only military areas are targeted, civilian and commercial structures are also damaged, because the real world isn’t separated into discrete, selectable items on a map. Bloomberg analyzed satellite imagery to estimate the type of areas damaged in the strikes.

Each detection was classified into one of six categories: military, industrial, civilian, commercial, government, or unclassified. We separated government facilities from the broader civilian category because these buildings may serve dual military-civilian purposes. Rather than forcing a single label, the analysis preserves the full mix of land use types around each detection — a site classified as “military” might also be 20% residential and 10% commercial, reflecting the mixed-use reality of urban areas.

Sets of Voronoi diagrams are used to show the percentage breakdowns for each detection.

Lower fertility is typically pitched as a bad thing, but it can be good in some ways, such as more women going to college and building careers or fewer unplanned teen pregnancies. For NYT’s the Upshot, Claire Cain Miller reports on the other side of lower birth rates.

One of the biggest drivers of the delay in childbearing is widely considered to be a success story: the decline of teen pregnancy, which had been unusually high in the United States. It reached its recent peak in 1991, at 61.8 births per 1,000 girls and women ages 15 to 19, before rapidly declining to 11.7 per 1,000 in 2025. The change is attributed to more effective contraception, education about pregnancy prevention and less sex among teenagers.

This week we highlight an overlay that obscures the useful bits and helps no one.

Shri Khalpada of PerThirtySix explains how GPS works using a set of small interactive globes.

The answer is in some ways simpler than you’d expect, and in other ways more complex. GPS is fundamentally a translation tool: it converts time into distance. A satellite sends a signal, your phone catches it, and the delay between those two events tells the phone exactly how far away the satellite is. Everything else is about making that measurement precise enough to be useful: accounting for bad clocks, satellite geometry, and eventually, Einstein’s theories.

So geometry is useful. Imagine that.

OpenAI announced their generative model ChatGPT Images 2.0. One of the new features is that you can generate more than a single image in a prompt, which means you don’t have to generate images one-by-one and stitch them together on your own.

So now everyone can generate research posters like the one above with a quick prompt. Blessed day. Although, the robots are going to eventually do all the work for us anyways, so I’m not sure what the point is.

Mortality varies widely by geography and demographic group. It has also changed over time with improvements in medicine or availability of resources. Our World in Data shows the differences with a treemap. Use the dropdown menus to select groups and a slider to shift time.

For low-income countries:

[N]on-communicable diseases account for 43% of deaths; that’s a much smaller share than in the world as a whole (75%). That’s not because death rates of these diseases are lower in poorer countries; adjusting for age, they’re actually higher than they are in rich countries.

The difference is that death rates from infections, injuries, and child and maternal mortality are far higher. One in three die from infectious diseases such as HIV/AIDS, malaria, meningitis, and tuberculosis.

Maybe the hardest number in this dataset to sit with is that one in ten deaths is a newborn or a mother leaving children behind.

The administration wants to build a 250-foot tall arch in Washington. That’s a pretty big arch. To show how big that is, Marco Hernandez and Anushka Patil, for the New York Times, used illustrations of the proposal against existing arches and structures.

Using inference with what you ask, how you write, and your phrasing, a complete profile is built from just a few sentences. For the Straits Times, Amanda Shendruk and Youjin Shen use a concrete example to demonstrate.

I like the build-up in this piece. It starts with a chat, and then highlights line-by-line and word-by-word to build a complete user profile that most people never think about.

Back in my day, companies used to collect data about you in more obvious ways, such as suggesting you fill out profiles or tracking clicks across various sites. They’d convince college kids to share links on their AIM away message. Later, people would be convinced that voice assistants like Alexa and Siri were eavesdropping to serve hyper-targeted ads.

Well no more. These days, a chatbot will do.