The New York Times estimated political bias in Grok’s models. They entered questions from an ideology survey and evaluated the chatbot’s answers over time.

By July 11, xAI’s updates had pushed its chatbot’s answers to the right for more than half the questions, particularly those about the government or the economy, the tests showed. Its answers to about a third of the questions — most of them about social issues like abortion and discrimination — had moved to the left, exposing the potential limits Mr. Musk faces in altering Grok’s behavior. Mr. Musk and his supporters have expressed frustration that Grok is too “woke,” something the billionaire said in a July post that he is “working on fixing.”

Musk continues to make “fixes” that align with his opinions. Not ideal.

A Tesla Model S, while on “Autopilot”, crashed into a parked truck, killing one and injuring another. Tesla claimed in a lawsuit that the data about the crash could not be found. So lawyers for the plaintiffs tried a different route. For the Washington Post, Trisha Thadani and Faiz Siddiqui report:

That’s when they turned to hacker greentheonly, who had a robust social media following for his work recovering data from damaged Teslas and posting his findings on X.

[…]

Inside a Starbucks near the Miami airport, the plaintiffs’ attorneys watched as greentheonly fired up his ThinkPad computer and plugged in a flash drive containing a forensic copy of the Autopilot unit’s contents. Within minutes, he found key data that was marked for deletion — along with confirmation that Tesla had received the collision snapshot within moments of the crash — proving the critical information should have actually been accessible all along.

Even if the data were corrupted or lost initially, which it seems like it wasn’t, Tesla should know their systems well enough to access the data in different ways. One would hope at least, for a company pushing for ubiquitous self-driving cars. Either that or Tesla should hire greentheonly to fix all their systems.

Pew Research asked U.S. adults if certain behaviors in public, such as cursing or smoking, were acceptable. The above are the results for four age groups.

For every behavior, the percentage of people who said it was rarely or never acceptable increased with age. Television and movies (and my own experiences) would tell you that sounds about right, but for some reason the clear trend surprised me. A quiz with the behaviors lets you get in on the action to see how crotchety you are.

At least we all mostly agree that taking a photo without permission is not acceptable. Now about all that public surveillance.

Gordon Hart put together a fun interactive atlas of space. Click on objects, pan, and zoom. It’s not comprehensive but not bad for a side project:

This was a fun side project over Winter Break 2024 to learn orbital mechanics, browser animation, serverless deployment options, and of course facts about moons, asteroids, and comets. Building for yourself is a treat that I haven’t properly enjoyed in some time. Source code is available on GitHub at @gordonhart/atlasof.space.

TSA PreCheck allows U.S. travelers to skip the part of airport security where you take your shoes off. For NYT’s the Upshot, Ben Blatt and Christine Chung analyzed wait times at a handful of airports to see if the program really shaves off minutes. The program seems to help most of the time, at New York airports at least.

At these airports — Kennedy, LaGuardia and Newark — PreCheck saves travelers five to 10 minutes on average, according to our analysis. But this will depend on the terminal you’re flying out of and the time of day you’re departing. Sometimes you may save over 30 minutes, sometimes none at all.

At some other major airports, the average time saved appears to be lower than in New York.

A histogram shows the distribution of wait times. The orange side on the right indicates less time waiting with PreCheck than with non-PreCheck. The blue side on the left indicates the same or more.

The U.S. federal government does not release official numbers for the cuts to health agencies, so ProPublica tracked changes to federal employee directories to estimate. They estimate 20,500 workers lost since January 2025. With a set of unit charts, each square represents five employees.

For WaPo’s Department of Data, Andrew Van Dam goes looking for American and British words, finding the most statistical action in spoken words over written.

We used our new measure to focus our book-word analysis on those words that people say out loud with at least some regularity. And with every crank of the out-loud-ness dial, we watch the two dialects get less and less similar, and more and more hilarious.

We start with relatively harmless terms. Footballers and whingeing stand out on one side, and statewide and nonfat stand out on the other. But it escalates quickly. British rises from manky and dodgy to shagging to knobhead and bruv. American goes from hoagie and doggone to homegirl and loogie to — at the slangiest echelons of the language — cornhole and bruh.

In the above, more British is on the left and more American is on the right, based on the OpenSubtitles dataset from Opus.

Then there’s the swear words.

Here are tools to use, datasets to poke at, and resources to learn from.

For Wired, Lauren Goode spent a day vibe-coding at Notion:

The next assignment was as open-ended as Lucy’s was specific: to build whatever I wanted. The freedom was unnerving, a Rorschach test for vibe coders. What did I see when I looked at the blinking cursor? I decided that there should be a way for Notion users to draft an “intelligent” to-do list in one step. They would be able to flick open the app and type “to do reorder pet food” and Notion AI would know what they meant. I also wanted this feature to avoid duplicate items from other recent to-do lists.

I was crushing it. I was a responsible babysitter for code, watching it cascade in front of my eyes and then toddle its way into the world. Except, my logic was wrong. My to-do list hack was somehow allowing for endless duplicates instead of avoiding them. Who was to blame: me or the AI?

A product designer named Brian talked me through it. “Pretend you’re talking to a smart intern,” he said. Again with the interns.

Kudos for not using “magic” even once.

New to me, someone anonymously scraped and published the Spotify playlists of a handful of celebrities and politicians. The project is called the Panama Playlists.

I found the real Spotify accounts of celebrities, politicians, and journalists. Many use their real names. With a little investigating, I could say with near-certainty: yep, this is that person.

I’ve been scraping their playlists for over a year. Some individuals even have a setting enabled that displays their last played song. I scraped this continuously, so I know what songs they played, how many times, and when.

The Panama Papers revealed hidden bank accounts. This reveals hidden tastes.

Will Smith posted a video showing highlights from his tour, but parts of it look AI-generated. Signs in the crowd are illegible, faces look smudged, and fingers are maybe a little too long. Andy Baio looked into it more closely, showing why the internet thinks the crowds are fake but are probably real and it’s more likely bad AI-assisted editing.
Read More

For the Washington Post, Kasha Patel and Naema Ahmed mapped the change in summer time.

The analysis, conducted by climatologist Brian Brettschneider, examined the hottest 90 days of the year from 1965 to 1994 and compared their frequency over 1995 to 2024. He found that the temperatures that used to kick off the hottest three months of the year are expanding beyond the calendar definition of summer.

I like the interactive bit at the beginning to enter your ZIP code. It’s speedy.

Speaking of the Correct the Map campaign, Miguel García Álvarez reminds that there is no such thing as the correct map, because all maps are a compromise to show a three-dimensional space in two dimensions.

And that is precisely where my problem lies. I am the first to argue that, in education, in newspapers and magazines, and in general, the cartographic projections used to represent the world map should be a good compromise. I think the Equal Earth projection is an excellent compromise. But as a cartography enthusiast, it pains me deeply every time someone talks about the “true” map, the “correct” map, or that the Mercator projection is “wrong.”

This is of the same ilk from Statistics that, “Essentially, all models are wrong, but some are useful.” Also, probability does not exist.

For Reuters, Catarina Demony and Ayendeng Bior report on the African Union’s push for the Equal Area projection (above) instead of the ill-used Mercator projection:

“It might seem to be just a map, but in reality, it is not,” AU Commission deputy chairperson Selma Malika Haddadi told Reuters, saying the Mercator fostered a false impression that Africa was “marginal”, despite being the world’s second-largest continent by area, with over a billion people. The AU has 55 member states.

Such stereotypes influence media, education and policy, she said.

The Mercator projection is good for location directions, but for world maps, the size of countries and continents is not proportional to reality. Africa in particular looks much smaller than it is with Mercator.
Read More

Soloham is a yo-yo style where there are two yo-yos off the string. They can be tossed in the air and thrown in patterns. It kind of looks like juggling with yo-yos. Typically, the style was used in a limited way because of difficulty level and scoring in competition, but in 2019, Hajime Miura used Soloham for his full routine.

Hajime had won the year before, and while the routine was a clear crowd-pleaser, he placed seventh. Shocking.. This explainer video shows why.
Read More

Adam Bonica analyzed the age of donors across different groups and politicians. For standard donations, ages tend towards middle age and older with some in between. On the other hand, spamming PACs use the a Nigerian prince strategy.

This is an important question because the answer informs our understanding of the problem at hand and what solutions are viable. Are the donors to these spam PACs just like any other political donors, or are they systematically different? I calculated the age of donors giving to these spam PACs, and the figure below compares their age distribution with that of ActBlue donors who gave directly to presidential campaigns.

What the data reveals is troubling: these spam PACs are part of a system engineered to target and extract money from seniors.

This week we make a solid, straightforward chart more readable and focused. It’s about the small things.

PlanScore uses four measures to define partisan gerrymandering, and they’ve made the data available over time.

Most of our federal and state legislators are elected from districts. Every ten years, state governments redraw district boundaries in a process known as redistricting. PlanScore promotes fairness in the redistricting process. We make it easy for policymakers and advocates to score new district maps and assess whether they’re fair or gerrymandered. We also provide access to the most comprehensive historical dataset of partisan gerrymandering ever assembled.

I think this could be the year that more people learn what gerrymandering is. Only a few thousand more AI-generated memes to go.

For the Washington Post, Drew Harwell reports on the budding industry of AI-generated videos that make no sense but people watch and enjoy anyway.

Their power has spawned a wild cottage industry of AI-video makers, enticed by the possibility of infinite creation for minimal work. Adele, a 20-year-old student in Florida who spoke on the condition that only her first name be used because she fears harassment, told The Washington Post she is taking a break from college to focus on making money from her AI-video accounts. Another creator in Arizona who went viral with an AI airport kangaroo said he made $15,000 in commissions in three months, speaking on the condition of anonymity out of concern over online harassment.

But the flood of financially incentivized “slop” has also given way to a strange new internet, where social media feeds overflow with unsettlingly lifelike imagery and even real videos can appear suspect. Some viral clips now barely rely on humans at all, with AI tools generating not just the imagery but also the ideas.

The novelty will wear off but generative text-to-video tools will also improve in speed and quality. There are curious times ahead.

My young son likes to draw, animate, write stories, and stitch together videos. He likes to make. I tell him to use manual tools so that he knows the process and can develop his imagination. But as things are now, I wonder if and when it’ll be time to introduce other tools.

Hannah Recht is tracking data through the U.S. Census Bureau APIs:

The Census Bureau doesn’t maintain a public changelog or regularly share updates about changes to their sprawling API universe. As a longtime Census data user and package developer I’ve found it nearly impossible to know when data is added or removed. This project attempts to publicly catalog the changes.

We need a government-wide changelog.