weird.stats

Wonderfully unnecessary.

How early could an attacker get control of a website?

Your login is one piece of a much larger system.

Reveal the discovery

Before the website’s own login code started.

PHP is software that builds web pages on a server, the computer that sends pages to your browser. In the demonstrated attack, researchers made PHP run their own instructions before the website checked a password.

The demonstration used Apache with mod_php, one way to run PHP. The attacker had the same permissions as the server program handling the request. That did not automatically give them control of the whole computer.

A simplified illustration, not a real server or memory layout.
Look beneath the login

The login is the part you see. PHP handles the request underneath it.

The flaw, in three steps.

Read straight through, or select a step to see it inside.

  1. 1. A request reaches PHP.

    A request is a message sent to a server. The attacker sends a specially made one, and PHP starts setting up the data to handle it.

  2. 2. PHP runs out of working room.

    Memory is PHP’s temporary workspace. If it fills up at the wrong moment, setup can stop halfway through, leaving data unfinished.

  3. 3. Cleanup can run the wrong instructions.

    PHP tried to clean up after the failure. It could use the unfinished data to decide which instructions to run next. Researchers exploited that mistake to make PHP run their instructions.

This illustrates the out-of-memory flaw demonstrated against Apache with mod_php. Other server setups have different attack conditions.

Why the login couldn’t help.

The attack happened while PHP was preparing the request. The website’s own code, including its login check, hadn’t started yet.

The PHP project released fixes on October 21, 2026. Email-alias service addy.io chose to take its website and web app offline while patching; its email delivery kept working.

A demonstrated attack does not mean every affected website was hacked. Other ways of running PHP have different attack conditions.

Open this discovery

Do your bath wrinkles make a new pattern every time?

You’ve seen your fingers prune. Have you ever looked closely at the lines?

Reveal the discovery

Some of the same lines return.

Researchers photographed the same fingertips after two soaks at least a day apart. The skin had smoothed out between soaks, but major wrinkles returned in similar places and directions.

This preliminary 2025 study involved three adults. Some wrinkles differed between soaks.

Could you draw it again?

Look closely at the lines. You’ll draw them from memory after the bath drains.

A painted fingertip in a tiny porcelain bathtub, with a rubber duck, warm tiles and a sunlit window.
Remember these lines

Drain the bath when you’re ready.

We made up the wrinkle pattern for this illustrated memory game. It can’t tell you whether your own wrinkles repeat.
The study behind the wrinkles

Three adults, ages 19–22, took part. Researchers photographed all five fingers on each person’s right hand after two 30-minute soaks in 40°C water, at least 24 hours apart. They matched visible wrinkles and compared their directions.

Some lines did not match. Matching wrinkles by hand leaves room for bias. The study did not establish a permanent, unique wrinkle pattern or prove the proposed blood-vessel explanation. Measuring several wrinkles on each hand does not increase the number of participants.

Laytin & German (2025): read the paper
Full paper (PDF)

See §§2–3.3 and Figure 3 for the methods and results. The paper was published on February 13, 2025. It does not report the exact dates of data collection.

2025 research · Checked October 8, 2026

Keep wandering
Open this discovery

One person.
An entire car.

A seat for you. A machine around you. Now multiply that bargain by everyone trying to get somewhere.

Deep Dive · Space, energy & the cost of empty seats

Read the Deep Dive
60 people60 cars
Roof cutaways reveal the people inside. Every pale dot is a traveler. Cars include drivers; the bus adds an operator. An illustrative snapshot of a queue on a one-way road, not a measured road or traffic-flow simulation.

Open this discovery

Sentenced to die.

4.1%

estimated to have been
wrongfully convicted.

U.S. defendants sentenced to death, 1973–2004.
A 2014 study’s estimate—not a count of innocent people executed.

Gross, O’Brien, Hu & Kennedy · PNAS, 2014

How certain is that number?

An estimate has edges.

The model’s 95% confidence interval is 2.8–5.2%. It does not capture every uncertainty about the assumptions or about who was innocent.

Estimated exoneration rate under continued threat of execution On a zero to six percent scale, the estimate is 4.1 percent and the 95 percent confidence interval runs from 2.8 to 5.2 percent. 4.1%2.8%5.2%06%
Point estimate and 95% confidence interval · scale 0–6%

Innocence doesn’t
unlock the door.

41 of 1,000

sentences illustrated in pale light · 4.1% estimate

An imagined prison cell, then 1,000 illustrated death sentences. At the study’s estimate, 41 represent wrongful convictions. Pale light shows that proportion—not identified people. Unlit cells do not establish guilt. These are not actual cells or a count of executions.

Off death row.
Still not free.

A reduced sentence can end the threat of execution without clearing someone’s name. The researchers found that exoneration became much less likely after that threat ended.

They modeled what might be uncovered if everyone remained under the intense scrutiny associated with a death sentence. That is how they reached 4.1%—which they considered a conservative estimate of wrongful convictions.

Escaping execution can still mean
spending your life in prison.

The study, the assumptions, the numbers

The study followed 7,482 defendants. Its Kaplan–Meier model estimated exoneration under continued execution threat at 21.4 years after conviction. It accounts for different observation times and departures from death row.

Exoneration is an imperfect proxy for innocence. Departures may be related to innocence, too. The authors examined those assumptions in sensitivity analyses. Their “at least” interpretation is not a statistical lower bound. This result does not describe today’s death row or all criminal convictions.

Values in the graphic · U.S. death sentences, 1973–2004
MeasurePercent
Model estimate4.1
95% interval, lower end2.8
95% interval, upper end5.2
Download the numbers (CSV)

Source: Gross et al., Rate of false conviction of criminal defendants who are sentenced to death. PNAS 111(20), 7230–7235 (2014). Figure 2; Analysis, Sensitivity analysis, and Discussion. These links refer to the same study.

Historical research · Sources checked October 5, 2026

Open this discovery

A change of direction

Why would an airport repaint a runway that hasn’t moved?

Same strip of tarmac.
A different number.

Show me why

Magnetic north moved. The runway stayed put.

Runway numbers generally describe their direction relative to magnetic north, rounded to tens of degrees. Tampa’s January 2011 notice changed 18R/36L to 19R/1L as part of a renumbering prompted by magnetic drift.

A historical example, not today’s flight information. Renumbering follows coordinated updates; it does not happen automatically whenever a compass reading changes.

184° → 186°. Runway 18 → 19.
The pavement stays in exactly the same place.

A runway number is a direction, not a count.Illustrative headings, not measurements from Tampa.
Orange arrow: magnetic north. N: true north.
Heading scale below the runway is magnified.

January 2011 example · Checked 2026-10-05

The receipts

Tampa International Airport’s January 2011 designation notice and the FAA magnetic-heading convention.

Runway designations are labels, not counts. Any future 184° → 186° illustration must be labeled hypothetical; no Tampa angle is inferred.

Open this discovery

April 1, 2020 Census

How many states combined have fewer people than California?

State borders shape representation in the Senate. Population alone doesn’t tell you how many seats a group of residents is represented by.

Show me

Resident population

California

39,538,223residents

21 states combined

37,174,921residents

Shared scale: 0–40,000,000 residents. April 1, 2020 resident counts. The second bar is the combined population of 21 states.

Nearly the same population. A very different number of seats.

Senate seats

California

2seats

21 states combined

42seats

One chair represents one allocated Senate seat, not a current officeholder. Counts: 2 for California; 42 for the 21-state group.

Both measures, side by side. Different units; separate scales.

See the numbers
Values behind both comparisons
MeasureGroupValueUnit
Resident populationCalifornia39,538,223residents
Resident population21 states combined37,174,921residents
Senate seatsCalifornia2seats
Senate seats21 states combined42seats
Download the data (CSV)

The 21 least-populous states, combined. They get 42 Senate seats. California gets 2.

In the 2020 Census, California had 39,538,223 residents. The 21 least-populous states had 37,174,921 combined—about 2.36 million fewer. Those states are allocated 42 U.S. Senate seats; California is allocated two.

This compares resident populations and allocated Senate seats using a fixed 2020 snapshot. It is not a count of eligible voters, a current population estimate, or a measure of an individual vote’s power. House representation follows a different system.

Checked 2026-10-04 · Data: April 1, 2020 Census

The receipts

California versus the 21 least-populous U.S. states in the April 1, 2020 Census. Ranked among the 50 states; D.C. and Puerto Rico are excluded because this comparison concerns state Senate representation.

Rank the 50 states by 2020 resident population and sum ranks 50 through 30: Wyoming, Vermont, Alaska, North Dakota, South Dakota, Delaware, Montana, Rhode Island, Maine, New Hampshire, Hawaii, West Virginia, Idaho, Nebraska, New Mexico, Kansas, Mississippi, Arkansas, Nevada, Iowa, Utah. Their total is 37,174,921; California has 39,538,223. Adding Connecticut, the next state, brings the total to 40,780,865, exceeding California. Thus 21 is the largest group of least-populous states whose combined population is below California’s. Seat allocation: 21 × 2 = 42 versus 1 × 2 = 2. The population bars share a zero-based 0–40 million resident scale. The seat objects use a separate literal count: one chair per allocated seat.

Denominator: Resident totals for California and the combined 21-state group, not eligible voters. Seat counts are allocated seats for each group: one state versus 21 states. No individual voting-power ratio is calculated.

Open this discovery

Could your ears make a chip seem fresher?

One recording, two sound treatments. Listen if you like, then see what happened in the experiment.

No sound? The answer works without it.

35%

Sound only plays when you tap Listen.

Which sounds fresher to you? Optional

Choose an impression, or go straight to the reveal.

An illustrative sound comparison. Listening alone does not recreate the experiment.

The receipts

Zampini & Spence, Journal of Sensory Studies (2004) ↗
The experiment manipulated auditory feedback during biting. The 15-of-20 package response is reported in the discussion, p. 357, as informal post-experiment questioning. It is not a population estimate or a measure of food freshness. Oxford abstract ↗. Our two clips are filtered versions of a separate CC0 recording by Joseph Sardin ↗, not study recordings. The chip image is generated.

Open this discovery

Next, a little pocket deception.

Keep wandering ↓

Which one has
more copper?

Pick a coin. Or go straight to the answer.

Modern U.S. penny & nickel · copper content by mass
The receipts

U.S. Mint specifications ↗
Penny: 2.5 g × 2.5% = 0.0625 g copper. Nickel: 5 g × 75% = 3.75 g copper. Ratio: 60. Based on rounded specifications, not individual coin measurements. The coin collection compares copper content, not face value or total coin mass.

Open this discovery
And somewhere in Arizona,

the mail has legs.

The last
9miles.

For Supai, home of the Havasupai people, the mail’s final leg is a mule train.

Illustrated mule carrying mail panniersSpecial delivery. Four-legged edition.

Getting there is
only half the trip.

Three hours down the canyon.
Five hours to climb back out.

10–22 mules. Five days a week.

The receipts

USPS Postal Facts, May 2026 ↗
USPS reports the nine-mile route, mule count, delivery days, and three-hour descent/five-hour return. Each mark represents one hour; playback is accelerated. The mule is an illustration.

Open this discovery

Let’s get very, very close.

One pixel.
0.005 mm of paint.

Scroll from the whole painting into a single square millimeter.

Start with
the whole thing.

The Night Watch, by Rembrandt. There’s a lot to look at. We’re heading for a very small part.

Overview reproduction
Overview of Rembrandt’s The Night Watch
The Night Watch · Rembrandt

1 mm × 1 mm

This tiny patch.
40,000 pixels.

At the museum’s image resolution, 200 pixels fit along each edge. That’s 200 × 200.

Enlarged scale diagram begins here.
200 pixels across · 1 mm of paint

0.005 mm × 0.005 mm

And this?
One pixel.

Each pixel in the museum photograph represents a patch of paint just five thousandths of a millimeter wide.

Magnified diagram, not a paint photograph.
The highlighted square is one pixel.

A very big photograph.
For very small details.

717 billionpixels in the photograph

8,439photographs stitched together

The zoom moves from an overview reproduction to mathematical diagrams. It does not display the museum’s full-resolution scan.

The receipts

Rijksmuseum, January 2022 ↗
The museum reports 717 gigapixels, 8,439 photographs, and 0.005 mm per pixel side. Calculation: 1 ÷ 0.005 = 200 pixels per edge; 200 × 200 = 40,000 per square millimeter. Overview reproduction via Wikimedia Commons. The diagram colors do not represent sampled paint.

Open this discovery