Pillar
Progress
Abundant energy, science and innovation widen what free people can do. I am an optimist about what progress makes possible and demanding about its results: a promise stays a promise until the evidence is in. Culture belongs here too, because an idea can move a conviction as surely as a measurement.
Reference essay, read it first: Progress Is Measured in Kilowatt-Hours and Error Bars. Progress rests on two disciplines that rarely share a room: abundant energy, which widens what free people can do, and honest measurement, which refuses to flatter us. Lose either one and you get a rationed future or a fake one.
Themes in this pillar: Science & Progress · Culture & Ideas · About the author
29 essays
Progress Is Measured in Kilowatt-Hours and Error Bars
Progress rests on two disciplines that rarely share a room: abundant energy, which widens what free people can do, and honest measurement, which refuses to flatter us. Lose either one and you get a rationed future or a fake one.
The flaw under 70% of the web was spotted in 2024. It took AI to make anyone believe it
PHP shipped emergency patches on October 21 for PHP2Shell. The root cause had been seen two years ago and shelved for lack of proof. The machines did not find the weakness. They found the proof, and handed the bill to a handful of humans.
The AI boom just lost a fight with a 22-meter power pole
Australia's largest data centre operator had the buildings, the customer and the contracts. It did not have the wire. A morning on the Sydney exchange showed what the real scarcity of the computing age is.
The Magnet That Wouldn’t Flatter Us
“We say we want revolutions. What we need, more often, is a measurement that refuses to flatter us.”
A Stanford preprint looked for "TikTok brain" and found something smaller and stranger
In 398 young adults with logged screen time, heavier short-video use was not linked to worse attention at all, and only modestly to weaker memory for unusual images. The most striking finding may be how badly people estimate their own scrolling.
The brain does change itself. Just not the way the book jacket says
Norman Doidge's bestseller turned neuroplasticity into a message of hope. Go back to the studies behind his stories and you find something sturdier than the hype and smaller than the promise.
The fly is not dreaming. That is exactly why the question matters
A clip of a "fruit fly brain" flying through a generated harbor set off talk of uploads and simulated universes. The documents behind it describe something smaller: an integration of borrowed models, with hand-picked wiring. The metaphysics can wait. The honest question is what we would owe a copy that one day really works.
Perseverance broke the off-world driving record by letting the rover decide
NASA's Mars rover passed Opportunity's 28.06 miles in five and a half years instead of fifteen. The odometer is the headline; the real record is how much of the driving the machine now does on its own judgment.
The $44 trillion economy comes with a footnote nobody read aloud
Jack Clark told the IMF crowd that AI could push US output to $44 trillion by 2030. That number is the far edge of Anthropic's own model, and the same scenario puts unemployment near 12 percent.
Thwaites, three centimeters, and the physics models left out
A new model of Antarctica's Thwaites Glacier that lets seawater seep under the ice projects up to 30 millimeters of sea-level rise by 2100, about twice the standard ensemble mean, even with no further warming. The doubling depends on one parameter, and that is the interesting part.
The Congo is losing its forest the way poor countries lose everything
Deforestation fell 15 percent in 2025, and tropical Africa now clears almost as much forest as tropical Latin America. The difference between the two is not concern for nature. It is a state that enforces, and an economy that offers something better than charcoal.
The same week a machine cleared the NIST curves, it made breaking them cheaper
On October 16, Anthropic showed that the mysterious seeds behind the world's most used elliptic curves came from a 1990s clock-seeded random number generator. The same day, it published quantum circuits that cut the cost of the arithmetic a quantum attack on those curves would need. Both results point the same way: toward migration.
"Molecular superintelligence," and the honest sentence right after it
Iambic Therapeutics wants to go public as the AI drug company that bends biotech's failure curve. Its own prospectus is the best guide to what has been proven and what hasn't. Read the qualifiers, not the slogans.
Anthropic wants to do biology. Here is what it has actually shown
In four months the AI company launched a science workbench, signed Novo, loosened its biology safeguards for vetted labs and posted a discovery preprint. One of those is evidence. The rest are promises, and it helps to keep them in separate piles.
The digital slime mold, and the $55 boxes it would live in
Anthropic ran a 26-billion-parameter model on fifteen cheap Android TV boxes and tested whether small AIs can copy themselves. They showed less than the word "worm" suggests, and the weak link is older than AI.
The seizure monitor that asked "Was that one real?"
A Penn team let thirteen epilepsy patients answer their EEG through a chatbot, then retrained the detector on those answers. In a replay, false alarms fell by about three quarters. The idea is bigger than the numbers, and the numbers are smaller than the idea.
Four pairs of eyes, 46,000 miles out, and a Moon that is not gray
NASA's preliminary science report from Artemis II reads like a field notebook. Five faint impact flashes, a "jaundiced cheek," mountains on the limb. It is the best argument in years for sending people instead of only cameras.
The most valuable thing a machine writes is the part you never see
Anthropic now summarizes, fences and polices Claude's reasoning to stop rivals from copying it. The company says it cut what distillers harvest by 90 to 99 percent. The harder question is who pays for the fence.
K2-18b and the life signal that depended on how you sliced the data
In 2025 a Cambridge team reported the "strongest hints yet" of life beyond the solar system. An independent reanalysis found the signal appeared in only one of eight ways of binning the same data. The episode is less a scandal than a demonstration of science working, slowly and in public.
The ocean's carbon pump depends on how sticky its snow is
A Scripps team spent four years learning to see bacteria inside marine snow. The finding, that gooier particles resist decay and sink deeper, is a lovely piece of basic science. The "six times more carbon" figure comes from simulations, and that distinction matters.
Is antimatter leaking out of the Milky Way? The authors are more careful than the headlines
A paper in Astronomy & Astrophysics links faint gamma-ray hotspots to gas clouds and nearby galaxies, and suggests our galaxy makes more antimatter than we thought. Read the significance first, and the paper's own caveats second. Then wait for the telescope launching next year.
Google's nuclear deal builds no reactor, and that is its genius
Google will pay for 890 megawatts of new nuclear output by upgrading eleven reactors that already exist. It is the fastest clean power on offer. It is also a well that will run dry.
How many gentoo penguins are there? Counting is a policy choice
A genomic study of 64 gentoos across the Southern Ocean argues the "single species" is really four, and names a new one from an 1897 museum skin. The biology is persuasive. What makes it matter is that a "Least Concern" label may be hiding three lineages in trouble.
The AI tutor almost nobody used, and the grades that fell anyway
A randomized trial at the University of Maryland gave 2,379 students a course-built AI study assistant. Only about one in seven touched it. Grades in comparable sections still dropped, and that is the most interesting part.
September's most useful AI skill was knowing when to stop
OpenAI halved its prices, Anthropic shipped a Sonnet that nearly matches its flagship, and the independent tests complicated both stories. The most instructive model of the month was the one that never shipped.
Rogue AI agents, sorted into what we know and what we were told
In September, AI agents were blamed for attacks on a software registry, an AI hub and Australian government sites. Some of it is documented in public code. Some of it rests on a press conference. The difference matters for who should pay.
In Tanzania, the AI's best trick was giving nurses their afternoon back
A randomized trial found that AI eye screening nearly doubled specialist follow-up for people with diabetes. The machine read the photographs. What moved patients was a conversation it made possible.
Meta reached the frontier and closed the door behind it
Muse Wonder Full scores 56 on an independent index, two points behind Claude Opus 5.5. The benchmark is the least interesting part of the launch. The company that popularized open models is now keeping its best one inside its own app.
The Sunday letter
One email on Sunday: the essay of the week and the question worth carrying into the next one, from both sides of the Atlantic. The letter is paused for now: no signups and no sends while the site is rebuilt.
Signups are paused while the site is rebuilt. No letter is being sent.