| ⬡ Deep Dive | How Physicists Track and Trap the Elusive Neutrino | 18 min |
| ◈ Science | NASA's Webb Finds the Strongest Evidence Yet for 'Black Hole Stars' | 7 min |
| ⬡ AI | Porting an Image Model to the Browser, Entirely Through Claude Code | 8 min |
| Browse | After 80 Years, a Famed Erdős Method Gets an Upgrade | 9 min |
| Browse | Why American Data Centers Can't Plug In | 18 min |
The leisurely centerpiece: seventy years of trying to catch a particle Pauli proposed in 1930 with the rueful verdict 'I have done a terrible thing — I have postulated a particle that cannot be detected.' Frantz follows the chase from Cowan and Reines parked beside a Savannah River reactor in 1956, through Ray Davis's chlorine tank a mile underground in the Homestake mine, to IceCube's cubic kilometer of instrumented Antarctic ice and JUNO's new 20-kiloton sphere. It's a detective story told through the absurd, beautiful machines the ghost particle forced us to build — and the recent JUNO and sterile-neutrino headlines land differently once you've read how we got here.
The 'little red dots' JWST keeps finding in the early universe looked like they might break cosmology — galaxies far too massive, far too soon. A team led by Vasily Kokorev now makes the strongest case yet that one of them, GLIMPSE-17775, is something stranger and more reassuring: a supermassive black hole swaddled in so dense a cocoon of gas that it mimics a star's spectrum. The tell is an 'iron forest' of 16 emission lines plus helium fluorescence that only a hot, accreting black hole behind a thick veil should produce — which means the masses needn't be impossible after all.
A small, concrete demonstration of where the frontier quietly moved. Willison saw a new 0.2-billion-parameter image-inpainting model on Hacker News — released, as usual, requiring PyTorch and a CUDA GPU — and had Claude Opus 4.8 convert it to ONNX, push it to Hugging Face, and wrap it in a WebGPU app that now runs entirely in your browser, no server, the 1.3 GB model cached locally. The interesting part isn't the demo; it's how routine 'take this research artifact and make it run anywhere' has quietly become.
Erdős's 1947 probabilistic method — don't construct the object you want, just show a random one works with nonzero probability — reshaped combinatorics and then refused to be improved on for the hardest Ramsey-number bounds for fifty years. A graduate student, Wujie Shen, with Jie Ma and Shengjie Xie, finally nudged the exponent: from 2.618^k to 2.618000000000001^k. The improvement is comically tiny and genuinely huge — the first crack in a fifty-year wall, opened with a geometric trick (placing points on a high-dimensional sphere) that, in hindsight, was hiding in plain sight.
The binding constraint on AI turns out not to be chips or capital but copper and transformers. This is the clearest explanation I've seen of why: interconnection queues now run years, high-power transformers carry five-year lead times, and grid 'congestion' means the cheap power is stranded on the wrong side of too few cables. The vivid case study is xAI's Memphis site, which came online able to pull eight megawatts from the grid — enough for a few thousand toasters — so it simply bolted on 422 MW of its own gas turbines and ran half off-grid. A useful corrective to abstract 'AI energy demand' takes.
Vafa's single micron-scale extra dimension, into which gravitons leak and grow heavy — turning dark matter and dark energy into two faces of the same geometry, and predicting both should fade over cosmic time. It held up as the rare string-theory idea the next decade of cosmology can actually test.
The finding that across nearly 19,000 conversations today's models already out-persuade humans at shifting real-world beliefs — and the effect grows with capability. What makes it stick is the pairing Clark draws with fresh evidence of self-sustaining AI development.
For a few months we got to sample pristine, deuterium-rich ice that froze 10–12 billion years ago, carried through the solar system on an interstellar comet from a star far older than ours. The kind of cosmic luck worth remembering after the news cycle moves on.
For the symmetry lover. Noether's theorem — that every continuous symmetry of a physical system corresponds to a conserved quantity, time-translation giving energy, spatial translation giving momentum — is arguably the most beautiful single result in physics, and the quiet scaffolding under everything from Lagrangian mechanics to the Standard Model. This essay sits with both the mathematics and the woman: a thinker who reshaped abstract algebra and physics while barred from a salaried professorship, lecturing under Hilbert's name because Göttingen wouldn't appoint her. Worth it for the reminder of how much of the field's deepest structure she laid down without credit.
A short, specialist pleasure: a review of Dieter Möhl's monograph on stochastic cooling — the technique Simon van der Meer invented to squeeze a diffuse cloud of antiprotons into a dense, usable beam by sensing each particle's deviation at a pickup electrode and kicking it back into line downstream. It's the unglamorous accelerator-physics machinery that made the W and Z bosons discoverable at all, and won the 1984 Nobel. The reviewer lingers on the Schottky-band analysis — integrated power constant, bandwidth growing linearly with harmonic number — exactly the kind of detail that rewards anyone who once cared about beam dynamics.