Gregor J. Rothfuss

Refusing National Statistics

As Hong Kong's Financial Secretary from 1961 to 1971, Sir John Cowperthwaite sent Whitehall statisticians home rather than let the colony collect macroeconomic data:
Asked by Milton Friedman in 1963 why he refused to gather national accounts, Cowperthwaite answered, "If I let them compute those statistics, they'll want to use them for planning"—turning descriptive metrics into tariffs, subsidies, and industrial gates. Over his decade in office without national statistics or industrial policy, Hong Kong's real wages rose 50%, acute poverty fell from 50% to 15%, and per-capita income climbed from 25% of the British level in 1961 to surpassing the United Kingdom by the early 1990s.

Submerged Swamp Roads

During the 13th and 14th century Northern Crusades, pagan Lithuanian and Samogitian defenders built secret underwater stone causeways across peat bogs to move cavalry and carts between fortresses:
Laid below the swamp waterline so they were invisible from the surface, these kūlgrindos deliberately twisted through low-lying mire rather than taking the shortest crossing—as archaeologist Ludwik Krzywicki found in 1903 when surveying 3 surviving stone stretches (250 m, 180 m, and 150 m) linking the fortresses of Medvėgalis, Paršpilis, and Šiuraičiai at Sietuva—so pursuing Crusaders unfamiliar with the turns would step off the submerged stones into deep bog mud.

Riemann's Shock Equations

Bernhard Riemann's published 1860 paper on compressible gas dynamics failed to conserve energy across a shock front because it assumed reversible adiabatic compression instead of entropy-producing jump conditions:
Auditing 12,000 pages of handwritten Nachlass manuscripts by Gauss, Riemann, Hilbert, and Grothendieck at Göttingen with Opus 5.5 agents uncovered 2 unpublished Riemann sheets that derive the exact Rankine-Hugoniot shock conditions in his own notation 10–20 years before William Rankine (1870) and Pierre-Henri Hugoniot (1887), alongside shock-wave relations conventionally dated to 1906–1937.

e-means Clustering

Lloyd's k-means algorithm requires a full pass over a dataset for every centroid update, while standard mini-batch k-means converges early to higher quantization error:
Adapting vector-quantized autoencoder moving averages to static MSE minimization, Meta's e-means updates running cluster counts and vector sums on a cosine decay schedule down to 10% of peak step size, splits donor clusters every 5 batches whenever a cluster's count drops below 10% of the mean, and sets batch and step sizes directly from scaling laws with 0 per-dataset tuning. On 1 billion 1,024-dimensional DINO embeddings (1,048,576 centroids), e-means reaches 25-epoch Flash-KMeans test error in 1.6 epochs for a 45.5x wall-clock speedup; on the full 10-billion-vector dataset with 16 million centroids, exact e-means finishes in 2.5 epochs (779 GPU-hours vs. 6,784 for 25-epoch Lloyd) with 0.014% lower test MSE, while an inverted-file approximate variant reaches 0.033% lower MSE in 23.2 GPU-hours on 1 machine.

Weight-Space Programming

Frontier LLMs can directly inspect and edit raw neural network weights in Python without gradients or training loops:
Asked to improve Qwen3.5-0.8B at math with no gradient descent or training, GPT-6 Astra uses a Python kernel over the model's 0.8 billion parameters and a budget of 50 forward passes to trace why 85 of the base model's first 96 practice answers run into repetitive thinking loops at the 32,768-token limit. Astra isolates the looping behavior to a single direction written by the MLPs of layers 15 and 19 and subtracts that direction from 64 output columns in each—natively editing 0.016% of the weights (131,072 parameters). That edit cuts token-limit looping on MATH-500 from 93.0% to 33.1%, raising MATH-500 accuracy from 7.9% to 48.1%, GSM8K from 13.8% to 67.1%, and out-of-distribution instruction following (IFEval) from 21.0% to 62.3%.
2026-11-12: Splicing code layers into training via phantom gradients
Ordinary weight gradients average across a batch, hiding token-level failures. Computing ∂L/∂h_ℓ with respect to every token's hidden state at every layer—the phantom gradient—maps where representation demand persists across checkpoints. By centering phantom vectors per target token, clustering failing tokens with spherical k-means, and splicing Python code layers into the residual stream wherever they beat a linear baseline, a coding agent inside training permanently lowers validation loss on a 12-layer, 286M-parameter nanochat model, raises CLRS-Text algorithmic reasoning accuracy on unseen intermediate lengths from 55% to 84%, and lets DeepSeek-V4.1-Flash build a linear probe across blocks 24–39 that rejects 80% of its own wrong answers (at a 6.7% false-rejection rate) to lift AIME and HMMT 2026 accuracy from 64% to 71%.

Neolithic Trepanation

Why 5-10% of Neolithic skulls across both the Old and New Worlds between 12 ka BP and 5 ka BP have drilled holes with post-operative bone healing, despite no contact between continents and no antibiotics:
Trepanation is an effective treatment for epidural hematoma: a patient ruptures a blood vessel in their skull, filling the cranial cavity with blood, and drilling a hole releases the blood and saves the patient’s life. In a prehistoric culture where people hit their neighbors with flint axes, hematoma rates were far beyond any in recorded history, and dramatic recoveries—a comatose person on the edge of death suddenly waking up—drove overgeneralization of the procedure to migraines and other ailments. Primitive surgeons survived without sterilization because they stopped at the dura partway between the scalp and brain, the scalp has unusually high blood flow, and primitive bacteria lacked the virulence of modern hospital-adapted strains. As head trauma rates declined with civilization, miraculous hematoma recoveries became rare and the practice died out.

Self-Cleaning LLO

Subsurface mass concentrations (mascons) perturb orbital eccentricity across the Moon until satellites outside 4 frozen inclinations crash into the regolith within months:
At the same scale, the Moon's geoid height (−540 to +571m on the GRAIL GRGM1200B model) is nearly 6x lumpier than Earth's (−107 to +86m): very uneven gravity is why almost all low lunar orbits are unstable, making low lunar orbit a self-cleaning orbit.

x86 Boot Ontogeny

Every boot on an x86 machine iterates through the entire sequence of computing evolution in a few milliseconds:
Stage 1: 16-bit Real Mode, acts like an 8086 from ~1978. 500ns; about 15-30 instructions. Goal of this stage is to create a writeable stack+heap.
Stage 2: 32-bit Protected Mode, acts like a 386 from ~1985. 10s of millions of instructions; takes forever on DDR5 with memory training. Temporarily uses L2/L3 “cache-as-ram” to kickstart main RAM (Intel calls this No-Eviction Mode, NEM)—the same concept as an APU/cart start, using the smaller motor (L2/L3) to kickstart the main turbine (main RAM) and execute the training process.
Stage 3: 64-bit Long Mode, acts like AMD64 from ~2003. 100s of millions of instructions; enumerate PCIe devices, GPU, bootloader, etc.
Stage 4+: Secondary core wakeup. Every single core has to go through the same process, but by this point it’s faster because main RAM is warmed up.

Actuator Tolerance Stackup

Electromagnetic CAD optimization is largely tapped out, leaving sub-millimeter manufacturing tolerance stackup—and replacing manual process engineering with closed-loop automated tooling across every SKU and gearbox—as the critical path for actuator torque density and precision:
Designing a Caldera-0-3115 motor to have a 0.08mm airgap instead of a 0.16mm airgap translates to ~8% max torque at 2.5kW max throttle—more impactful than any electromagnetic design improvement from Ansys. But it would die within hours of flight time because the stator and magnets would start colliding. Current manufacturing processes create error tolerance stackup from:
1. Flipping/refixturing of the multi-step lathing process of the rotor bell not being perfectly coaxial
2. Press fits
3. Adhesives
4. Tolerancing the grinding of magnets
5. At 22k rpm, the remaining 0.x milligrams of imbalances after dynamic balancing that sway the bell from 1 side to another
6. Warping from repeated thermal expansion
Improving them requires CNCing with different tools, cutting operations, and fixtures, testing different parameters for press fit, and adhesives that respond differently to thermal expansion. And that's just the airgap: every feature has its own optimizations, across every SKU, with the added complexity of a gearbox. Neither the US nor China will reach actuators with the precision to perform surgery on cells by doing process engineering with humans; what matters is automated design and process iteration—specialized CNCs with factory telemetry that automatically close the loop to cut better parts.

Endogenous Mammalian Octasulfur

Mammals store and use elemental sulfur as S8, the ring-shaped octasulfur form that is the bright-yellow form of the element seen in nature. It was already known that prokaryotes handle sulfur in this form, but no one had seen it in higher animals until now. Researchers detected the S8 signature using Raman spectroscopy and a mass spec method that captures the molecule in a host-guest form inside a polyaromatic capsule structure. That technique shows that the molecule is found in human mitochondria and in cellular lipid droplets, with levels reaching 100mM or more in cancerous tissue. A normal human carries around 10s of grams of elemental sulfur in their cells, synthesized as an independent product of endothelial nitric oxide synthase (eNOS), where its hydrophobic nature prevents lipid peroxidation as an antioxidant backstop to glutathione.

Fixed-Width Hot-Patching

Because AArch64 mandates fixed-length 32-bit instructions aligned to 4-byte boundaries, the instruction pointer can never land in the middle of an instruction during live execution. Windows exploits this by inserting 12 bytes of padding before every function entry point for an atomic 3-instruction trampoline using scratch register xip0, completely avoiding the multi-byte instruction-boundary decoding and 2-byte relative jump gymnastics required by variable-length x86.

Display Memory Tradeoffs

In 1964, Donald Bitzer invented the plasma display for UIUC’s PLATO system because cathode ray tubes required external computer memory to refresh the screen 30 times a second—an unbearable cost when memory was priced by the byte. Plasma’s bistable gas discharge cells acted as their own physical memory elements, eliminating the refresh bandwidth bottleneck. Once DRAM rode Moore's Law down the cost curve, the core economic reason for plasma vanished, leaving it to fight LCDs solely on picture quality until LCD manufacturing scale and yield learning curves wiped it out.

Compiling Matter from Code

Micromanufacturing has historically been trapped between traditional mechanical machining (limited to 10s of microns) and semiconductor lithography (planar, mask-bound, and restricted to silicon chemistry). Atomic Machines emerged from 6 years in stealth with the "Matter Compiler," an all-digital platform that fabricates 3D micro-machines with moving parts down to single-digit microns directly from code without masks, molds, or per-product tooling. Their first commercial device is the PrimeSwitch (PS-150): an electromechanical power relay for 800-volt AI data center architectures that opens in 50 microseconds, measuring 9.5 mm across and 3 mm thick.

Direct-to-Cell Beam Synchronization

Standard cellular protocols assume stationary base stations, causing packet collisions and dropped calls when communicating with low-Earth orbit satellites moving at 17,000 mph. Because Starlink uses phased-array narrow beams, a standard smartphone at the center of a spot beam is physically closer to the satellite than a phone at the perimeter, creating microsecond propagation discrepancies across beam footprints. SpaceX patent US 12,757,141 B1 solves this on the satellite rather than the handset: an onboard digital buffer dynamically calculates geometric distance for each active beam, holding packets from faster, closer beams so that signals across all beams arrive simultaneously with the slowest beam. This creates a uniform artificial latency plane that allows unmodified LTE smartphones to maintain synchronized connections from orbit.

Divergent AI Strategies

While the American AI strategy models development as a winner-take-all race toward artificial general intelligence through LLM scaling, Chinese policy explicitly rejects this framing. China’s AI+ Plan does not mention America or race dynamics; instead, it subsidizes open-weight foundation models via local compute vouchers to drive broad industrial diffusion into manufacturing and domestic supply chains. The divergence is between a centralized moonshot for frontier superintelligence and decentralized technological integration into physical industry.

Iron and Bronze Collapse

The question is not why Bronze Age empires fell, but why nothing like them returned for centuries. The causes of the Bronze Age collapse are debated — droughts, the Sea Peoples, internal revolts. But a decisive factor was iron. While the tin required for bronze was concentrated in rare deposits and easily monopolized by imperial courts, iron is the 4th most abundant element in Earth’s crust. In contrast to bronze, iron democratized violence. Because iron could be produced locally, imperial centers permanently lost their monopoly on the tools of war, ending the centralized equilibrium that had sustained Bronze Age states for 1,000 years.

Agentic Verification

Across 26 prompt conditions tested on Rust implementations (TLA+, Alloy, Verus, Kani, fuzzing), default unprompted coding outperformed most specialized verification prompts. Agents overwhelmingly wrote code first rather than modeling upfront: 75 out of 80 agents told to use TLA+ implemented in Rust first, then used TLA+ post-hoc on trivial functions where pass rates were already 99.6%, while completely failing to model concurrency, event queues, and state machines where the actual failures occurred.

Post-Proof Mathematics

We cannot suddenly move the goalposts and pretend that proving theorems was never the point, or that open problems were merely proxies for mathematical understanding. For quite some time, the implicit social contract has been that society pays us to exercise our intellectual curiosity and we, in return, provide skills and practical knowledge. In a world where models match or exceed humans at proving theorems, that social contract needs to be renegotiated with humility rather than entitlement. Even if this future supports fewer professional mathematicians, it could nevertheless support a much richer mathematical life.
2026-10-20: Settling the quasi-Riemann hypothesis via overlooked connections
Where the Riemann hypothesis posits that all the nontrivial zeros are on the vertical line at 1/2, the quasi-Riemann hypothesis asks to prove that they are all contained in a vertical strip narrower than the known strip between 0 and 1, between 1/8 and 7/8. The vast majority of consequences of the Riemann hypothesis already follow from quasi-RH. Contrary to what mathematicians had long expected, the proof didn’t need to introduce any new concepts — it instead unearthed hidden connections between old ones, approaching the Riemann hypothesis by considering the zeta function as 1 member of a broader family of functions using metaplectic forms. Mathematicians who study this branch of number theory were mistaken in thinking that their tools weren’t strong enough to resolve the quasi-Riemann hypothesis, convinced that proving it would require techniques from another branch of number theory entirely.

Premodern Urban Commerce

In second-millennium Europe, city growth correlated strongly with non-absolutist governance: because medieval feudal elites were rural rather than urban, city population served as a direct proxy for trade rather than imperial tax extraction. London expanded as a commercial river crossing before royal institutions centralized at Westminster, while Hanseatic towns grew along maritime grain networks as self-governing commercial republics.

Agents, Computers, and Volition

The key thing to understand about agents is that they are an AI with access to a computer: probabilistic LLMs leveraging deterministic computers to do computer things. When distribution became free, publishing became abundant and curation was scarce. With agents, the raw capability to execute tasks is becoming abundant; what remains scarce is human volition. While base models are substitutable commodities, agents improve with accumulated user context, and that persistent context becomes the dominant switching cost.

Browse All 10,819 Posts in Archives →