21 October 2026 — Idlewave

Computing. In an independent evaluation on twenty-nine months of hourly ERCOT prices, the t0 forecasting models cut lagged-price baseline error by thirty-eight percent. Independent evaluations of forecasting models are rarer than the leaderboards they are meant to check. Preprint.

$t_0$: A Time-Series Foundation Model for Forecasting with Context

Computing. One step of recursive harness editing improved GLM-5 and Gemini 3.5 Flash agents by up to 11.5 points on WorkBuddy-Bench. The model weights did not change; only the scaffolding around them did, and it was edited without a ground-truth checker. Preprint.

Harness²: Recursive Agent Harnessing for an Open World (Google Research)

Physics. A protocol using local projective measurements and unitary feedback prepares ground states of frustration-free gapless quantum systems. Gapless states carry long-range correlations, which is exactly what slow adiabatic preparation struggles to reach in finite time. Peer-reviewed.

Digital Dissipative State Preparation for Frustration-Free Gapless Quantum Systems

Physics. Exact unstable Navier-Stokes solutions, usually computed in minimal boxes, have been carried over to spatially extended turbulent flows. A minimal domain holds one coherent structure at a time; real flows hold many, and they interact. Peer-reviewed.

Exact Coherent Structures as Building Blocks of Turbulence on Spatially Extended Domains

Physics. Interaction cross sections were established as a quantitative structural probe of hypernuclei, benchmarked on the hypertriton. The hypertriton holds its lambda particle so loosely that its size has been hard to pin down by any method. Peer-reviewed, single author.

Precision Hypernuclear Structure from Interaction Cross Sections

Materials. A first-principles framework computes internal conversion rates in optically active spin defects, which common approximations underestimate by orders of magnitude. Internal conversion is a dark decay path, so it quietly sets how bright and readable defects like the nitrogen-vacancy centre can be. Peer-reviewed.

First-Principles Calculations of Internal Conversion Processes in Spin Defects

Materials. An analysis of the Kubo formula found a mechanism that may explain insulating out-of-plane resistivity in layered metals like strontium ruthenate. These materials conduct like metals in the plane and like insulators across it, which textbook transport theory had no tidy way to allow. Peer-reviewed.

Insulating Out-of-Plane Transport in Anisotropic Metals: The Case of ${\mathrm{Sr}}_{2}{\mathrm{RuO}}_{4}$ and Related Layered Oxides

Chemistry. A thermodynamically consistent machine-learning potential simulates electrochemical interfaces at constant applied voltage, treating electrons grand-canonically. Standard machine-learning potentials hold the electron count fixed, which is not how an electrode under voltage behaves. Peer-reviewed.

Thermodynamically Consistent Machine Learning Potentials for Constant-Potential Simulations

Mathematics. A family of polytopes called in-in zonotopes organises the contributions to scalar equal-time correlators in flat space. It continues the habit, begun with the amplituhedron, of finding that messy sums of diagrams are really the faces of one shape. Peer-reviewed, single author.

Geometry of In-In Correlators

Computing. A thirteen-billion-parameter model trained to predict several future tokens solved twelve percent more HumanEval problems than a next-token model. The extra output heads cost no training time, and models predicting four tokens also ran up to three times faster at inference. Preprint.

Better and Faster Large Language Models via Multi-Token Prediction (2024)

Materials. Long-range orbital effects in titanium come from orbital accumulation converted to spin current and back, not long orbital diffusion lengths. Long orbital diffusion lengths were the usual reading of titanium's reach; on this account the orbital signal hands the travelling to spin current. Peer-reviewed.

Orbitronics: Orbit-Spin Interconversions for Long-Range Orbital Propagation Enabled by Spin Current

Computing. A theoretical paper argues complexity has no causal influence on generalisation and only appears to through confounding. Simpler models generalising better is close to an assumption in machine learning; here the correlation is attributed to weak constraints taking simple forms. Preprint.

Is Complexity an Illusion? (2024)