research/machine-learning
-
Essay
The Natural Order Hypothesis for LLMs All rights reserved
We propose a natural order hypothesis: language models trained on substantially shared information become able to solve problems reliably in approximately the same order, with training methods chiefly changing how quickly and how far learning proceeds. We postulate this weak form and conjecture a strong one, in which a single capability scale would also govern how often different labs’ models make the same discovery. Evidence from open mathematical conjectures and chess training runs shows often nearly nested solved sets, the cross-sectional signature expected from approximately common order, and substantial descriptive order agreement on one chess bank; the strong form has little support. We give one dated forecast and propose tests that separate the forms.
-
Essay
Verified Inference Between Adversaries Revised from 29 August 2026
An operator can fabricate execution logs and hold approved weights while running something else. VerInf investigates proofs of language-model inference that mutually distrustful parties can verify on their own hardware. This living document explains what the proof certifies, the system I started from, and my work on profiling, prover optimization, and scale-out during the MARS V fellowship.
Added Anthropic's prior-work citation; engineering status remains as checked on 8 October.
-
Essay
As we approach AGI, the increase in the ability of Artificial Intelligence models to infer a robust specification from a sparse prompt will lead to a devastating trend of homogeneity. We argue that this is the primary concern regarding the interaction of AI and human intelligence, rather than blanket claims that “AI reduces human cognitive ability.”
-
Essay
3D pose-estimation and kinematic-analysis system for neurological-recovery research, developed in Liqi Shu’s laboratory at the Brown University Department of Neurology. Python/TensorFlow inference, MATLAB-based statistical post-processing, Rust backend with HTML/JS frontends. Four externally-funded sub-projects since 2023; clinical-implications manuscript in preparation.
-
Essay
A deep learning model using ICD-10-CM diagnosis codes with a permutation-invariant Deep Sets aggregator improved 30-day unplanned readmission (AUC 0.7496 vs 0.6553 for CCI) and 30-day postdischarge in-hospital mortality (AUC 0.8557 vs 0.7844 for age-adjusted CCI) compared with Charlson and Elixhauser comorbidity-index benchmarks in a national claims database of over 113 million adult hospitalizations.