↓Skip to main content

Caihao Cui

Caihao Cui

Knowledge Grounds Action; Learning Nurtures Virtue.

Recent

Abstract response streams split under a reward signal, with one path following a constrained frontier and another drifting away from the true objective

Higher Scores, Worse Models? The Theoretical Limits of LLM Alignment

15 November 2026·4233 words·20 mins

An industrial monitoring screen peels away a single-score panel to reveal diagnostic evidence: a matrix, residual curve, distributions, and cross-validation traces.

Misreading Charts: How Correct Numbers Lead to Wrong Decisions in Machine Learning

14 November 2026·5214 words·25 mins

Conceptual TimesFM 3.0 rolling retail experiment with four demand scenarios, multiple forecasting models, and prediction intervals

Time Series Foundation Models - Google TimesFM (Part 2): Synthetic Retail Experiments and Forecast Analysis

8 November 2026·3847 words·19 mins

Conceptual architecture of Google TimesFM: from 1D continuous patching and local normalization to asymmetric long-horizon output heads, 3.0 CPM multivariate decoding, and quantile prediction intervals.

Time Series Foundation Models - Google TimesFM (Part 1): Model, Algorithms, and Engineering Practice

7 November 2026·7767 words·37 mins

A crystalline decision tree forest stretches across the landscape, with an optical Tower of Babel rising against the distant horizon

Should We Build a Foundation Model in a Forest of Trees? A Pragmatic Guide to Time Series Foundation Models

1 November 2026·Updated: 3 November 2026·13629 words·28 mins