Blog

In the age of AI, the ability to evaluate, judge, and appreciate is becoming the new form of productivity. Human definitions of 'good' and 'beautiful' are reshaping the essence and value of creation itself. From the complete creative loop of the craft era to the evaluation-driven paradigm of the AI age, we're witnessing a fundamental revolution in how we produce and create.

Multi‑epoch training wastes time on easy, well‑learned samples. InfoBatch dynamically prunes data and rescales the loss to keep accuracy while speeding up training by 20–40% across vision and language tasks.

Continual fine‑tuning of vision‑language models can damage zero‑shot transfer. ZSCL adds simple constraints in feature and parameter space to keep zero‑shot ability while improving downstream performance.

Training large language models uses a lot of memory. CAME cuts memory use to Adafactor levels but keeps Adam-like performance.

Large language models (LLMs) possess a remarkable ability to anticipate the length of their generated responses. By leveraging this capability, we propose a novel technique called Sequence Scheduling to enhance the efficiency of LLM batch inference.

A step-by-step derivation of the backpropagation algorithm for multi-layer perceptrons.