Exploring Claude/GPT Knowledge Cutoffs and Pre-Training Timelines

Title: Exploring Claude/GPT Knowledge Cutoffs and Pre-Training Timelines Source: hackernews Probing frontier models with niche facts and token breakdown analysis allows for estimating parameter counts, dataset mixtures, and training timelines. These Incompressible Knowledge Probes and Data Mixture Inference techniques attempt to reverse-engineer pre-training checkpoints and distillation processes via official APIs.