Alec Radford tests Hassabis’s AGI challenge with a model trained on pre-1931 data

哈萨比斯出的难题，GPT之父接上了：用一个知识停在1930年的模型

Alec Radford’s team trained 13B talkie on 260B English tokens dated before 1931. They tested surprise on nearly 5,000 historical events and used HumanEval for lower-contamination code evaluation. The key issue is time leakage: the 13B model still has vague post-WWII knowledge.

Why it matters: HKR-H/K/R all pass: the 1930 cutoff is a sharp hook, the post gives 260B tokens and ~5,000 event tests, and the finding targets data leakage. This is strong research, not a model or platform release, so 78–84 fits.