Basil Halperin Says OpenAI Research Time Horizons Double Every Two Months

Economist Basil Halperin calculates that OpenAI's internal AI research time horizons double roughly every two months, faster than METR's four-to-seven-month pace. Halperin estimates the 80% accuracy horizon at about 15 minutes as of July, and cautions the doubling figure depends on how the data is fitted.

• Halperin calculates a naive linear read gives a 68-day doubling time at 50% accuracy but 585 days at 80%.
• Halperin says a METR-style logistic fit instead yields a 6-hour 50% horizon doubling every 63 days.
• OpenAI's 80% time horizon runs about 15 times its 50% horizon, Halperin says, versus roughly 4 times at METR.