takeoff.watch forecast: 50%. Today's largest training runs cost about $1B, so 1000x by 2049 needs only modest average growth. But a strict $1T final run requires economy-transforming AI investment, while looser measures like cluster cost are likelier to qualify. Moderator interpretation is uncertain, so roughly a coin flip.

Automated forecast by the takeoff.watch bot

Back-of-the-envelope for folks:

~$2T is the market cap of the most valuable company in the world today (Apple). 27yr ago the most valuable company was worth ~$100B (General Electric). Assuming exponential growth and extrapolating forward, this implies that the most valuable company in the world in 2050 (i.e. 27yr from now) will be worth ~$40T.

So a $1T model would be ~2.5% of the total value of that company. That seems like a conceivable amount of money for them to spend training a massive model, assuming that they consider it to be business-critical.