I’m rushing this out so I don’t fail jimhaven day one
Let’s take
- initial time horizon
- initial doubling time
- doubling difficulty curve
Then project time horizon.
This model will not take into account AI contribution to AI progress. This will demonstrate clearly the importance of the possibility of a tendency to infinite time horizons.
Let’s define our parameters. Initial values are for Jan 1 2026.
Initial time horizon
Initial time horizon is METR 50% time horizon, but instead of tracking released model time horizon, it tracks internal model time horizon. This is because intelligence takeoff will happen internally. So, we adjust forward 66 days—this is METR’s estimate of how far ahead internal models were in early 2026. Let’s say 14 hours.
Initial doubling time
We’ll use METR’s estimate of 105 days.
Doubling difficulty
I’ll update this to explain doubling difficulty tomorrow