I made a spreadsheet to conservatively project the cumulative years of continuous driving (both fully manual and partially autonomous) that Tesla’s fleet of HW3 cars does. The formula for each month is (1 hour * 30 days * the number of cars) / 8760 hours. Then you add up the result for each subsequent month. I start counting in April 2020 — by which time, hopefully the new neural networks designed for HW3 will be at least running passively in shadow mode on the fleet. At that point, the HW3 fleet will be doing over 1,000 years of driving per month.
I didn’t account for any retrofitted HW2 cars in this spreadsheet. I assumed production of 105,000 cars per quarter (8750/week) until July 2021, when I added 4000/week Model Ys.
Google Sheets: Tesla HW3 Fleet Years of Continuous Driving (projected)
The reason I care about this is imitation learning on the fleet. For comparison, AlphaStar Supervised was trained using pure imitation learning on human-played StarCraft II games and it is better than 84% of human players.
DeepMind used about 1 million games to train AlphaStar Supervised. At an average game length of 25 minutes, that's only about 50 years. As I mentioned, in April 2020, the HW3 fleet will be driving over 1,000 years per month.
If 0.5% of driving is used in imitation learning, 50 years will be used by October 2020. If 0.1% is used, 50 years of driving will be used by December 2021. If 0.05% is used, the training dataset will get to 50 years by October 2022.
Driving and StarCraft are, of course, disanalogous in multiple important ways. I don't think we can assume the same number of years of imitated behaviour will result in the same level of competence. This is just the only example of imitation learning I'm aware of that is remotely comparable in scale.
I didn’t account for any retrofitted HW2 cars in this spreadsheet. I assumed production of 105,000 cars per quarter (8750/week) until July 2021, when I added 4000/week Model Ys.
Google Sheets: Tesla HW3 Fleet Years of Continuous Driving (projected)
The reason I care about this is imitation learning on the fleet. For comparison, AlphaStar Supervised was trained using pure imitation learning on human-played StarCraft II games and it is better than 84% of human players.
DeepMind used about 1 million games to train AlphaStar Supervised. At an average game length of 25 minutes, that's only about 50 years. As I mentioned, in April 2020, the HW3 fleet will be driving over 1,000 years per month.
If 0.5% of driving is used in imitation learning, 50 years will be used by October 2020. If 0.1% is used, 50 years of driving will be used by December 2021. If 0.05% is used, the training dataset will get to 50 years by October 2022.
Driving and StarCraft are, of course, disanalogous in multiple important ways. I don't think we can assume the same number of years of imitated behaviour will result in the same level of competence. This is just the only example of imitation learning I'm aware of that is remotely comparable in scale.
Last edited:
