AI Safety Spending in Two Fake Graphs

Thanks to Boaz Barak for the inspiration.

The normal philanthropic foundation scales up, plateaus, and then spends down, often ideally over 20 years, though some take longer.

I claim that AI spending should be almost exactly the opposite of that: an inverse plateau. Spend as fast as you can now on safety before the most dangerous time (the dawn of superhuman AI), then save until very late and spend more when technology is radically better (and ideally after you’ve accrued lots of interest) because at that point there will be much better things you can do with your wealth.

You may be wondering why “AI safety spending by default” is a gradual ramp rather than the foundation plateau. This is because most philanthropic foundations become active in spaces with very mature ecosystems so it’s easy to spend down. But philanthropic capital coming online for AI now will dwarf the existing ecosystems, making it much harder to scale up spending quickly. So I expect by default, without extremely radical interventions, AI safety spending will look like a gradual ramp for the next 20 years (assuming we have that many) instead of a plateau.