A model like this shows you some constraints around what you can reasonably assume. If you assume a lower fatality rate, the implications are that either WAY more people have it than we think, or it's doubling at a much slower pace. For example, using a slighlty more simple version of his model if I assume a fatality rate comparable to the flu and a time to double of 6 days, roughly 1/3 of the US has already been infected and we're well on our way to herd immunity. Even in that scenario 120K people still die.

Now let's assume a more conservative but still optimistic fatality rate of 0.75%. In this scenario, about 5% of Americans have or had this if I assume a doubling of every 6 days. How many people end up dead? Still 120K. How can that be? Using this method, you're inferring how many people had it weeks ago based on how many you observe dead today. So a lower fatality rate assumption means many more had it it a few weeks ago, many more will have it now. So the fatality rate is lower, but it's multiplied by a proportionally bigger number and the absolute number of deaths is the same.

This model is just mathematical reality. You can change the assumptions how you like. What you have to take-away:

1. Low fatality rates imply either a long time to double or a really huge number who already have it. Flu-like fatality rates can easily project more cases in the US today than there are US citizens.

2. Although the models have assumptions, the assumptions are interrelated in the sense that changing one in one direction requires that you change another one in another direction or else your model quickly becomes unrealistic. Some people want to just make up numbers and don't realize the full implications of their made-up assumptions on all the other assumptions.