EDIT: I see that CarlShulman brought up essentially the same points, which I somehow managed to miss while posting this. I've left up the original comment for posterity but feel free to ignore.

a worker has more capacity than a donor does to learn whether small probability failure modes prevail in practice, and can switch to a different job if he or she finds that such a failure mode prevails.

I have some questions about this.

First, at least for the example small-probability failure modes that you gave as an example in your previous article, an individual worker would not be in a much better position to assess them than a donor (or at least, would be in no better position to assess them than a donor's inputs, like GiveWell or other third parties). Can you give some examples where workers would be in a better position?

Second, donors seem to be in a much better position than workers to react to failure. If you're doing something that hasn't been done before--which seems to be the place where direct work is most obviously better than earning to give--then you need to spend a lot of time figuring out things that nobody else has figured out yet. This means that the costs of switching to a different cause are quite high. On the other hand, a donor can simply change their beneficiary organization, which is much easier.