I’d like to present a simple argument about the ineffectiveness of a specific type of research, whose results would not have immediate applications upon completion. The argument could be useful to anyone who is interested in evaluating similar research projects.
The only assumption required by the argument is that, in the future, some kind of research technology (RT) will enable us to do the type of research in question, or maybe even all kinds of research, much more efficiently than today. For example, the technology may save us time or other resources.
Consider a research project satisfying the inequality:
where
- is the time at which the research technology becomes available
- is the interval of time that is necessary to carry out the research project if RT is exploited
- is the time at which the results of the research project will be applied.
I claim that it is inefficient to start now a research project satisfying the inequality: it could simply be postponed after the research technology becomes available, because the results of the project wouldn’t be used before even if it was completed earlier than .
Take the following fictitious research project as an example: Project Uranus aims at understanding how we could best exploit the natural resources of Uranus, once we’re able to reach it. Given current technologies, the production of a satisfactory solution to the problem requires 5 years. Let’s assume you believe that, relatively soon, a research technology that would speed up Project Uranus will be developed—this could be something like a general-purpose AI, or a more limited AI technology that would significantly improve our modelling capabilities. Assume you also believe that we won’t be able to use the natural resources of Uranus until long after that research technology is developed—this could be because using those resources requires developing advanced space-travel technology. In that case, according to the presented argument, you would conclude that it is better to postpone Project Uranus after RT becomes available.
Applications
In practice, things aren’t so neat: we don’t know the specific values of and , and sometimes the results of a research project will be useful to various degrees at different times, making also difficult to estimate.
However, the argument can still be used in the evaluation of research projects. For example, doing research on some specific long-term scenarios now may be inefficient, if the results of such research can't be applied soon enough.
I prefer to not give concrete examples for two reasons:
- I don’t want readers to narrow their attention on the evaluation of examples; instead, I want to put the emphasis on the argument itself
- I don’t think I’m the right person to do it: I have a mathematical background and I started studying AI safety somewhat recently. My knowledge in other domains such as strategy and economics is really limited.
If you have in mind a specific example that you believe it’s worth to point out, you’re welcome to leave a comment about it!
This work was supported by CEEALAR (EA hotel). Thanks to Michael Aird and Kris Gulati for feedback.
Another consideration might be that if you expect this research technology to be developed but had not taken that into consideration for estimating your impact, you may be underestimating the likelihood that someone else would do that same research before tapp anyway since if the research becomes easier, others are more likely to do it. You could be overestimating your counterfactual impact.