How Planning Helps

Longtermism is an ethical stance which gives priority to improving the long-term future. (More)

Recently some have criticized longtermism, saying that it is often quite hard to have much confidence in how our acts today effect the distant future. Others have noted that it is also often hard to have much confidence in how personal acts effect global outcomes soon, or in how they effect personal outcomes decades later. All this is said to argue against consequentialist accounts of responsibility, compared to rule based accounts.

Many also say that “strategic planning” at big firms today is too detailed and infrequent; they’d do better with smaller more frequent problem-dependent plans. Big charities also do too much planning, it is said. Many say that engineers design too much, instead of iterating prototypes. Recently “agile” software engineering has been replacing “waterfall” planning with rapid prototyping. Very large planned projects seem to consistently fail. (I first learned of this in Freeman Dyson’s ’88 book Infinite in all Directions.)

Nuclear over-regulation seems mainly due to excess demands for planning. Law often excuses firms from harms if they had a plan to avoid them, even if those plans were ineffective. For example, ineffective sexual harassment training is seen as reducing liability in cases of actual sexual harassment. Politicians are often criticized for “not having a plan.” The bias toward an “inside” over an “outside” view also seems to be due to trusting a plan over background stats.

All of this seems to suggest that we are often overly inclined to rely on overly complex plans. Relative to the alternative to repeatedly doing stuff and then reacting to seeing what happens as a result. Maybe we do this because planning helps more to impress each other, or to justify ourselves to each other.

But, that doesn’t mean that planning is useless. Consider these quotes:

Plans are useless, but planning is essential. – Eisenhower

The value of a plan lies in the act and effort of planning: in doing so, you gain understanding about your client, the project’s goals and objectives, and the abilities of your team. Planning encourages situational awareness through learning and discussing strengths, weaknesses, opportunities, threats and a flexible strategy that can tie them all together. A good plan establishes a vision, with goals and supporting objectives, and provides context and directionality so the team can move forward and be supported without being locked in. (More)

In fact, it occurs to me that planning is useful even in the extreme case of our long term global future, where concrete plans seem the least useful. The reason, I think, is that planning aids the process of reactive adaptation, i.e., trial and error. Successful trial and error requires not only that you try stuff and see what succeeds, but also that you be sufficiently well focused on the right parameters of the trials and success. Only then can you notice the key correlations between trial parameters and success, so that you can learn which direction to set the parameters to get success.

Thus planning can be useful just because it helps you to to think about what are the important parameters that you should be tracking and comparing to success. And for this purpose, you actually don’t need to have any confidence that your plans can predict which direction to act in order to promote which outcomes. All you need to think is that the parameters that turn out to stand out as relevant in a planning process correlate with the parameters that you should be tracking in order to make use of trial and error.

Note that my main personal criticism of recent AI risk efforts is that now is too early to do much; we should save resources until we see more concrete examples of big AI control problems. It is fine for a few to think about AI risk now, as a way to get clues on what parameters we should start to track regarding AI products, activities, and outcomes. But we should expect to mostly have to wait until we see concrete things go wrong, try things to fix those problems, rinse and repeat.