Set deep goal

Part of Explore domain. Why look farther into the future?

Said another way, do you plan to profit from the action in the long-term? That is, will it produce something reusable? If so, consider how valuable that value-producing artifact (investment) will be over its lifetime and how much it will cost to build. See also a discussion on the value in reproducibility in Estimate improvement value.

Future costs affect today

Because there is more than one way to solve every problem, you can often find ways to solve two problems with one tool (e.g. abstraction). By planning ahead, you can often skip intermediate solutions and go straight to what looks like the final solution.

You may not want to design a perfect solution when a better one is coming. Another person on your team or your company or the open source community may be designing a better solution. It’s often prudent to wait for others to automate and standardize rather than getting into complicated businesses while they are still evolving.

Examples

While it may have added value to write custom code to orchestrate production in 2010, most of your code would likely have been eliminated with the advent of containers and Kubernetes.

To dig back into your dependency chain to the hardware could likely only have been done with AWS in 2015, but by 2020 other tools like Terraform and GKE were providing similar services and stabilizing the cloud API. The MinIO interface was only created after AWS experimented with variations on the s3 API.

RNNs were the best way to solve NLP problems a few years ago, then LSTMs, and now Transformers are taking over.

Frequentist statistics was the right way to describe uncertainty through the 60s and 70s, but now Bayesian methods are taking over (this is “math” which you assume to be a stable topic to learn about).

Historical perspective

Said another way, it’s likely that (given stable human values) we’re doing a lot of guessing about the right next “step” in a particular domain. For example, we often see many deep learning papers based on roughly the same idea come out at about the same time, building on some recent breakthrough. These are like a variety of alternative “steps” that could be taken from the breakthrough; only time will tell which one is the best. Time may indicate that another step (that takes longer to discover) is actually better.

Learning older concepts is not without value, however, in the same way that failed experiments contain information (unfortunately there are many more ways to fail than succeed). Many pedagogical materials are essentially a compression of history, skipping over “less successful” results in an effort to focus on the big improvements (similar to a well-compressed git branch).

Are the tools you want to use still in the latest libraries on the topic, e.g. PyTorch? Library writers have a limited resources as well, and only write the most important functions.

Future needs affect today

Future features may change the design of the features you implement today. For example, you may discover you don’t need to do something if you do something else (that’s useful in another context as well) instead.

You don’t want to maintain code you plan to remove. If you don’t understand the value of features and company direction, you aren’t going to be able to make smart decisions about what code to invest more time in. You want to invest more time in code that will be around for eight years than one year.

Said another way, if you can predict what will be valuable (have value) in the future, you may be able to consider it as part of today’s plans. Stable values make planning easier because plans can be reused, and you’re less likely to make errors planning too far ahead.

You can look into your dependencies (who you buy from) in one direction, or into your dependents (who buys from you). If you push into your dependencies, you’ll be able to produce more efficiently (lower costs). If you push into your dependents, you’ll be able to achieve value more efficiently.

You need to not only “merge” information from the outside world about costs (e.g. how to solve a problem) but also about values. One purpose of Scrum is to make businesses more responsive to a changing market.

Time to brainstorm

When you eventually actually do work, you will have had many chances to look at it and sleep on the “how” of solving it if you’ve seen it in planning before. That is, sleeping on ideas (or having them in the back of your mind as you work) can help you generate more plans that reach the same state, perhaps more cheaply.

Test

In the words of Plan, extend the “time frame” of your plan. That is, take the time to add c (and G2) on top of b in this graph:

* c: G2
* b: G1
* a

Estimate cost

Increased uncertainty

Should you consider some time past when you are surely gone and completely forgotten, or only to your retirement? Should you consider your impact on future humanity, indirectly through family, coworkers, friends, etc? Year 2200? Or should you use some kind of future discounting? The farther in the future you’re looking, the more uncertain your estimates of the world will look are going to be. If you have less value uncertainty, and less cost uncertainty, then you can look a little farther into the future, but there’s a lot of world state you don’t control.

You really can’t plan anything out to 2200 only because you don’t know what will happen in the next 5 years with much certainty (war, hard AI, family, investments, etc.). I think looking forward 5 years is good enough for many problems, but with that attitude an undergraduate won’t be getting e.g. a PhD that takes 6 years.

How much uncertainty do you add looking farther into the future? It depends on the volatility of both your market, and the volatility of your dependencies i.e. supply chain (costs). Either way, it’s not necessarily linear. For a generic suggestion (without considering your domain), see Time value of money.

Set goal

See Set goal.