Learning and Doing — what’s the balance?
In school we don’t do enough, at work we don’t learn enough
Contents
∘ Adults Don’t Learn Enough
∘ Learners Don’t Do Enough
∘ Why learning without doing is harmful
∘ Synthesis — Learning by Doing
∘ How can we balance learning and doing?
∘ Conclusion
Having been in adult working life for nearly a year, I’ve often found myself balancing the difficult tightrope between learning and doing, while navigating the complexities of work and personal development. This balance is essential to an effective life, yet it is so elusive.
This theme has preoccupied me recently, so I felt compelled to pen my thoughts down in this article — talking about why we lean towards learning or doing at different life stages, and the benefits of finding a balance between both. I then offer a few tips that have proven helpful in my journey so far. I admit I myself haven’t perfected this balance (my frustration with my failure to do so sparked this whole introspective journey anyway).
Let me briefly introduce myself — I work as a public sector economist, but also build AI products on the side, such as Podsmart AI, a podcast summary app. I have only started working full time last year, and am definitely in the phase of rapid learning on-the-job, with an exciting future of opportunities and options available for me to define. Rapid advancements in technology such as AI make it even more crucial these days to learn continually even while working, as new techniques that can 100x your productivity, or lucrative new opportunities, are constantly being discovered.
Adults Don’t Learn Enough
As working adults, doing is easier than learning, and we do not learn enough. When we start working, we join established organizations (banks, consulting, big tech or the government) and are given tasks and responsibilities to serve the organization.
The first reason why we do rather than learn, is that organizations face immense pressure to maintain the status quo, which results in an over-emphasis on ‘business-as-usual’ work that perpetuates what is already known and tested, leaving little room for innovation.
This is definitely true of my AI endeavors — there always seems to be a never-ending list of bugs to fix! As a result, I’m often unable to invest time into improving the RAG and summary algorithm — which are nearly a year-old and definitely outdated, given the rapid innovation of AI question-answering techniques.
I also find examples of this in my full-time job. A colleague, who had made significant contributions to a major project on modeling future economic resource needs, now finds himself inundated with inquiries from various agencies about the model’s outputs. His days are consumed by managing spreadsheets full of data and responding to endless emails, which considerably delays his progress on research projects.
Why isn’t he frustrated with this stagnation in research? This brings me to my second point: incentives. In many organizations, there is a pronounced bias toward upholding the status quo, which is often more rewarded than innovation. Engaging in work that supports the ongoing success of established processes leads to immediate, certain, and tangible results, especially if the status quo is the centerpiece contributing to the organization’s success.
This incentive system leads to many being blind to the benefits of innovation on overall knowledge and productivity, as they chase the reward signal toward upkeeping current ways of getting things done. In fact, this colleague was recently promoted. Furthermore, it can blind people to their own value-add. As an economist, I contribute value to my organisation by injecting economic rigour into policy decision-making. I had recently suggested to this same colleague a method of estimating productivity that was economically more rigorous than the current method. Yet he preferred the flawed current method over mine, as the workstream’s output had been agreed on by senior management.
Learners Don’t Do Enough
On the other hand, during periods traditionally associated with learning, e.g. college, or when you take an online course to upskill while working, we have no stake or responsibilities in maintaining the status quo, freeing us up to learn.
In such periods, we tend to not do enough, as doing takes more effort. An obvious example myself and many college students can relate to — merely flipping through the textbook to prepare for an exam the next day. I would mostly be lazy to do the practice exam, but in the few times I did so, the learning I internalized, through correcting my mistakes and putting my mind through the motions of actually thinking, was orders of magnitude more effective.
I can think of high school — those dreaded lessons when the teacher gives us to do a practice exam to complete in-class. Though those were painful, it definitely helped us to learn.
An example more relevant to adult life is when we take Data Science or ML online courses to upgrade our skills. The tendency is to just run the sample code and existing examples (Boston house pricing or fraud detection), rather than adapting the code to do an exploration that you are interested in.
Additionally, in these cases, doing seems like the riskier, uncertain option — an inversion of the workplace where learning is riskier. Running the pre-made Colab notebook is guaranteed to work. What isn’t guaranteed to work is trying to adapt the Colab code into personal explorations. So many times I have endured this frustration. That of searching through confusing documentation or data APIs, dealing with bugs in the package or incompatible versions, and so on. Entire days can consumed in investigating, often being fruitless, and you start to question your life for why you decided to be brave and go off the beaten path.
The insidious aspect of learning is that it can sometimes be a form of procrastinating from doing too, tricking us into thinking that we’re progressing by consuming educational content but not retaining any of it.
I know I spend too much time scrolling thru Twitter, bookmarking tweets on interesting AI models or methods to read ‘later’ (more like never), or listening to too many productivity or self-help podcasts on how to build an audience or build successful products.
Ryan Holiday said this in an interview on Ali Abdaal’s podcast (summary here) — “the irony is that obsessing about productivity is a form of procrastination, it’s giving you the sense that you’re serious, that you’re making progress, but you’re not. You’re avoiding the hard thing, which is doing the thing”.
Why learning without doing is harmful
Firstly, as I alluded to above, we often fail to internalize the knowledge we learn by just reading. The implications of this failure is particularly acute in technical disciplines — where you need deep understanding of the matter to be useful at all.
For example, while writing the proposal to a research project at work, I had read a slide explaining the staggered Difference in Difference econometric regression, and had thought I fully understood it. Only when I had to implement the methodology in code did I realize that so many subtleties (what year as the base year? calendar time or event time fixed effects? etc.) Another example is when building software and using packages with high level of abstractions (many different modules and classes).
Secondly, learning too much without doing could risk irrelevant learning, which is a waste of time and mental effort. When taking Multivariate Calculus as an Econ major, we get taught completely irrelevant things like Stokes and Green’s theorem. Because college classes are catered to students of broad interests and disciplines, most of them do not suit your specific interests or needs. Even taking an online AI/ML course (because everyone says “it’s important to learn about AI now”) is a surefire way to irrelevant learning. There are many subfields of AI — computer vision, traditional ML methods for tabular data, and natural language processing. Even within NLP, there are many applications e.g. story generation, info extraction, summarization. I’ve learnt that I’m much more interested in NLP, and summarization and information extraction, and so hone in on those resources.
Of course, there are cases, such as introducing students to a new discipline, and learning for entertainment or self-fulfillment sake, where a broad range of potentially irrelevant learning is desirable, but this applies to the minority of the cases.
Lastly, learning without doing could lead you to learning the outright wrong thing — especially in rapidly-evolving and general fields like AI.
One common case is that who make ML/AI/software educational resources sometimes come from companies who want to promote their software tools.
In other cases, these educational resources optimized for the average audience. Another example I alluded to in my earlier post, is that people tend to make YouTube coding tutorials centered around demos, probably catered to college students or people who want to pick up web dev as a hobby, hence teaching you Bootstrap, React, Javascript, manually writing fetchers etc. This advice is detrimental to those wanting to learn production-grade software practices.
Additionally, people on social media often say sensationalist or controversial things that may not be fully true. When Gemini 1.5 Pro launched, people would have you believe that RAG has become obsolete due to the 1.5 million context window, but I can assure you that this is not the case. Only by doing, or being intimately involved in building an AI product, will you realize that RAG is so cost-effective and fast that it will definitely not go away just because of million-plus token LLMs.
Synthesis — Learning by Doing
If we sometimes tend to do too little, and other times learn too little, what is the solution? It is to maintain a balance between learning and doing.
Unfortunately, there is no law written in the stars of the optimal balance — it differs across people and in different circumstances.
This dilemma between learning and doing has parallels in the exploration vs exploration trade-off in Reinforcement Learning (RL). One common method when training RL models is the epsilon greedy strategy — take a completely random action with probability epsilon, and take the current optimal action the other times. By varying epsilon over time, this strikes a balance between learning the best strategy, and executing the learnt strategy to achieve rewards from it.
However, there are subtle differences between RL and our life — unlike an RL agent which can be trained for millions of epochs, our life has finite epochs (think years), hence the learning vs doing trade off is especially important. Also, better ways of doing things emerge in later epochs i.e. innovation, hence we cannot just decrease our rate of exploration over time. Lastly, in life we optimize many things (e.g. cost, time, satisfaction, sustainability), so it’s not clear how to evaluate success.
Merely saying that we should balance learning and doing does not say anything about the interaction between them. The better answer, then, is to learn by doing — maintaining an iterative cycle between learning and doing.
A better ML parallel to the interaction between learning and doing is the Expectation-Maximization (EM) algorithm used in unsupervised learning. Briefly, the E-step estimates probabilities of data points belonging to certain classes based on initial parameters, while the M-step recalculates parameters to maximize likelihood of data given class relationships.
Learning corresponds to the M step, while doing corresponds to the E step. Learning is effectively us adjusting our knowledge parameters or mental models, hence corresponds to the M step. Having learnt all this new knowledge, we then move to the E step, employing existing parameters to do our work effectively (classifying data points in this analogy). The process then loops back to the M step, taking the feedback from E step to iterate, eventually converging on an optimal solution that couldn’t have happened without the two. (Aside: When I asked ChatGPT to relate learning and doing to the EM algorithm, it suggested that E represents learning and M represents doing. Hence, you can view it the other way round too!)
When you use your doing to help you decide what to learn, you gain actionable knowledge relevant to your use case, reducing the risk of wasting time on irrelevant learning (that you would likely forget anyway! Knowledge depreciates fast when it’s not used). And by doing, or acting on what you learn, you internalize the learnings, as well as improve your work processes by producing work more efficiently or of a higher quality.
How can we balance learning and doing?
Let me share 5 tips which I am trying to incorporate into my life:
Seeking out work projects that are opportunities for learning. For example, deliberately choosing projects which would use methods that I’m not too familiar with. I’ve not had experience doing microeconometric studies, and hence took a recent research project as an opportunity to deeply learn difference-in-difference regressions and the latest advances. Because they are work projects, they will necessarily involving doing, and doing relevant stuff to the organization, hence this is guaranteed to achieve a somewhat desirable learning-doing balance. This tip is paired with the next tip of…
Working with and learning from competent colleagues willing to learn and innovate. Some colleagues tend to think of creative and inventive new solutions, and so I try to learn from them and join their projects. A counterexample is some colleagues who are competent in the status quo methods with a low ceiling for innovation potential. For example, one colleague prides himself on his superior Excel VLOOKUP skills, many times patronizingly asking me why I don’t know VLOOKUP, and using Excel for increasingly large datasets. I could choose to learn VLOOKUP, or choose to learn data science and AI skills to effectively scale up and automate these workflows.
Approaching learning with ‘what can I do with this new knowledge?’ by drawing connections to existing work. This is paired with the next tip of …
Maintaining a breadth of learning sources — e.g. following subject matter experts on Twitter/X (which I use to learn about macroeconomics/finance, as well as AI), having friends in a range of disciplines (e.g. SWE in big tech, startups, lawyers, PMs etc.) and engaging them in curious conversations about their work, listening to podcasts (interviews with experts give amazing insight into their mental models and worldviews).
These two tips have exposed me to a wide range of potential things to learn, enabling me to filter down to a small group of things to learn, based on what I think will help me the most in what I’m currently doing (e.g. focusing on latest advances in reranking, because I want to improve the RAG algorithm for my AI app), as well as sparking new ideas of things I can potentially do in the future (e.g. seeing many tweets using LLMs for knowledge graphs inspire me to build an AI app to parse essays into knowledge graphs).
Experimenting with ‘doing sprints’ — when encountering ideas from your learning, being able to quickly hack together a proof-of-concept is useful — in proving feasibility, getting feedback, and also finding out if you’re interested in taking this idea further.
The beautiful thing is that it’ll get progressively easier to do these sprints! Over time, I’ve learnt to become more efficient in these ‘doing sprints’ — e.g. knowing where to find sample code, the existing frameworks to rely on, enable me to spin up an Interview Coach AI prototype in 3 hours, compared to weeks initially building Podsmart AI.
Just being mindful of learning vs doing in your life. Finally, just reading this article is already a step in the right direction! It has (hopefully) gotten you thinking about the balance of learning vs doing in your life and how the above suggestions may or may not apply. This is similar to how being aware of the importance of a balanced diet (e.g. of proteins, carbs, vegetables) nudges us into generally making better food choices.
Conclusion
Learning vs doing is a balance that we will probably struggle with through our whole adult lives. Curious to hear your thoughts on this mental model, as well as your worldviews about self-improvement? What is your optimal balance of learning and doing? What strategies have helped you to achieve this balance?
If you liked this post, please consider following me on Medium or Twitter/X — I do more AI / economics / personal productivity content!
