Imagine you have a paper list of your monthly expenses and want to find the total. A skilled comptometer (mechanical calculator) operator could perform this calculation much faster than is possible today. The reason lies in the input method and the operator’s skill. Entering numbers on a numeric keypad is a slow process, which is why barcodes and RFIDs are commonly used in shops these days. Comptometers allow pressing multiple keys simultaneously using different fingers. For instance, the number 233 can be inserted in a single three-finger press. Each number is immediately added to the total below, eliminating the need to press the “enter” and “+” keys. Comptometer schools graduated more than 20,000 operators per year1 in the 1920s, those skills are now lost.
Credit: Carlos Baquero
Let’s consider another profession, computer programming. Until quite recently, programmers acquired their skills by (ideally) starting with a mathematical and abstract problem-solving background and practicing programming. First, they would write simple programs and then gradually move on to more complex ones, learning how to break down and compartmentalize larger problems along the way. Traditionally, this was done in plain text editors such as VI and EMACS.
This traditional approach had some drawbacks. Once you’re familiar with writing a main function in C, such as “int main(int argc, char *argv[])”, it can be time-consuming to retype it in every new code snippet. However, manually typing the code and being lazy can also be beneficial, as it prompts the programmer to consider what’s truly necessary. In many cases, if the program arguments aren’t required, the code can be simplified to: “int main()”. The advent of Integrated Development Environments (IDEs) in the 1990s automated the creation of code scaffolds, which simplified programming but could also introduce unnecessary code fragments. Manual typing requires a certain level of frugality and emphasis on efficiency, skills that have certainly been impacted as programmers migrated to IDEs.
With the dawn of the new century, platforms like Stack Overflow brought about a transformative change. Programmers now had access to a wealth of recipes and solutions to help them solve coding challenges and incorporate them into their code. However, some programmers resorted to simply copying these solutions without fully understanding them or adapting them to their own code contexts. In C++, it became possible to work around the use of new approaches, such as migrating from printf() to std::cout, without grasping the concept of I/O streams. Nevertheless, the most skilled programmers embraced this new information from code snippets as opportunities to learn about new APIs and enhance their efficiency.
With each new step that helps automate code creation, the further away programmers’ stride from the initial roots of understanding the core principles of each language, conceptualizing the solution and the programming. A rift started to form among copy-paste, trial and error programming and programming from scratch. This separation would further accelerate after 2020 with the introduction of AI code assistants.
With large language models trained in vast code libraries, programmers can now specialize in providing elaborate textual prompts (the modern alchemical incantations) that spawn complete code solutions to a target problem. Or it seems so. Unless the problem is absolutely mundane, the produced code is often not 100% aligned with the intended task to solve. Someone needs to detect this and adjust it, and that person needs to be a programmer. But if someone specializes in prompt engineering and has less time for grassroots programming, they might lack the skills to detect and correct the mismatch. What we might get is Stack Overflow copy-paste on steroids, now for whole projects.
Craftsmen and Vibe Coders
The term “Vibe Coding” was recently coined by Andrej Karpathy. It refers to a coding approach that involves presenting natural language descriptions of a problem to solve and relying on an AI for the actual code generation4. This contrasts with the traditional Craftsman approach of programming by hand from reasoning and memory.
The rift among the two types of programmers, Craftsmen and Vibe Coders, will keep growing. As with someone, able-bodied, who faces the decision of taking the elevator or the stairs, knowing the impacts on fitness and cardio-vascular health, each programmer now faces the decision of how much to delegate to an AI or craft by hand.
Vibe Coders are opting for the easy way out, even if it hinders their learning and cognitive development. If this choice is made early on, they might not even learn how to program. We are already witnessing instances of this in some undergraduate courses, where individuals managed to deceive their way in without any proficiency in writing or programming. This problem extends beyond academic evaluation and work performance reviews; it’s a detrimental strategy for the individual themselves. If one becomes merely an AI supervisor and provides minimal additional value, they risk becoming mere cogs in a machine, to be supervised by an AI.
Technological advancements don’t necessarily lead to human decline. For instance, we wouldn’t consider a Formula 1 car driver to be less skilled than a racehorse jockey. Similarly, a carpenter using power tools will become more efficient, but they’ll likely still do some finishing touches by hand. Programming is also a craft that requires training and experience. The Craftsman approach in programming will still choose AI when faced with a task that doesn’t hinder planning and thought processes. Prompting can become very precise and include the solution strategy, and use the AI to automate contextual application to the code, followed by careful quality control. All this requires a deep understanding of the solution and how it can be programmed. Sometimes, the AI might still show a better way, which can aid the learning process. (Some of the best programmers in the world are actively learning how to further improve themselves with AI integration. The Coding with AI series2 by S. Sanfilippo is a great example.)
A balance must be struck between AI assistance and manual programming skills. Coming back to the carpenter metaphor, carpenters don’t ask the power tool to make them a nice table, but instead use it to cut pieces in precise preplanned dimensions. Otherwise, they would not be carpenters and would soon be replaceable.
The Last Solo Programmers
As time passes and generations are replaced in the workforce, there will be less and less of the first generation of programmers that learned from reading a programming language book or attending programming classes and gradually developed their skills by tackling problems of increasing complexity. History has taught us that skill sets often disappear faster than their relevance in deployed systems. Just consider all the Cobol code that is still in production.
It is important that the new generations of programmers make their own balanced choices on how much to depend on programming tools (the same applies for writers, as the note below attests). As always, the problem is never about the tool but about how we use it. If this is done well, we will not see programming become a curiosity skill of the past, like the 1920s Comptometer operators. There is plenty of future in the programming profession, even if transformation is inevitable as it always has been in this industry.
Notes: This text was planned and written by hand; then proofread, and in cases, paragraphs improved by Apple AI writing tools. The draft was reviewed in Perplexity.ai with the query “Read this text and tell me if any important point is missing and how to improve it,” followed by minimal manual improvements from this feedback.
Acknowledgements
I want to thank Cristina Videira Lopes, from the University of California, Irvine, for bringing my attention to the new term “Vibe Coding” and António Coutinho, from Eurotux.com, for discussions on the topic.
References
1 The Comptometer Operator. John Wolff’s Web Museum. https://www.johnwolff.id.au/calculators/comptometers/Operator.htm
2 Coding with AI. Salvatore Sanfilippo. https://www.youtube.com/watch?v=_pLlet9Jrzc
3 The End of Programming as We Know it. Tim O’Reilly. https://www.oreilly.com/radar/the-end-of-programming-as-we-know-it/
4 Vibe Coding. https://en.wikipedia.org/wiki/Vibe_coding
Carlos Baquero is a professor in the Department of Informatics Engineering within the Faculty of Engineering at Portugal’s Porto University and is also affiliated with INESC TEC. His research is focused on distributed systems and algorithms.