In October 2026, Attention Is Still All You Need

Work in progress. The section on consistent engineering principles is still to come.

It has already been six months since the last update to Undertide. It was not until now that I realized my speaking and writing skills had deteriorated to an unbelievable level: even expressing a basic idea properly without making grammatical errors or misspelling anything seemed impossible. If the CELPIP test weren’t approaching, I would never have realized that things had become this bad. However, spending time practising writing and speaking just for the sake of a good score on a language proficiency test feels miserable to me. What gives me the biggest headache about these questions is figuring out what they are actually asking me to say, rather than the English itself. Even if you asked me to answer these questions in my mother tongue, Chinese, I would be stuck for a few minutes without knowing how to give a proper answer.

I haven’t cooked a single meal since the spring of 2025. I also haven’t had conversations with people about life skills, relationship advice, urgent everyday matters, and so on. When I’m asked about these topics, I don’t even know how to answer properly. But the ability to express your ideas and thoughts is universal. Regardless, you will need it sooner or later. It has become even more important in 2026, when you can use natural language to turn ideas into actual implementations, thanks to the rise of agents. However, the upper bound of their capabilities remains your intelligence rather than theirs. Your cognitive resources are limited, which means the amount of input you can handle is limited. You could have thousands of threads running simultaneously, but you would only be able to verify a few of them a day. Do we have a solution to this problem? My answer is that we don’t. We are only human, and this is part of life. This brings me to the topic I want to discuss today: in 2026, what skills make a human stand out and go beyond what used to be possible? If the next-best solution is to maximize what agents can achieve with your involvement, I do have an answer: spend time and effort practising your ability to express yourself in English and applying consistent engineering principles rooted in your thinking. If you only want the answer rather than my full reasoning, those are the takeaways.

I have wanted to write technical articles for a long time. Because of academic pressure and a whole lot of other messy matters, I didn’t get a chance to do that. I either let my agents write articles for me, giving them only the basic skeleton of my ideas, or roughly translated Chinese articles I had written into English. Of course, I have deleted literally every one of those articles. I’ll get back to the reason later. Anyway, this may be the first time I have intentionally drafted a technical article directly in English without including concrete code or sharing a specific technical skill. Writing this article is itself practice for the first half of my answer. This is new to me, but it is definitely worth remembering and practising more in the future.

Attention Is All You Need

As the title suggests, I want to talk about attention. I know this may sound like more slop: it feels as though the slogan has been flooding the internet since the 2017 paper “Attention Is All You Need” was published, changing the industry’s direction in the years that followed. I’m not referring to softmax or QKV here. Attention simply means your attention. More precisely, I mean your cognitive resources, which are consumable, cannot be replenished within a given period, and are extremely limited. As an analogy, attention is like the gasoline in your car. You want to reach a destination, and your fuel determines how far you can go before getting stuck halfway there. When you run out of attention, your brain stops taking in further information meaningfully, or you may even begin to make poor decisions. Never make decisions when you have run out of attention: that is the basic principle I want to establish at the beginning of this article. Given this premise, what can we do to allocate attention in order to maximize productivity? As I said earlier, your ability to express yourself deserves the largest share of your attention.

The ability to express yourself

What is the ability to express yourself? Is it equivalent to communication? In my definition, not quite. It consists of three parts: your knowledge, your ability to describe something, and your ability to connect the two.

My background is in mathematics and statistics. However, in the final year of my studies, I took a course at U of T called Software Design, with the course code CSC207. Although it was taught using Java and related concepts, the course placed a strong emphasis on technical and non-technical communication. The same was true of my time series analysis, data visualization, and many other statistics courses. At first, I thought it would be easy to explain something I knew to other people. It seemed intuitive that I could naturally express what I had learned and understood. Then I found myself repeatedly struggling to do so. After thousands of failures, I realized that understanding something does not mean you can express it well. But if you can explain it in plain language, technical language, or a combination of both, using your own words, you have a solid understanding. This reminds me of the Tower of Babel: meanings can diverge even when people refer to the same thing. Tiny nuances in wording can affect your communication and lead others to interpret the same idea differently. To prevent this and make our communication more precise, with fewer misunderstandings, we introduce terminology. But terminology is a collection of abstractions whose meanings can differ across contexts.

For both established terminology and your own descriptions, what you ground them in determines their meaning. The third element you need is concrete examples. In my experience, keeping abstractions and implementations clearly separated, with distinct roles, is crucial to clear communication. The interpretation of the word “aesthetics” can depend entirely on which of two different examples you refer to. But to avoid repeating large amounts of concrete detail, you need abstractions. When another set of rules or styles must apply to the same category of objects, you can extend those abstractions with details, just as we define Java classes and interfaces, extend classes, and implement interfaces with concrete classes. If you have no idea what I am talking about when I mention Java, it may be harder for me to explain this using other examples. But for someone who has already learned Java, I can skip countless examples and explanations, and we can still understand each other. How do you know what a Java class is? You learn about it. How you learn about it is not the topic of this article. What learning takes is your attention. How do you express things clearly to your agents and human teammates? You start by understanding the subject and knowing how to organize your thoughts. What’s more, you use language to organize thoughts that are supported by that knowledge. Together, these three skills produce a clear, deliberate, and precise expression of an idea that can be turned into an implementation and checked against the idea in your head. I still need to emphasize that what you think and what you express are fundamentally two different things. No one other than you knows what you are thinking, but people can interpret what you express and form their own understanding. The difference between your original thoughts and their interpretation is something you need to minimize. Bringing them as close together as possible is one purpose of your attention. When the recipient is an agent rather than a human, the comparison narrows to your thoughts and the result. There can be differences between thoughts and results, and because these are different kinds of things, some difference is inevitable. Yet a result that differs from your intention but can be revised into the desired form is very different from one that is unpredictably different or even misleading. Fortunately, the ability to express yourself can help you address the first situation. Developing and using that ability demands your attention.

Given how much attention it takes to express yourself well, you may ask, “Are there no other things to spend attention on?” Why do we have to use natural language to communicate with agents? Why don’t we just do the work ourselves? Why not spend attention on thinking and learning, then on writing code? The fact is that, whatever you think, the world is shifting from programmer-led repositories to agent-led repositories under human supervision. Fundamentally, agents are far more efficient and productive than humans. They don’t get tired as humans do. They finish work that takes countless hours in minutes, and they make ten disposable prototypes that you can use to refine your ideas. This is also practice in expressing ideas, but in the other direction: they express something, and you interpret it. Then you give them instructions for further refinement. I have often used this process without knowing exactly what I wanted at the beginning, only discovering it while working with agents. This lets you iterate quickly and approach levels of efficiency and speed that would have seemed mythical before the GPT-3 era. However, we all know the jokes about LLM hallucinations and how a few careless mistakes can throw everything into irreversible chaos. As agents evolve, we begin to observe that the better you are at expressing yourself, the more reliably and efficiently they work. Agents’ ability to avoid hallucinations has improved dramatically over the past eight months, from February to October. Improvements in instruction following make improvements in your ability to express yourself even more visible in your work. Having someone with no technical background and poor skills of expression work on the same task can produce an even greater difference in results than differences in model capabilities. It is a matter of narrowing the scope into small chunks and putting the pieces together, rather than working with an ambiguous scope and unclear divisions of responsibility. Putting those pieces together is something a person who expresses ideas well is capable of doing. As I said, understanding something does not imply that you can express it properly, but the reverse does hold.

To sum up, the first share of your attention should go into developing what you need to express yourself well and applying those skills when you collaborate with agents. Investing your attention wisely in fundamental building blocks remains irreplaceably important, but that importance now extends beyond the code you write by hand.

Consistent engineering principles

To be continued.