How I approach learning

AI is changing how we write, build, and work. And we're going to keep learning the same way? I don't buy it. We can do more than get the same explanations faster. We can rethink the teaching itself.

I want a way of learning that starts from what I already understand, follows a coherent path, and changes when my reasoning reveals a gap. I want that understanding to change what I can do outside the lesson: notice something I would have missed, question a claim, or make a better decision when nobody has labelled the problem for me.

The approach has two priorities. The teaching should fit my current knowledge and the understanding I want to reach. The effort I spend should help me learn the subject: following an argument, trying an idea, noticing a mistake, and working out why it happened. I want help arranging the work so I can give those things my attention.

The goal is to reason from the principles behind a method. In Zero to One, Peter Thiel and Blake Masters describe “thinking about business from first principles instead of formulas.” I want to apply that idea to learning. If I learn a formula, I want to understand its assumptions, why the steps follow, and what changes when an assumption changes. That gives me something to build on when the next problem looks different.

Start from where I am

A book or course gives a subject a structure. I value that: a sequence, prerequisites, and a standard to work toward. But the same chapter can explain something I already know and then rely on something I have never properly understood. A useful tutor should help locate those differences.

That means beginning with questions. What can I explain? What can I use? Where does my reasoning stop? A few answers give us a starting point. They are evidence to keep revising as we work through the material; guessing correctly or recognizing a familiar question can make that first picture misleading.

Prior knowledge also changes what kind of help is useful. In a study of university statistics instruction, the effects of different activities on immediate conceptual understanding depended on learners' prior knowledge. Learners with less prior knowledge benefited from worked examples compared with generating arguments. That supports adapting the help, while leaving open exactly when and how to change it for a particular person.

I want both the path and the individual explanation to respond. If a step depends on an idea I cannot yet explain, we should return to it. If I already have the relevant foundations, we can work on a harder connection or application. A learning path is useful when it shows why a step belongs there and can change when the evidence changes.

Put the effort where it helps

Learning from several sources can mean repeatedly translating notation, finding my place, and reconstructing what I was doing. I want a continuing study space that helps organize that work, while keeping the books, papers, and other sources available.

Cognitive load theory concerns demands on working memory: our limited capacity to hold and manipulate information while thinking. Some demands come from the material; others come from how it is presented. In experiments on irrelevant information, mismatching pictures initially impaired learning of action words compared with matching pictures; that effect disappeared with task experience. These experiments do not establish that bringing several sources into one AI interface improves learning. That part is a design preference I would need to evaluate.

The boundary depends on what I am trying to learn. Formatting a progress note can be delegated. Comparing evidence may be part of the subject itself. If I am learning to read research, deciding whether a paper supports a claim is valuable work for me to do.

I also want the amount of help to change. An example can make an unfamiliar problem possible to think about. A hint can leave me with a useful next step to find. As I become more capable, I want to take on more of that reasoning myself. The amount of struggle alone tells me little; I care about what I am doing with it.

Connect new ideas to something I can use

Learning researchers use schema for an organized structure of prior knowledge. It helps describe why the same explanation can mean different things to different learners: they bring different relationships and expectations to it.

A familiar framework can help memory while leaving details uncertain. In a brain-imaging study of object–scene memories, people remembered the context better for information that fitted a familiar schema. Over time those memories became coarser: people could retain the kind of setting while losing its specific details. The researchers observed related changes in patterns of brain activity. This was a laboratory memory task, so it supplies no direct test of an AI tutor or a teaching method.

What I take from it is a useful question for learning: how does this new idea fit what I know, and which details might that familiar framework hide? An analogy can give me a way in. I still want to check where it stops working, what assumptions it introduces, and what makes the new case different.

In How People Learn II, the National Academies describe knowing when knowledge applies as an essential part of learning. They discuss transfer: applying what we have learned in a new context. Understanding the underlying principles can help, but using them elsewhere still needs checking. (Chapter 1, p. 14)

If I'm learning statistics, for example, I want to use it to judge whether an apparent improvement in a product is convincing. What was measured? What was it compared with? What does the evidence leave uncertain? That gives the concepts somewhere to go beyond the exercise.

Teach through a conversation that checks

A good teaching conversation gives me opportunities to contribute reasoning. Before an explanation, I might make a prediction and explain why I expect it. Afterward, I can compare the explanation with my attempt, identify the difference, and try again.

This connects to pretesting: attempting a question before studying the answer. Experiments comparing tests before and after study found benefits that depended on the assessment; in their delayed experiment, the pretesting advantage was confined to repeated questions. That makes a brief prediction worth considering, without making it a requirement before every unfamiliar step. Sometimes I need an explanation first.

I also want to explain the reasoning back. Researchers call generating an account of how the material works self-explanation. It can make the connections I am drawing available for inspection. The original circulatory-system study found greater gains with explanation prompts, although that group spent more time and could consult materials during the posttest. It offers qualified support, rather than proof that explaining is always the best use of study time.

For my purposes, explanation creates something to examine. Does each step follow? Did I leave out an assumption? Can I explain a changed example? The AI's job is to respond to that reasoning, identify a possible gap, and help check it against the source. Its agreement alone gives me little reason to trust an answer.

Check what I can do with it

The distinction between assisted performance and learning matters especially with AI. In a randomized high-school mathematics study, unrestricted GPT assistance improved practice performance but reduced scores on the subsequent exam without assistance. A tutor with safeguards largely mitigated that reduction; it did not establish a positive independent exam benefit. The way the help was designed mattered.

I therefore want to return to ideas after putting the explanation away. Retrieval practice means bringing knowledge to mind instead of consulting it. In a study comparing retrieval with concept mapping, retrieval combined with restudy produced better performance a week later under the tested conditions. That supports including independent recall alongside the work of building an explanation.

I would also check what happens in a less tidy situation. Can I work out which ideas matter, identify missing information, and choose a reasonable next step? I want to try something, examine what happens, and revise my reasoning when the result challenges it.

Sometimes I should attempt that without help to see what I understand. In actual work, I also expect to use books, documentation, and AI. I want enough understanding to judge their suggestions, catch mistakes, and explain why I trust the result.

Books, AI, and Markdown each have a place in this approach. Books give me a foundation and a coherent account to return to. AI gives me a way to question that account, ask for another explanation, and work through my own reasoning. Markdown keeps the path, sources, attempts, and progress available when I resume.

I want that record to distinguish what I covered, what I could explain independently, and what I could use in an actual task, including the help I needed. The research informs the ingredients; the combination remains a working approach. I would judge it by whether the learning changes how I understand a situation and what I can do about it.