You can outsource your thinking, but you can't outsource your understanding.
Andrej Karpathy shared this thought at AI Ascent 2026, and it struck a chord.
Today, it's never been easier to outsource cognitive tasks. We use AI to research, summarise, and draft. It drives incredible productivity. But can we achieve a deep understanding of a complex topic without doing the thinking ourselves?
I don't believe so. Understanding is inextricably linked to deliberate, rigorous thinking.
After recently revisiting Cal Newport's Deep Work, I realised how easily we fall into the trap of shallow reading—especially with highly technical content. To test this, I set myself a goal: revisit the foundational 2017 paper, "Attention Is All You Need," and drop down into the actual mechanics.
I had skimmed it over the years. I knew the vocabulary. But skimming gives you the illusion of knowledge; deep work gives you the architecture of understanding.
Dropping into the math of the transformer revealed entirely new dimensions:
🔍 Positional Encoding: A surface-level read assumes position is simply implied. The math reveals that self-attention is inherently a set operation—it has no concept of sequence. Without positional encodings, the model is essentially looking at a scrambled bag of words.
🔍 Multi-Head Attention: Why multiple heads instead of one massive one? Splitting the embedding dimension allows the model to process information in parallel across different representation subspaces. A single head would dilute the signals; multiple heads allow the model to track distinct features simultaneously (grammar, entities, context) without interference.
This brings us back to Karpathy. He recently shared his workflow of having an AI agent automatically maintain a personal wiki based on his reading. This is a brilliant, high-leverage way to outsource knowledge management.
In addition to doing this, I also use LLMs as Socratic sparring partners.
Instead of asking a model to "summarise this section," I test my understanding: "I think multi-head attention works like [X]. Where is the flaw in my mental model?"
I've written previously about how frontier models are encroaching on the domain of teaching. But if we remove the friction of teaching or struggling through a concept, we lose the ability to identify gaps in our reasoning.
AI can gather, organise, and synthesise information, but the neural pathways of understanding must be forged by the individual.
Use AI to handle the shallow work of knowledge management, but fiercely protect your cognitive cycles. True comprehension requires deep work.
How are you using AI for your own learning—are you using it as a summariser, or a sparring partner?
