Attention Is All You Need

This page was translated by AI

Preface

As the Lunar New Year approached, I finally got a fifteen-day break. As usual, when I got home, I picked up my phone and tablet and let my mind drift into the unconscious stream of content platforms. Every day I told myself I would only keep a TikTok “spark” going, yet I would not leave Douyin until an hour or two later. Whenever I had nothing to do, my fingers would instinctively open Douyin and start scrolling.

At first, I thought I was simply surrounded by the short-video algorithm. Then I noticed how hard it had become to settle down and complete a long task: watching a film, consciously finishing my tasks, or turning plans in my head into action. Without realizing it, our brains have been reshaped by short videos.

In neural networks, an architecture called the “Transformer” opened the door to large language models (LLMs). Its paper, Attention Is All You Need, explains the architecture’s core: the attention mechanism.

So even a super-brain as powerful as AI needs a weighting algorithm to identify keywords and allocate its precious attention to them. How much more do we?

Next, I will explore how short-video algorithms steal our attention and what changes they make to our brains.

A reward system precisely targeted by “weights”

In the Transformer architecture, attention weights determine which tokens receive more computing power as a sequence is processed. Short-video algorithms, at their core, are searching for the “high-weight features” in our brains.

Research reports suggest that short videos are so hard to stop watching because they perfectly fit the psychological principle of variable-ratio reinforcement. Every video is a probabilistic gamble: you do not know whether the next one will make you laugh or astonish you. That uncertainty keeps stimulating dopamine release.

Just as AI constantly optimizes its path to minimize loss, short-video algorithms calculate your flow state in milliseconds. When you swipe past something uninteresting, the algorithm quickly adjusts the “weight” of the next video, keeping the feedback loop highly activated. Over time, our brains are trained into machines that accept only intense, immediate rewards and grow accustomed to this extremely short feedback path.

Brain reshaping after pretraining: the prefrontal cortex loses control

If human willpower is the CPU’s central controller, then in neuroscience this role is mainly handled by the prefrontal cortex. It governs our higher cognitive functions: planning, decision-making, and impulse inhibition.

Yet studies published in journals such as NeuroImage have found that in people who are heavily immersed in short videos, brain regions responsible for reward processing, such as the ventral striatum, are unusually active, while functional connectivity in prefrontal regions responsible for self-control declines.

This explains why the plans in your head are hard to put into practice. Once our brains are “reshaped” by short videos, top-down attention driven by willpower—reading, coding, or watching a film—is overwhelmed by the bottom-up stimulation of short videos. Put simply, after long-term algorithmic fine-tuning, our brains gradually lose their ability to handle low-frequency tasks with long feedback cycles.

Fragmented attention and task overload

In Attention Is All You Need, attention is computed in parallel. But in human physiology, deep attention carries an extremely high switching cost.

When we grow accustomed to information arriving in fifteen-second cycles, our brain’s “attention window” shrinks significantly. This is a physical change: studies indicate that excessive reliance on short videos can cause marked cognitive overload when people face tasks requiring sustained attention.

That is why, even when you want to watch a two-hour film, your brain sends a “switch now” signal every few minutes. Under the algorithm’s conditioning, it can no longer tolerate ten minutes of input without a reward signal. Countless fragmented short videos fill up our attention bandwidth, leaving too little processing power for a long task that requires depth and a coherent logical loop.

graph TD
    %% Global styles
    classDef algo fill:#e1f5fe,stroke:#01579b,stroke-width:2px,color:#01579b;
    classDef brain fill:#e8f5e9,stroke:#2e7d32,stroke-width:2px,color:#2e7d32;
    classDef fail fill:#fff5f5,stroke:#c62828,stroke-width:2px,color:#c62828;

    %% Top: two parallel causes
    subgraph TopRow ["Why You Just Can't Stop"]
        direction LR
        subgraph Algo ["The Algorithm Guesses What You Want"]
            direction TB
            A1[Every video you swipe past] --> A2(The algorithm furiously calculates your preferences)
            A2 --> A3(Precisely feeds you content that feels good)
        end

        subgraph Brain ["Your Brain Is Changing"]
            direction TB
            B1[Dopamine makes you increasingly excited] --> B2[Your brain gets used to this fast pace]
            B2 --> B3(Self-control gradually “goes on strike”)
        end
    end

    %% Bottom: the outcome
    subgraph Fail ["The Result"]
        direction LR
        C1[Your attention shatters into pieces] --> C2[Your mind wants to work but cannot sit still]
        C2 --> C3["Movies and coding both feel exhausting"]
    end

    %% Connections
    A3 --> Fail
    B3 --> Fail

    %% Vicious cycle
    C3 -.->|You feel empty and open your phone again| A1

    %% Apply styles
    class A1,A2,A3 algo;
    class B1,B2,B3 brain;
    class C1,C2,C3 fail;

A suggestion

If you are preparing for an exam or need time to study with a clear mind, I recommend uninstalling short-video apps.

In everyday life, consciously reduce both how often and how long you use short videos, and avoid becoming absorbed in highly stimulating content for extended periods. This gives the brain’s reward system a chance to “cool down” and gradually regain sensitivity to ordinary stimuli. For users already showing signs of compulsive use, regular digital detoxes—such as setting aside several days each week without watching short videos at all—can be an intervention that lets overstimulated dopamine circuits gradually cool down.

Postscript

Douyin has a “Spark” feature that was originally meant to record the number of consecutive days you and a friend have interacted on the app. Little do people realize that it has also become one of the most effective ways to bind users to the platform.

“Sparks” attempt to make relationships concrete and quantifiable. But relationships that are quantified are often easier to manipulate. You worry that losing a spark will hurt the relationship; you think you have kept it going for hundreds of days, so there is no reason to end it; renewing it takes only a moment anyway….

But connections that are truly worth our time have never depended on a number to survive.

Douyin does not need “Sparks” to make money. For the company, it merely gives you a reason to open the app. Once you do, you are faced with Pandora’s box, carefully prepared by countless algorithm engineers behind internet companies.

So in the new year, why not tell yourself: Attention Is All You Need!