How Brain Rotted Are We?
Yesterday, I had a conversation with a friend, and it made me rethink how badly AI is affecting our brains. Are we facing a brain rot epidemic?
So when I was 9 years old, I always wondered why I couldn’t use a calculator to do math when it was quite obvious that everyone uses calculators in the real world. Now I realize that not giving me a calculated answer gave my brain a chance to grow, to think, and to improve my intellect.
We are way past calculators now and have agents doing everything for us which we used our brains for. The role of a software engineer has evolved to become a prompt orchestrator. The question is, how can we orchestrate a workflow which we are not experienced in?
I am lucky to have completed my foundations in Computer Science prior to LLMs becoming popular. Prior to when LLMs were a thing, I used to spend countless hours on competitive programming and making personal projects. This gives me the upper hand in having a pre-established understanding of the foundations of software engineering.
The same cannot be said to those who are graduating now. I believe that most graduates do not have firsthand experience building a project piece by piece. So what sets them apart from agents?
There are 3 parts to this answer from my perspective:
- Intuition
- Understanding of cause and effect
- Pre-established technical expertise
While intuition is innate, understanding cause and effect and gaining a foundation in technical expertise requires active problem-solving. But most of us are delegating problem-solving to AI to maximize productivity. In return, we are building up cognitive debt.
What’s wrong with letting AI problem solve for us?
Let’s take a real-world scenario. Ask AI to generate an essay about the solution to climate change. Then do this 100 times over. The chances are it will generate the same content worded differently multiple times. This is a real world study showing this phenomenon.
TIME, on the MIT Media Lab study
The group that wrote essays using ChatGPT all delivered extremely similar essays that lacked original thought, relying on the same expressions and ideas
LLMs are complex pattern matching systems, not intuitive thinkers. They will produce solutions based on what they were trained on. But the world innovates due to novelty. AI struggles to produce novel ideas. So, as a result, we will end up in a world where we hinder originality. What if there was a complex problem you wanted to solve, and you had a solution which an LLM has never come across?
To understand why cause and effect are important, we need to go back to the 1990 study on pneumonia patients.
The 1990 pneumonia model
In the 1990s, researchers trained a model to predict which pneumonia patients were likely to die so hospitals could decide who to admit. The model concluded that asthma patients had a low risk of dying, so it gave them low priority. However, asthma makes pneumonia more dangerous and any doctor would spot this. The reason asthma patients had a lower death rate was because they were sent to the ICU early. The data the model was trained on recorded the outcome but not the reason. The model found a real pattern but arrived at the wrong conclusion since it did not have an understanding of cause and effect.
Therefore, if people who use AI lack the foundation to understand the problem they are solving, how can we trust the output it arrives at?
What’s the solution?
We need to accept that we cannot eliminate our reliance on LLMs. This will put us at a competitive disadvantage among others. Whether we believe in the future of LLMs or whether we think it’s doomed to fail, we need to ride the wave till the wind settles.
However, we need to find a mechanism to repay the cognitive debt LLMs are creating, and this requires us going back to our roots. How did we learn Computer Science 10 years ago?
Writing a good React app is not where our talent lies, but architecting the internal scaffold to allow us to scale and hold its structure for the next decade without being obsolete is where true talent lies.
Want to make a portfolio site? Go use Claude for that. But spend your free time solving intellectual puzzles. Whether it’s an LLM that’s 50% more token efficient than ChatGPT or a deterministic engine to audit an application or even the infamous LeetCode problems.
What makes LeetCode effective is not watching NeetCode’s videos on YouTube and reimplementing but you taking the time to think how to best solve the problem in the lowest time complexity theoretically possible. Learn what data structures are and use the hardest programming language you can to build the solution to convince your brain to work even harder.
I am back on LeetCode. I hope more engineers join me in this journey. Here’s to the 184th problem I will solve soon.