A Life Filled With LLMs
I spent 20 billion tokens building projects.
That sentence sounds exaggerated, but it is close to how this period of my life has actually felt. LLMs have entered my workflow deeply: writing code, changing code, reading documentation, looking up problems, explaining errors, generating plans, refactoring projects, writing tests, writing docs. Their shadow is almost everywhere.
The strange thing is that, in theory, I should feel lighter. When such a powerful tool enters the production process, it should reduce my workload, reduce my working hours, and reduce the pain of facing complex systems. Reality has been exactly the opposite.
My workload has not gone down.
My working hours have not gone down either.
Even worse, my brain feels more exhausted than before.
It Did Not Do the Work for Me. It Changed How I Work.
Many people imagine LLMs like this: you hand a task to the model, and it hands the result back to you. It sounds like a new kind of automation. But after actually using it to build projects, I increasingly feel that LLMs have not done my work for me. They have only transformed the shape of my work into something else.
In the past, when I built projects, I faced code, documentation, compilers, and debuggers. The problems were complicated, but the context was continuous. If a bug blocked me, I would follow the call chain, waveforms, logs, and source code all the way down. It was painful, but my brain stayed inside the same problem space.
Now it is different.
I have to constantly describe problems to an LLM, organize context, filter information, judge whether its answers are reliable, and then map its suggestions back into the project. When it writes a piece of code, I cannot trust it directly. When it gives an analysis, I cannot accept it directly either. I have to check whether it misunderstood the architecture, invented an interface, missed an edge case, or turned a local optimization into a global disaster.
So the continuous thinking I used to have gets cut into many fragments.
On one side is the actual context of the project; on the other is the LLM’s understanding of the project. On one side is the real structure in my head; on the other is the structure it temporarily assembles from prompts. I have to keep switching between these two worlds.
This is not relief.
It is a new kind of mental labor.
Context Switching in the Brain
What really tires me out is not writing code. It is context switching.
A complex project already requires holding a large amount of state in your head: relationships between modules, interface contracts, historical design decisions, unresolved issues, attempts you just made, paths already proven not to work, and bugs that look like bugs but are actually expected behavior.
After LLMs enter the process, these states do not disappear. They multiply.
I not only have to remember the project context. I also have to remember “what I just told the LLM,” “what it currently knows,” “what it does not know,” “what it may have misunderstood,” “whether the conclusion from the previous conversation is still usable,” and “whether the code generated in this round conflicts with the previous round.”
It feels like I already had a heavy system running in my head, and now I have launched a huge collaborative process beside it. Sometimes it speeds things up. Sometimes it consumes memory like crazy.
The scariest part is that it does not truly remember for you.
It looks as if it knows everything, but in reality it easily forgets, confuses, and drifts. You have to keep feeding it context, correcting it, and reminding it. In the end, you realize that the person carrying the long-term memory is still you.
The LLM generates. You take responsibility.
Memory Overflow
I have become more and more aware of a kind of “memory overflow.”
Before, when I worked on projects, I was tired too, but the tiredness was purer. Today I debugged this module, tomorrow I debugged that module, and there was still a relatively clear thread in my mind.
Now, LLMs make tasks move faster. In one day, you may read more code, try more approaches, generate more patches, open more branches, and discuss more possibilities. On the surface, this looks like improved efficiency.
But human working memory has not been upgraded along with it.
The faster tasks move, the more unclosed contexts accumulate in the brain. This approach was tried but did not completely fail. That module was changed but not fully verified. A sentence in one answer may be useful but has not been confirmed yet. A bug may have been fixed, or it may only have been hidden.
All of these things hang in the mind like memory that has not been released.
In the end, a person becomes extremely tired. Not physically sleepy, but mentally overflowed: sticky, deep, and impossible to organize.
LLMs make information production too easy, while the human ability to digest information has not improved at the same pace.
Faster Does Not Mean Less
This may be my greatest disillusionment with LLMs: they make many things faster, but they do not make things fewer.
Code is written faster, so requirements increase.
Documents are generated faster, so documents increase too.
Plans arrive faster, so the number of plans that need comparison also increases.
Prototypes are built faster, so people begin to expect you to produce in one day what used to take a week.
Technological progress does not naturally bring rest. More often, it raises the standard. In the past, you could only do one thing in a day. Now you can do three, so three things become the new normal workload.
So LLMs have not liberated people from work. They have only compressed work at a higher density into the same amount of time.
That is also why, after using so many LLMs, I increasingly miss the state before they appeared.
Back then things were slower, but the slowness had order.
Electronics Have Been Pulled In Too
LLMs are not only changing software development. They are also changing the whole environment of technology consumption.
Now every electronic product seems to need AI. Phones need AI. Computers need AI. Headphones need AI. Even keyboards and mice seem eager to have AI. Manufacturers have finally found a new reason to raise prices: stronger NPUs, larger memory, higher compute, smarter systems.
But as a user, do I really need this much AI?
Most of the time, I only want a stable, durable, cheap, quiet tool. But the market will not stop for my plain needs. AI has become a new marketing entrance, and also a new way to pass costs on to users. You may not need those features, but you still have to pay for them.
The more ironic part is that products have not become easier to use because of it. Many features are only wrapped in another layer: more buttons, more pop-ups, more cloud services, more subscription entrances. They look smarter, but the actual experience becomes more complicated.
Technology was supposed to return tools to their nature as tools. Instead, tools increasingly feel like shopping malls that keep trying to sell themselves to the user.
Media Quality Is Declining
Another problem LLMs bring is that the threshold for producing content has dropped sharply.
Articles can be generated automatically. Scripts can be generated automatically. Video copy can be generated automatically. Thumbnails and titles can be optimized automatically. Even comment sections can be filled automatically. There is more and more content, but the amount of content truly worth watching has not increased in proportion.
On short-video platforms, more and more content has a familiar taste: the title is provocative, the structure is complete, the tone is smooth, but after watching it, nothing remains. It feels as if it has been polished by the same machine: smooth, full, and hollow.
This is frightening.
In the past, low-quality content at least exposed a person’s roughness. You could see a person’s limitations, and through those limitations you could also see something real. Now, low-quality content is polished by LLMs until it looks respectable. It no longer appears crude. It may even look professional. But at its core, it still has no thought, no experience, and no real judgment.
When everything becomes more like content, truly valuable content becomes harder to see.
Information pollution does not happen simply because there is too much information. It happens because too much information looks real.
I Wish It Had Never Appeared
If I look at it only rationally, of course I know LLMs are powerful. They can indeed improve efficiency, help with learning, and lower the barrier to many things. Without them, I probably could not have advanced some projects so quickly, or encountered so many tools and methods in such a short time.
But human feelings do not always obey reason.
When I realize that my workload has not decreased, my working hours have not decreased, and yet my brain has become more tired; when I realize that every day I am switching context, organizing context, feeding context to models, and then rescuing context from model output; when I realize that technology products are becoming more expensive and media content increasingly resembles industrial waste, a very strong thought appears in me:
I wish LLMs had never existed.
This is not because I hate technology.
On the contrary, it is because I love technology too much.
The technology I love is the kind that makes people freer, clearer, and closer to the essence of things. Not the kind that pushes everyone toward higher speed, higher density, and louder noise.
Maybe LLMs will eventually become true productivity tools.
But at least right now, what they bring me does not feel like liberation. It feels like an accelerated predicament.
They have made the world faster.
But I am not sure the world has become better because of it.