Why We’re Returning to a Token-Saving Approach

Looking at trends in the AI industry, one interesting pattern emerges.
Ultimately, the industry is returning to a “token-saving approach”.

The atmosphere was different in the early days.
When Jensen Huang (CEO of NVIDIA) proposed the direction of “using lots of tokens,”
many companies competitively moved toward using more tokens.

But when you actually try it, you quickly realize:

Using more tokens doesn’t necessarily lead to better results.

Problems You Notice When Trying Automation

I’ve had a similar experience while setting up various automation tasks.

If you give AI a lot of autonomy
and have it automatically handle multiple steps, it might seem efficient at first…

But a problem arises.

The AI ends up handling too many tasks that I wasn’t aware of
Ultimately, I have to review the entire result myself
The verification costs end up being higher than expected

In conclusion,

Instead of saving time, it actually adds time spent on double-checking.

My Takeaway: Excessive Automation Is a Double-Edged Sword

Some concepts that have been trending recently—

Harness
Loop engineering
Multi-agent automation

When you actually try using these structures yourself, their efficiency is less clear-cut than you’d expect.

The reason is simple.

A system that isn’t fully controlled by humans ultimately requires human intervention anyway.

So, as things stand now, the most realistic approach is something like this:

A person gives clear instructions
AI returns results immediately
A person provides rapid feedback
Quality is adjusted through context engineering

In other words,

“Rapid interaction” is more efficient than “full automation.”

Technological Advancement vs. Human Limitations

This leads me to a more fundamental thought.

Computers have advanced almost exponentially over the past 100 years.
But what about human input methods?

Still the keyboard
Still the mouse

Input interfaces have remained virtually unchanged.

This is a significant signal.

Technology keeps getting faster, but human processing speed remains the same.

Imbalance in the AI Era

AI has become capable of handling an ever-increasing number of tasks, but
it is still humans who must understand and verify the results.

This is where the problem arises.

AI’s processing power is increasing rapidly
Human comprehension speed is limited
Ultimately, humans are the bottleneck

So these days, I find myself thinking:

“Perhaps it’s not that AI has advanced, but that we’ve reached a stage where humans can no longer keep up?”

What Matters in the End

What will matter going forward is neither more tokens nor
more complex agent structures.

Rather, the following will become more important:

How clearly can we give instructions?
How quickly can we verify results?
How cost-effectively can we iterate?

It all comes back to this:

“Concise input + fast feedback + appropriate context”

This is the most practical way to use AI at this point in time.