In semiconductor manufacturing,
there is one metric that matters more than anything else.
That metric is
Yield.
Yield means
the proportion of chips that function correctly out of all chips produced.
For example,
if 100 chips are produced
and 90 of them work,
the yield is 90%.
This number is often called
the lifeline of the semiconductor business.
Why?
Because even a change of just a few percent in yield
can dramatically change profitability.
For decades,
the semiconductor industry has built an enormous amount of technology
all for a single purpose:
to increase yield.
But there is something interesting here.
The way semiconductor manufacturing improves yield
is remarkably similar
to how AI systems should be designed.
Especially when it comes to one concept:
Boundary.
Semiconductor Factories Operate Through Boundaries
As discussed in Overview of Semiconductor Manufacturing Technology and the Application of AI Technologies, semiconductor manufacturing is an extremely complex process.
It involves processes such as
-
photolithography
-
etching
-
thin-film deposition
-
CMP (chemical mechanical polishing)
-
inspection
These steps continue for hundreds of process stages.
At this point, one thing becomes critically important.
Every process step has an acceptable range.
For example:
Film thickness
±2 nm
Temperature
±0.5 ℃
Alignment error
±1 nm
These acceptable ranges are called
process windows.
In other words,
each step has a boundary that defines
“this range is still acceptable.”
Yield Is Determined by Boundary Control
In semiconductor factories,
if even one process step exceeds the boundary,
the chip can fail.
For example:
The resist film becomes slightly too thin.
The exposure alignment shifts by just 1 nm.
Even such tiny deviations can destroy the circuit.
In other words,
yield is the proportion of chips that remained within the boundaries.
Put differently,
increasing yield means:
-
designing boundaries correctly
-
monitoring those boundaries
-
ensuring the process stays within them
AI Systems Have the Same Structure
The same problem appears in AI systems.
AI systems
-
make predictions
-
generate scores
-
perform inference
However,
AI systems do not understand their own limits.
In other words,
AI systems do not naturally have boundaries.
As a result,
AI systems simply continue.
They go as far as they can.
Probabilities keep updating.
Optimization never stops.
And eventually,
the system breaks.
In semiconductor terms,
this is equivalent to
operating outside the process window.
Boundaries Are the Process Window of AI
What AI systems need is
Boundary.
A boundary defines
how far AI is allowed to make decisions.
For example:
Confidence < 0.7
→ send to human review
Transaction amount > ¥1,000,000
→ require human approval
Unknown input
→ stop automatic decision
These rules define
the process window of AI.
In other words,
they define the safe operating range
within which AI is allowed to act.
The Philosophy of Yield
In semiconductor factories,
there is no such thing as a perfect process.
Every process contains
-
noise
-
variation
-
error
Therefore semiconductor factories are designed
with the assumption that errors will occur.
The key is not to eliminate errors.
The key is
to keep errors inside the boundary.
AI systems are exactly the same.
AI systems
will make mistakes.
The problem is not that mistakes occur.
The real problem is
when those mistakes escape the boundary.
AI Yield
From this perspective,
the quality of an AI system can also be understood in terms of yield.
That is,
the proportion of AI decisions that operate safely.
To increase AI yield,
model accuracy alone is not enough.
What is required is
Boundary design.
The Essence of AI System Design
The semiconductor industry spent
more than 50 years
becoming an industry of
boundary design.
The AI industry
has not yet reached that stage.
Many AI systems today focus only on
capability.
They pursue better models
but fail to design boundaries.
However,
as AI becomes embedded in society,
the most important factor will not be the model.
It will be
Boundary.
An AI system is not merely a model.
An AI system is
a factory of decision processes.
コメント