36.【Physical AI Design】💥 Why Directly Connecting an LLM Always Fails

Latency, Non-Determinism, and Loss of Control

tags:

• Physical AI
• LLM
• Control Engineering
• FSM
• System Design

💥 Why Directly Connecting an LLM Always Fails

Latency, Non-Determinism, and Loss of Control

In the previous article, we clarified that
Physical AI is not a problem of intelligence, but a problem of design.

So why do so many teams still fall into the same idea?

“LLMs have become this smart—
why not just let them handle control and decision-making too?”

Here is the conclusion, stated plainly:

The moment you do that, the system collapses structurally.

This article explains—
not by intuition, but by structure—
why directly inserting an LLM into a control loop is guaranteed to fail.

🧠 The Conclusion First

Direct LLM integration fails for three fundamental reasons:

• ⏳ Latency
• 🎲 Non-determinism
• 🔀 State destruction

This is not a matter of tuning or refinement.
It is a combination that cannot work by design.

💣 The Triple Failure Set

⏳ Latency: It Misses the Deadline

In Physical AI, control loops operate in real time.

Examples include:

• Motor control
• Attitude stabilization
• Voice interruption handling
• Safety stop decisions

These require responses on the order of milliseconds to tens of milliseconds.

Now compare that with an LLM:

• Response time fluctuates
• Depends on network, load, and generation length
• You cannot guarantee when the next token arrives

Any component that may or may not respond in time
is disqualified the moment it enters a control loop.

🎲 Non-Determinism: It Cannot Be Reproduced

In control engineering, the most dangerous property is this:

Under identical conditions, the system behaves differently.

LLMs are inherently:

• Sample-based
• Probabilistic
• Context-dependent

Which means:

Same input
Same state
→ Different output

This is normal for an LLM.

In control systems, however, this means:

• Impossible to test
• Impossible to verify
• Impossible to certify

“Sometimes it works” is not success.
It is failure by design.

🔀 State Destruction: It Breaks the FSM

Most Physical AI systems—explicitly or implicitly—rely on
finite state machines (FSMs).

Typical states include:

• Initializing
• Idle
• Active
• Error
• Safe stop

When an LLM is directly connected, it tends to:

• Skip states based on semantic judgment
• Ignore “is this allowed now?” checks
• Logically override safety states

This is not a “bad decision”.

It is the inevitable result of
placing a component that does not understand the FSM
inside the FSM itself.

🔧 What This Looks Like from a Control Perspective

From a control engineering viewpoint, an LLM is not a control element.

• No defined transfer function
• No fixed response time
• No stability analysis possible

And yet, if you insert it into a control loop, the outcome is clear:

Instability is guaranteed.

This is not a tuning problem.
It is a structural design error.

🔗 Connection to Earlier Failures

We have already seen this failure pattern in other domains.

• 🎮 Game design structured by FSMs
  → Ignoring state boundaries causes instant collapse
• 🔊 Voice AI interruption failures
  → Injecting decisions during Speaking breaks the system

Different domains, same root cause:

👉 An uncontrollable element was placed inside the control core

Only the surface changes.
The failure structure is identical.

❌ Common Counterarguments — and Why They Fail

“But humans operate this way”

→ Humans are parallel, redundant, and adaptive controllers.
LLMs are not.

“We can just add stricter rules”

→ That means you are building an FSM.
You should place it explicitly—and let it stay in control.

🧭 Summary

LLMs are intelligent.
But they are not controllable.

Physical AI design is about
where intelligence is placed, not how powerful it is.

In the next article, we will show:

• Where an LLM can be placed safely
• How to separate PID, FSM, and LLM roles
• Why a three-layer architecture is the only stable solution

At that point, the boundary between
“usable intelligence” and “forbidden intelligence”
will finally become explicit.