In Terminator 2, the T-800 is one of the most capable autonomous systems ever put on screen.
It navigates complex environments. Adapts to changing conditions. Makes decisions in milliseconds. And it doesn't stop to reason through every action from first principles. It runs fast — and only recalculates when something breaks the pattern.
This is, surprisingly, a good model for how AI systems should work. And it's almost the opposite of what the industry is building right now.
The default assumption is wrong
When LLMs produce bad outputs — hallucinations, confident nonsense, wrong answers — the reflex is to add more reasoning.
Chain of Thought. Verification steps. Multi-agent pipelines. More layers, more structure, more computation.
The assumption: LLMs fail because they don't think enough.
But look at what actually happens. The T-800 doesn't fail because it has too little reasoning capacity. It fails — in the first film — because it has the wrong mission objective. The reasoning is fine. The values are wrong.
And standard LLMs don't fail because they reason poorly. They fail because they have no signal that they're failing at all.
How humans actually make decisions
Kahneman described two systems decades ago.
System 1: fast, automatic, pattern-based. This is intuition — the accumulated patterns of experience that fire without deliberate thought. You recognize a face, navigate a familiar route, sense that something is off in a conversation. No reasoning required.
System 2: slow, deliberate, effortful. This is conscious reasoning. It's expensive and humans avoid it when they can.
The key insight: humans don't switch to System 2 randomly. They switch when System 1 sends a signal — something feels wrong, something doesn't fit, there's a mismatch between expectation and reality.
That signal is the error trigger.
Without it, System 2 never activates. With it, the system knows when to slow down and reconsider.
The T-800 runs the same architecture
Fast path by default. Pattern matching drives most decisions. The mission objective acts as the anchor — the equivalent of S1 in any reasoning architecture.
But when the environment throws something unexpected — a new threat model, a mission conflict, missing information — the system recalculates. Not always. Only when triggered.
The result: fast execution most of the time, reliable correction when it matters.
What LLMs are missing
Here is the diagram. Three systems, same basic challenge.
Human and T-800: fast path by default, error trigger activates only when needed. Output is fast and grounded.
LLM: reasoning layer on top of reasoning layer, always. No mechanism to detect when the pattern is failing. Output is slow and expensive — and just as likely to be wrong, sometimes more so, because errors compound across steps.
The missing piece is not more reasoning. It's the trigger that tells the system when to reason.
What an error trigger looks like in practice
It doesn't have to be complex.
Low confidence on the generated output. Contradiction between parts of the response. Missing data that the question requires. Unexpected structure in the input.
Even simple heuristics work: "I don't know" thresholds, cross-checks between outputs, lightweight verification.
The architecture shift:
Instead of:
pattern → reasoning → reasoning → reasoning → output
Try:
pattern → error trigger → (optional) correction → output
Where the trigger fires only when the pattern is uncertain or inconsistent. Not on every query. Just the ones where something is actually off.
Why adding reasoning layers doesn't fix this
More reasoning assumes the model knows it needs to reason. But the failure mode is precisely that the model doesn't know it's failing. It generates fluently either way.
Add reasoning steps and you get longer outputs, more structured errors, higher costs — and sometimes a more convincing wrong answer, because each step built confidently on a flawed premise.
The T-800 doesn't run a full threat analysis before opening every door. It pattern-matches. The analysis kicks in when the threat model breaks.
LLMs should work the same way.
The mental model shift
Instead of: LLMs need more reasoning to be reliable.
Think: LLMs need a signal that tells them when their pattern matching is failing.
Fast systems win. Reliable systems scale. The goal is both — not more layers on top of an architecture that has no way to know when it's wrong.
The T-800 figured this out in 1991. It's time for the rest of the field to catch up.
Top comments (0)