I feel like System 1 models have changed the way I think about AI app architecture.
They simplify the process a lot.
Previously, you had one of two routine options.
Either you trained a neural network to behave as a specialized classifier, or you used an LLM for classification and devised an elaborate workflow to verify the results and make sure they were accurate.
Classifiers still need work for real accuracy, but the process is a lot simpler now.
The interesting thing is that classifiers are actually pretty good at simple decisions, ranking, and estimating probabilities.
What LLMs aren't all that great at is navigating elaborate decision trees with lots of nested if/then conditions.
Adding classifiers makes it a lot easier for your agentic workflows to conform to your decidedly deterministic flow charts.
How's it work?
Well, instead of asking a reasoning model to navigate the entire logic tree, you let your deterministic software manage the branches and use classifiers to evaluate each condition on a stepwise basis.
Think of a decision rubric as a series of nested if/then statements.
Each classification determines which branch to take next, and each branch gets evaluated independently. So you only need to run the branches that matter.
Suddenly, you've got a system that's considerably more deterministic than asking a probabilistic reasoner to figure out the whole thing on its own.
And because you're evaluating individual decisions, you can measure confidence, compare rankings, identify disagreements, and selectively verify results without constructing an elaborate reasoning workflow.
Then you hand the collected evidence back to the reasoning LLM to make the final judgment. And that's important, because a collection of individually correct classifications doesn't necessarily produce a correct final decision.
That's how you separate classification, procedure, and judgment.
The probabilistic reasoning is still there, but it's no longer responsible for running the entire show.
And you save billions of tokens doing it.
Or, if you wanted to get even more aggressive, and totally remove the LLM from the picture, you could always use one classifier to classify what the last classifier did as success, or not. And then use your quorum of classifiers to classify the classifiers, and you've classified your classification classically.
That's what we call... consensus!
Obviously.