Predictive Software Neural Networks
Discover how predictive software is changing engineering by moving from manual rules to pattern matching at scale.
I used to think AI was just very smart if/else logic. Then I started digging into how a refund decision gets automated on a food delivery app, and realized that the entire game changes when you move from writing rules to discovering them.
That shift, from logic to prediction, is what this whole series is about.
From Rules to Predictions
In traditional software, I write the rule. If X happens, do Y. I know the condition before I write a line of code.
Prediction flips this. I don't know the rule. Instead, I have examples: thousands of past refund requests, each with an outcome already decided by a human. My job is to find a set of rules that explains those outcomes and then apply those rules to future requests where I don't have the answer yet.
This moves the problem from writing logic to discovering it. And that discovery process is what machine learning is.
Sample to Population
The core setup is always the same: a sample and a population.
A sample is the data I have: previous refund requests with known outcomes. A population is everything I care about, including every future request I'll ever need to make a decision on. The symbol $\forall$ means "all of them."
The goal: find rules in the sample that generalize to the population. If the sample is representative, the rules will hold. If not, the whole thing falls apart, and we'll see exactly how that failure looks.
The Scale of Compute
Finding those rules isn't cheap. The search space for possible rule combinations explodes fast. This is why NVIDIA is valued in the trillions: raw compute power is the bottleneck for everything in this space.
But hardware alone doesn't get you there. The algorithmic breakthroughs matter just as much. Two Nobel Prizes this year went to researchers in this field. One of the landmark algorithms, backpropagation, is central to how neural networks learn. We'll build up to it from first principles by the end of this series.
ExpandRule-based engineering vs machine learning: two different approaches to producing an output from input
What Gets Built
Software engineers sit at the intersection of all of this. The data scientist figures out what rules work. The ML engineer gets those rules running in production at scale. And we are the ones wiring it into a product that a user actually touches.
In the next post, we'll build the first piece of that pipeline: a concrete model that decides whether a refund request is legitimate, using real features and a real sample. The full production workflow -- from prototype to deployed API -- is where this series ends up.
The Essentials
- Predictive Engineering: A shift from writing explicit logic to finding rules that match patterns in known data.
- Sample vs. Population: We use a known subset (sample) to develop rules that generalize to the entire set ($\forall$), including future, unseen data.
- The Role of Compute: Modern ML requires immense processing power to search the space of possible rules efficiently.
- Models as Converters: A model is a set of rules that converts background information into a prediction.
Further Reading and Watching
Keep reading