← WRITING·2025-11-09·5 minAI

Notes on learning AI as a backend dev

I've been a backend developer for four years. I understand distributed systems, databases, queues, APIs. I did not understand AI beyond "it's a neural network and it learns from data."

Here's what actually helped me build a mental model.

Start with the math you already know

If you've done any statistics — even just enough to tune a Prometheus alert threshold — you have more foundation than you think. Linear regression, loss functions, gradient descent: these aren't mysterious. They're optimization problems.

The one thing that genuinely surprised me: backpropagation is just the chain rule from calculus, applied recursively. Once that clicked, a lot of the "magic" became mechanical.

The inference-vs-training split

For robotics specifically, I care most about inference: taking a trained model and running it on constrained hardware in real time. Training is a batch job you run once (or occasionally). Inference is the thing that has to work at 50Hz on an edge device.

This framing helped me prioritize. I don't need to understand distributed training at scale. I need to understand model quantization, ONNX export, and latency profiling.

What I'm still learning

Control theory is adjacent to ML in robotics — it predates it, and in many cases outperforms pure ML for low-level control. I'm working through Åström and Wittenmark's Computer-Controlled Systems alongside the ML material. The combination is more useful than either alone.

The field moves fast. My strategy: understand fundamentals deeply, follow a few practitioners closely, and resist the urge to learn every new technique the week it's published.