Why I'm Still Excited About "Old School" ML in the Age of LLMs

Look, I get it. Everyone's obsessed with LLMs right now, and I'll be the first to admit that I spend way too much time playing with ChatGPT, Claude and DeepSeek. They're incredible tools that have fundamentally changed how I think about human-computer interaction. But here's something that's been bugging me lately: in our rush to embrace these shiny new AI interfaces, we risk overlooking the incredible value that "traditional" machine learning still brings to the table.

Let me share a recent experience. Last month, I was helping a construction tech company build out their site safety monitoring system. While we were using an LLM to generate daily safety reports and communicate findings to project managers (which was pretty cool), the real heavy lifting was still being done by good old-fashioned ML models. These models were processing feeds from dozens of site cameras in real-time, detecting whether workers were wearing proper PPE, identifying unsafe equipment operation, and spotting potential hazards that even experienced safety managers might miss. No amount of prompt engineering could replace that kind of specialized visual pattern recognition.

And you know what? This isn't an isolated case. I keep running into situations where traditional ML approaches – especially CNNs and RNNs – are quietly doing the crucial work behind the scenes. Those Instagram filters you love? CNNs. Your Spotify recommendations? Traditional ML. That predictive maintenance system keeping the factory running? RNNs all the way.

Here's what I think a lot of people are missing: it's not about choosing sides. The real magic happens when you combine these technologies. I recently worked with a retail client who perfectly illustrated this. Their system uses:

• CNNs to process product images and categorize them automatically
• RNNs to predict inventory needs based on historical patterns
• Traditional ML for their recommendation engine
• An LLM to handle customer queries and generate product descriptions

Each piece of this puzzle is crucial, and each technology is doing what it does best. The LLM isn't trying to predict inventory levels, and the CNN isn't trying to write product descriptions. It's a beautiful symphony when you get it right.

I'll be honest – I'm tired of the "X will replace Y" narrative in tech. It's not about replacement; it's about evolution and integration. Sure, LLMs are revolutionary, and I'm as excited as anyone about where they're heading. But I'm equally excited about advances in traditional ML. Have you seen what modern CNNs can do with medical imaging? It's mind-blowing stuff.

The teams I admire most are the ones building expertise across the full spectrum. They're not jumping on the LLM bandwagon and forgetting everything else – they're thoughtfully considering which tool fits each problem. Sometimes it's an LLM, sometimes it's a CNN, and often it's both working together.

So here's my take: by all means, get hyped about LLMs. They're transformative. But don't forget about the incredible value of traditional ML approaches. The future isn't about one technology winning out over the other – it's about building systems that leverage the best of both worlds.

The next time someone tells you traditional ML is dead, ask them how their site safety monitoring works. Or their structural analysis system. Or their equipment optimization algorithm. The answers might surprise them.