The AI world has been obsessed with scale. Bigger models, more data, insane compute power—it has felt like an arms race where only the giants can compete. But the tide is shifting. Small Language Models (SLMs) are proving that you don’t always need a trillion parameters to get the job done. And that’s where focus is headed: making AI smarter, not just larger.
Why Smaller AI is the Future
Let’s talk about efficiency. Large Language Models (LLMs) are brilliant, but they’re also energy-hungry beasts. Training them takes massive GPU farms, vast amounts of electricity, and an environmental cost we can’t ignore. SLMs, on the other hand, are leaner, faster, and more sustainable. They can run on lower-powered machines, fit within the compute budgets of smaller enterprises, and still deliver strong results—especially when fine-tuned for specific tasks.
China’s DeepSeek project is a great example. Instead of chasing ever-larger architectures, DeepSeek is focused on efficiency—optimizing models so they perform well without requiring supercomputer-level resources. It’s a shift in mindset: how can we do more with less? Techniques such as sparse parameter activation, using reduced numerical precision ops, compressed models, reinforcement learning, optimized cross-chip communication, and design for standard GPUs etc. all enabling the innovative DeepSeek.
The Perfect Blend: SLMs and Classic ML
For too long, we’ve been caught in an either/or debate. Classic ML vs. Deep Learning. Traditional AI vs. Generative AI. But the reality is, the future isn’t about picking one side—it’s about blending both approaches to find the sweet spot.
Take fraud detection as an example. A Generative AI model might be great at simulating rare fraud scenarios, helping a system learn about edge cases. But when it comes to real-time transaction monitoring, classic ML models like decision trees and gradient boosting are still the best at efficiently analyzing structured data and flagging anomalies. The smartest companies will use both—leaning on generative AI for creativity and training, while relying on traditional methods for speed, reliability, and explainability.
This hybrid approach isn’t just smart—it’s necessary. SLMs won’t replace classic ML; they’ll work alongside it, complementing its strengths.
Making AI More Sustainable
Sustainability isn’t just about energy consumption. It’s about building AI that is accessible, cost-effective, and practical—solutions that companies of all sizes can deploy without breaking the bank. We need models that run efficiently on local hardware, that don’t require an endless stream of GPU clusters to function.
The last two years were about scaling up. 2025 will be about scaling smart.. This year I plan to revisit the basics of classic ML and Data science, and AI engineering using a hybrid approach. Stay tuned, more to come..