A reflection on why we need both builders and understanders in the age of AI
There’s a moment that happens in every AI conversation that we rarely think about. You type a question into ChatGPT, Claude, or any other language model, and within seconds, you get a response that feels remarkably human. Coherent. Thoughtful. Sometimes even wise.
But in that brief moment between your question and the AI’s answer, an invisible symphony of complexity plays out. Billions of calculations. Weights and biases dancing together. Attention mechanisms focusing and unfocusing like a mind learning to see.
And behind all of that complexity? Humans. Quiet heroes who spend their days not just building AI, but understanding it.
The World Sees the Magic, Not the Mechanics
We live in an age of AI democratization. Everyone’s building agents. Crafting prompts. Deploying chatbots. The focus is rightfully on application – how can we use these tools to solve real problems, boost productivity, create value?
But there’s another conversation happening in the background. A deeper one.
While the world marvels at GPT’s eloquence and celebrates another “fit-for-purpose” AI agent launch, there are those who ask different questions:
- How does a machine learn to understand context from nothing but text?
- Why does increasing model size lead to emergent capabilities?
- What happens when alignment goes wrong, and how do we prevent it?
These aren’t just academic curiosities. They’re the foundational questions that determine whether our AI future is built on solid ground or shifting sand.
The Gardener’s Wisdom: Making AI More Purposeful
Recently, I spent time creating educational content about Large Language Models – breaking down the mechanics that most people never see. What struck me wasn’t just the technical elegance, but the profound human wisdom embedded in every optimization technique.
Take model compression, for instance. Engineers have developed three main approaches:
Quantization – like compressing a photo, reducing precision while preserving the essential image. It’s not about making something smaller; it’s about finding the minimum viable fidelity.
Pruning – removing the parts of a neural network that don’t contribute meaningfully to the final output. Like a gardener cutting away dead branches to help the tree focus its energy.
Distillation – teaching a smaller “student” model to mimic the behavior of a larger “teacher” model. The student learns not just the answers, but the essence of how to think about problems.
A master gardener once told me: “First remove what’s dead, then shape what remains, and finally graft the essence to a more efficient form.”
The AI engineers optimizing inference pipelines are doing exactly this – not just making models smaller, but making them more purposeful.
The Consultant’s Intuition: Experience and Exploration
But efficiency isn’t just about compression. It’s about intelligence in how we use our resources.
Consider two more techniques that fascinate me:
Caching – remembering previous computations so you don’t have to repeat expensive work. Like a seasoned consultant who says, “I’ve seen this exact problem at three other companies” and pulls out a proven framework.
Speculative Decoding – generating multiple possible next words in parallel, then selecting the best path forward. Like a consultant who walks into a meeting, immediately senses three possible client concerns, addresses all of them simultaneously, then doubles down on whichever resonates.
The pattern is beautiful: Experience provides the shortcuts, intuition guides the parallel exploration.
These aren’t just technical optimizations. They’re reflections of how human intelligence actually works – building on past knowledge while exploring multiple possibilities.
The Cost of Every Word
Here’s something most people don’t realize: every single word an AI generates costs money. Real money.
When you ask GPT-4 to write you a poem, that’s not just a creative exercise – it’s dozens of matrix multiplications across billions of parameters, running on expensive hardware, consuming actual electricity.
The economics are brutal:
- Cost: Every token generated burns compute cycles
- Latency: Users expect instant responses, but bigger models are slower
- Deployment: Models need to work everywhere – from massive server farms to mobile phones
As one researcher put it: “There are no perfect solutions, only trade-offs.”
This is why the quiet heroes matter so much. They’re not just making AI work – they’re making AI workable. Accessible. Sustainable. Democratic.
The Alignment Challenge: Teaching Machines Human Values
But perhaps the most profound work happening in AI isn’t technical at all – it’s philosophical.
How do you teach a machine to be helpful without being harmful? How do you embed human values into mathematical functions? How do you ensure that intelligence serves humanity rather than replacing it?
The alignment researchers are grappling with questions that feel almost theological:
Reinforcement Learning from Human Feedback (RLHF) – literally teaching AI what humans prefer through reward signals. Like raising a child, but the child can read a million books per second.
Constitutional AI – giving models a set of principles to follow, then training them to critique and improve their own outputs based on those rules.
Safety Fine-tuning – explicitly teaching models to say “no” when they should, to be cautious with sensitive topics, to err on the side of helpfulness rather than harm.
The researchers doing this work aren’t just preventing AI disasters – they’re actively shaping what kind of intelligence we want to create and live alongside.
Why Understanding Matters as Much as Building
After spending time breaking down the mechanics of LLMs into something accessible, one thought kept surfacing: “The future of AI isn’t just for researchers. It’s for everyone.”
But here’s what struck me deeper: While we race to deploy and scale, someone needs to understand. Someone needs to ask not just “Does it work?” but “How does it work? Why does it work? What could go wrong?”
The world needs both kinds of people:
The Builders who craft beautiful prompts, design elegant user experiences, and deploy AI solutions that solve real problems.
The Understanders who dive deep into attention mechanisms, debug alignment failures, and ensure that our AI systems remain interpretable and controllable.
Neither is more important than the other. But we tend to celebrate the builders and forget the understanders.
The Invisible Infrastructure of Intelligence
Every time you have a seamless conversation with an AI, remember: there’s an entire ecosystem of human intelligence that made that moment possible.
The researchers who figured out how to make transformers attend to the right parts of your prompt.
The engineers who optimized inference so your response comes back in seconds instead of minutes.
The alignment specialists who ensured the AI tries to be helpful rather than just technically correct.
The systems designers who figured out how to serve millions of queries simultaneously without breaking the bank.
They’re the quiet heroes. The understanders. The ones who ensure that as AI becomes more powerful, it also becomes more human-aligned, more accessible, more trustworthy.
A Personal Reflection
Creating educational content about AI has taught me something profound: the most important conversations about technology aren’t happening in boardrooms or product launches. They’re happening in research labs, in debugging sessions, in late-night discussions about the nature of intelligence itself.
These conversations matter because they shape what AI becomes. Not just how we use it, but how it uses us. Not just what it can do, but what it should do.
The Call to Action
Whether you’re building with AI or seeking to understand it, you’re part of something historic. We’re not just creating tools – we’re creating new forms of intelligence and figuring out how to live alongside them.
If you’re naturally drawn to building, keep building. Create agents that solve problems. Design interfaces that feel magical. Show the world what’s possible.
But if you’re drawn to understanding – if you’re the kind of person who reads technical papers for fun, who wants to know how attention mechanisms actually work, who loses sleep over alignment problems – know that your curiosity is just as valuable.
The future needs both the prompt engineers and the transformer architects. Both the product managers and the safety researchers. Both the users and the understanders.
The Question That Keeps Me Up
Here’s what fascinates me most: We’re building intelligence that we don’t fully understand, using methods that sometimes surprise even their creators, for a future we can’t entirely predict.
And somehow, this distributed network of human understanding – researchers, engineers, philosophers, and curious minds – is collectively figuring it out. One paper, one experiment, one insight at a time.
The AI revolution isn’t just about the machines getting smarter. It’s about humans getting better at understanding intelligence itself – artificial and otherwise.
What fascinates you more – building with AI, or understanding how AI builds its responses?
The world needs both kinds of curiosity. Both kinds of heroes. Both kinds of intelligence.
The future of AI isn’t just for researchers. It’s for everyone – like you. Whether you build, understand, or simply ask better questions, you’re shaping what comes next.
Credits and Acknowledgements: This piece was created through human-AI collaboration that embodies its central message.
[1] Tong Xiao and Jingbo Zhu, authors of Foundations of Large Language Models (NiuTrans Research & NLP Lab, Northeastern University), for their
structured, accessible insights. GitHub link
[2] ChatGPT Study Mode by OpenAI, for its step-by-step, interactive learning approach — helping concepts stick, not just pass by.
[3] Google NotebookLM, for turning dense texts into digestible overviews, mind maps, and even audio/video summaries that made exploration
more visual and fluid.
[4] Claude by Anthropic, for early brainstorming support and thoughtful, reflective answers that offered helpful contrast and context + Helping refine and crystalize my thoughts with words.
[5] The open web including but not limited to YouTube videos, democratized models and tools, blogs and forums.
Leave a comment