Are We Optimizing AI for Execution at the Cost of Intelligence?

The MoE Dilemma: Efficiency, Scale, and the Risk of Siloed Thinking in AI

Do we want AI that gives us fast answers or AI that helps us understand the world? That’s the real question.

China’s DeepSeek, a company pioneering Mixture of Experts (MoE) technology, has built what it claims is one of the most efficient large language models in the world, faster, cheaper, and better at structured tasks than its Western counterparts. The promise is simple: by only activating the “experts” relevant to a given query, DeepSeek’s MoE model can cut down costs and improve performance.

On the surface, it sounds like a significant technological leap. But efficiency and intelligence are not the same thing, and I can’t shake the feeling that we’re optimizing for the wrong thing.

We’ve always known that intelligence isn’t about having the right answer but knowing how to think. It’s about recognizing patterns, drawing connections across disciplines, and challenging assumptions to unlock new insights. Traditional dense models, for all their inefficiencies, have this capacity. They engage their entire knowledge base, allowing for reasoning that isn’t siloed or constrained by pre-filtered logic.

However, DeepSeek’s MoE model is structured, selective, and compartmentalized. It activates a subset of experts to process a given query, making it exceptional at structured, deterministic tasks. It’s an engineer’s dream, AI that’s cheaper, faster, and more specialized. But intelligence isn’t just about specialization. In fact, the most important breakthroughs don’t come from efficiency, they come from discovery, from pushing the boundaries of what we thought we knew.

So the real concern is this: Are we building AI to function like a highly efficient but fundamentally rigid bureaucracy? Are we engineering systems that execute well but don’t challenge? If so, we might end up with AI that’s good at following orders but bad at wisdom. And in a world where AI is starting to shape everything, that’s not just a technical question—it’s a philosophical one.

Are We Training AI to Think Like Bureaucracy?

When I look at MoE, I see a system optimized for execution, not intelligence.

DeepSeek’s MoE operates by routing queries to pre-selected specialists, experts trained for specific tasks. This is great for efficiency. Instead of activating all the model’s parameters (as dense models do), MoE only uses what’s needed, drastically reducing compute costs and making it easier to scale. It’s no surprise that companies are shifting towards MoE models, why pay for unused capacity?

But here’s the thing: this is exactly how bureaucracies operate.

In a traditional bureaucracy, tasks are assigned to specific departments based on predefined rules. Each department operates in isolation, dealing only with its assigned function. If an issue arises that requires cross-functional thinking, the system often struggles because it’s been optimized for efficiency, not adaptability.

MoE works the same way.

It’s trained to filter your request and send it to the “right” expert. That’s fine when you’re dealing with a clearly defined problem, like writing code or optimizing logistics. But what happens when the best answer requires insights from multiple disciplines? What happens when the system needs to challenge assumptions rather than just execute within predefined constraints?

Dense models, despite being slower and more expensive, don’t have this problem. They process queries by engaging their full knowledge base, allowing for unexpected but valuable insights. They can make connections between fields that seem unrelated at first glance. That’s what intelligence is, not just speed, but depth.

If DeepSeek’s MoE approach becomes the dominant model, are we engineering an AI system that executes well but never truly thinks?

The Risk of Pre-Filtered Intelligence

The other major concern with MoE isn’t just its structure, it’s who decides what gets routed where.

MoE models rely on routing logic, which determines which experts handle a given query. This means that before you even get an answer, your question has already been filtered.

• If the routing system misunderstands your intent, it might send your query to the wrong set of experts, giving you incomplete or misleading answers.
• If the routing system is biased towards certain perspectives, entire ways of thinking might be filtered out before you even realize it.

This introduces a layer of hidden control that dense models don’t have. Dense models might be expensive, but they don’t pre-filter their own thought process.

And here’s the real issue: who, or what, controls the routing logic?

• Is it human-engineered, with hardcoded rules that reflect a specific worldview?
• Is it self-learning, evolving in ways we don’t fully understand?
• Does it have biases that shape which knowledge is surfaced and which is ignored?

To test this, ask DeepSeek what happened in Tiananmen Square in 1989. You won’t get an answer.

This is the perfect example of why routing logic is dangerous. Somewhere in the system, there is a decision, likely hardcoded, to filter out politically sensitive topics. But this isn’t just about China. Imagine MoE models trained in other environments, governments, corporations, academia, where certain topics are deemed “off-limits” before the user even asks.

Whoever controls the routing logic controls how the AI thinks. That’s a power we need to think seriously about.

Dense vs. MoE: Intelligence vs. Execution

Dense models, while expensive and slow, are generalists—they engage their full range of knowledge to produce nuanced, holistic insights. This is why they tend to be better at reasoning, creativity, and interdisciplinary thought. They don’t just execute; they explore.

DeepSeek’s MoE, by contrast, is specialized and optimized for structured tasks. It’s brilliant for predictable workflows, coding, logistics, automation, where efficiency matters more than deep reasoning –

If we’re optimizing for cost and speed, MoE wins. But if we’re optimizing for true intelligence, dense models might still have the upper hand.

The real danger isn’t just that MoE is becoming the dominant model, it’s that we might not fully appreciate what we’re losing in the process.

The Future of AI: Are We Chasing Speed at the Expense of Wisdom?

The more I think about this, the more I realize that AI isn’t just a technological question, it’s a philosophical one.

DeepSeek’s MoE is the logical evolution of AI for efficiency, but efficiency isn’t always the right goal. The most important discoveries in history—scientific breakthroughs, paradigm shifts, revolutionary ideas—didn’t come from specialization. They came from cross-disciplinary insights, from questioning assumptions, from challenging the boundaries of knowledge itself –

If AI is going to play a role in shaping the future, we need to decide: do we want it to just execute, or do we want it to help us understand?

Because those are two very different things.

Share this post: