On the creation of narrow AI: hierarchy and nonlocality of neural network skills

The Big Picture

Imagine hiring a tax accountant who, to be great at taxes, also had to learn surgery, basketball strategy, and Mandarin poetry. That’s essentially how today’s most powerful AI works: to get a model that excels at one complex skill, you must train a model that absorbs nearly everything.

This isn’t just wasteful. It’s a safety concern. A coding assistant that secretly understands how to synthesize dangerous chemicals has capabilities nobody asked for. Researchers at MIT and IAIFI set out to understand why narrow AI is so hard to build, and they found two deep structural reasons.

The team of Eric Michaud, Asher Parker-Sartori, and Max Tegmark identified two obstacles. First, skills build on one another in layers: mastering complex abilities requires first learning simpler foundations, which forces training to be broad. Second, once a skill is learned, it isn’t stored in a neat, removable compartment. It gets woven throughout the entire network, making surgical removal messy.

Key Insight: Neural networks sometimes need to learn broadly before they can learn narrowly, and even when they do learn a skill, that skill tends to be distributed across the whole network rather than isolated in a single removable section.

How It Works

To study these problems rigorously, the team designed a synthetic task called compositional multitask sparse parity (CMSP): a deliberately simple, controlled problem set arranged in a hierarchy. Some puzzles are “primitive” and solvable independently; others are “composite” and require mastering the primitives first. Think long division: you can’t do it without multiplication.

Training only on the harder composite tasks led to failure, or agonizingly slow learning. But training on the full distribution, including primitive tasks, produced dramatically better results. The broad training acted as an automatic curriculum, a structured sequence of simpler-to-harder problems that let the network build foundational skills first and use them as scaffolding for compositional ones.

This isn’t a quirk of their toy setup. It points to a real reason why giant general models outperform narrow specialists: the broad sweep of internet text acts as a natural curriculum that teaches foundational skills no narrow training set can replicate.

The second challenge is subtler. Once you have a powerful general model, can you carve out just the coding ability and discard the rest? The researchers tested this using pruning: identifying neurons whose removal has the smallest impact on a target skill, then cutting them away.

The bad news: skills don’t live in isolated neighborhoods. They’re distributed across the network with overlapping representations, so clean surgical removal is nearly impossible. Prune based on coding performance, and you’ll inadvertently damage the abstract reasoning that was quietly supporting it.

The team then tried a technique to actively reshape how skills are stored: group lasso regularization applied during additional training. This mathematical penalty pushes the network to consolidate each skill into as few neurons as possible, forcing spread-out representations to reorganize into something more modular. After regularization, pruning worked far better. Target skills were retained while unwanted ones were successfully removed.

The team ran a head-to-head comparison across their experiments:

• Training from scratch on narrow data: often fails due to the curriculum problem
• Knowledge distillation: a smaller student model learns to mimic the output distribution of a larger teacher
• Pruning (with and without group lasso): directly removing neurons based on ablation scores, which measure how much each neuron’s removal hurts performance on a given skill

Pruning after group lasso regularization consistently beat distillation at producing compact, specialized models. That’s counterintuitive. Distillation is the dominant approach for model compression, but it doesn’t explicitly disentangle skills, which makes selective preservation difficult.

Why It Matters

The safety picture is sobering but not hopeless. You probably can’t skip broad training if you want state-of-the-art performance on complex tasks, because hierarchical skill dependencies mean the general training is doing genuine work. But once you have a powerful general model, there may be a principled path to extracting a lean, narrow version: fewer dangerous capabilities, a smaller computational footprint, better interpretability.

These results also sketch a new design pattern for efficiency. Rather than fine-tuning one monolithic model for every application, future workflows might involve a deliberate “narrowing” step: take a foundation model, apply group lasso regularization, prune aggressively, and deploy a model with only the skills it actually needs. That could reduce inference costs and bring capable AI to resource-constrained environments like edge devices, embedded systems, and scientific instruments where a GPT-scale model is out of the question.

Open questions remain. Can these techniques scale to models with hundreds of billions of parameters? Can clever synthetic curricula circumvent the hierarchy problem? And how do we precisely specify which skills to keep versus discard in a real-world LLM?

Bottom Line: Building strong narrow AI requires first understanding why it’s hard. The answer involves both how data is structured (hierarchically) and how neural networks store what they learn (nonlocally). Pruning with group-sparsity regularization offers one way forward, outperforming distillation in controlled experiments.