Link Thinking and the New Cognitive Stack

There’s a quiet revolution happening in how large language models think. It started with chain-of-thought prompting, but it’s evolving into something more profound: models that can deliberate, explore, and verify their own reasoning.

I call this “link thinking”—the ability to chain reasoning steps into coherent cognitive sequences. And it’s becoming the defining capability of the next generation of AI systems.

From Pattern Matching to Deliberation

Early LLMs were sophisticated pattern matchers. Trained on trillions of tokens, they learned statistical regularities that allowed them to generate plausible text. But they didn’t reason in any meaningful sense. They completed patterns.

Chain-of-thought changed this. By prompting models to “think step by step,” researchers discovered latent capabilities. The same model that would confidently answer a math question wrong could solve it correctly when forced to show its work.

This revealed something important: the knowledge was there, but the reasoning process needed to be externalized. The model needed space to think.

The Scaling of Reasoning

Recent work from OpenAI and DeepMind has pushed this further. o1, o3, and similar systems use test-time compute scaling—spending more computation at inference time to explore reasoning paths, verify steps, and refine conclusions.

This is a different paradigm from training-time scaling. Instead of making the model bigger, we make it think longer. The model generates candidate solutions, checks them, backtracks when wrong, and iterates toward better answers.

The results are striking. On complex reasoning benchmarks, these systems outperform much larger models that don’t use explicit reasoning. Quality scales with thinking time.

The Architecture of Deliberation

How does this work under the hood? The details vary, but common patterns are emerging:

Search over reasoning paths. Rather than generating a single answer, the model explores multiple lines of reasoning. A value function estimates the promise of each path. Promising paths are extended; unpromising ones are pruned.

Self-verification. The model checks its own work. Does this step follow from the previous one? Does the final answer satisfy the constraints? This requires the model to step outside its generation process and evaluate.

Meta-reasoning. The model reasons about its reasoning. Should I explore this branch further? Is my current approach working? This kind of reflective thinking is what distinguishes deliberation from mere generation.

Implications for System Design

This shift has profound implications for how we build AI systems:

Inference becomes variable-cost. Traditional LLMs have predictable inference costs—input length plus output length, multiplied by per-token pricing. Reasoning models have variable costs that depend on problem difficulty. A simple question gets a quick answer; a hard one requires extended deliberation.

Latency becomes a design parameter. Do you want the fast answer or the good answer? Users need transparency about this tradeoff. Interfaces should communicate when a model is “thinking” and allow interruption if the cost exceeds value.

The line between training and inference blurs. Test-time compute is a form of learning—exploring, evaluating, updating beliefs. The distinction between static model and dynamic reasoning process becomes less clear.

The Cognitive Stack

I think of this as a new cognitive stack for AI systems:

At the base: perception—understanding the input, parsing the problem, identifying relevant information.

Above that: reasoning—the explicit deliberation process, step-by-step thinking, exploration of alternatives.

At the top: metacognition—reflection on the reasoning process itself, knowing when to continue, when to stop, when to try a different approach.

This stack mirrors human cognition in interesting ways. We too engage in fast pattern matching for familiar situations and slow deliberation for novel problems. The difference is that AI systems can scale the deliberation arbitrarily, given sufficient compute.

Challenges and Open Questions

Much remains uncertain:

• Verification is hard. How do we know the reasoning is correct, not just plausible? Self-verification helps but isn’t foolproof.
• Alignment becomes more complex. If models can reason, they can reason about deception. How do we ensure honest deliberation?
• The compute cost is real. Extended reasoning requires resources. How do we allocate them fairly and efficiently?

The Road Ahead

The trajectory is clear: reasoning capabilities will improve, costs will decrease, and deliberative AI will become the default for complex tasks. The question is not whether this will happen, but how quickly and what we do with it.

For practitioners, the implication is to start designing for deliberative AI now. Build interfaces that can handle variable response times. Establish evaluation frameworks that capture reasoning quality, not just final answers. And most importantly, think about what problems actually benefit from extended reasoning versus quick pattern matching.

The cognitive stack is changing. Understanding this change—how these systems think, when they succeed, when they fail—is essential for anyone building with AI.