Source: The Intelligence Staircase
Technical AI papers can look intimidating from the outside. Lots of architecture names, benchmark tables, and acronyms behaving like they pay rent. But the useful question is usually simple: what bottleneck is this trying to remove?
The AI debate has a vocabulary problem. We argue about whether AI is “truly intelligent” or “just pattern matching” as if intelligence were a single dial. The Intelligence Staircase makes a compelling case that this framing is the source of most of our confusion. The framework proposes ten distinct tiers of intelligence, from basic behavioural competence hardcoded stimulus-response, like a thermostat up through adaptive learning, generalisation, meta-cognition, and recursive self-improvement. Each tier solves a problem the one below it can’t handle. And crucially, performance on tier N tells you very little about capability on tier N+2. The Clever Hans Effect is the most important warning in the framework: a horse that appears to solve arithmetic by reading subtle cues from its trainer demonstrates tier-1 behavioural competence but zero tier-2 understanding. Current AI systems do something analogous constantly — they produce outputs that look like higher-tier cognition but are generated by mechanisms that don’t require it. What I take from this practically: when evaluating an AI system, ask not just “can it do the task?” but “how is it doing the task, and does that mechanism generalise?” A system solving hard problems by brute-force pattern matching over training data is at a different tier than one building and manipulating internal representations.
My takeaway: architecture matters because it decides what kind of intelligence is affordable, reliable, and usable outside the lab. The benchmark is the beginning of the story, not the ending.