The AI industry is betting hundreds of billions on a single hypothesis: scale large language models enough and general intelligence will emerge. More parameters, more compute, more data. I think this bet has a blind spot.
The architectural limit
Scaling gives you better pattern matching. It does not give you dialectical reasoning — the ability to hold contradictory hypotheses, argue with yourself, and refine through disagreement. A monolithic system, no matter how large, is still one voice. One voice cannot debate.
Human civilization didn’t achieve collective intelligence by growing one giant brain. It emerged from thousands of limited brains competing, contradicting, and correcting each other across generations. Language, institutions, and scientific method are coordination mechanisms for distributed cognition. The intelligence we attribute to « humanity » doesn’t live in any single human — it lives in the interaction.
The tacit knowledge problem
Current AI learns from observation — text, images, data. But experts know things they have never written down. The senior engineer who « feels » that a system will fail. The lawyer who knows which argument will resonate with a specific judge. The doctor who sees what the test results don’t show.
This tacit knowledge cannot be observed. It can only be captured through correction — when an expert sees a wrong answer and fixes it. That delta between the generic response and the expert correction is where real intelligence lives. Architectures that only observe will never capture it.
The alternative path
Distributed cognitive architectures take a different approach. Instead of one massive model, multiple agents with bounded capacity propose hypotheses, compete for validation, and refine through disagreement. Intelligence emerges from the interaction, not from any single agent.
This mirrors how the brain actually works — not as one unified processor, but as thousands of cortical columns that compete and cooperate. It mirrors how science works — not as one authority declaring truth, but as competing theories refined through falsification.
The bet
The question is not whether distributed architectures are interesting. The question is whether we can afford not to explore them seriously — while the entire industry doubles down on scaling alone.
I might be wrong. But if scaling has architectural limits, we should find out before we spend another hundred billion discovering them the hard way.