Once we accept distributed inference + different learning timescales, we’re not really talking about models anymore. We’re designing a distributed cognitive system. First jump: frontier models become sparse infrastructure. They don’t answer, inspect, watch, or execute every thing. Most cognition happens locally through smaller models, specialist agents, memory, classifiers, tools. frontier gets called when something exceeds local competence. Instead of: human → GPT-11 → answer you get: human → persistent local intelligence → specialist network → frontier when necessary Question becomes: how much of civilization’s cognitive workload can happen w/o invoking the frontier at all? Radically different compute and energy curve. Smaller models, locality, caching, specialization, reduced data movement, and selective escalation make useful intelligence much cheaper. Learning changes too. Different layers adapt on different timescales: context and working state in seconds, memory and routing over hours, consolidation overnight, adapters or distillation over longer periods, frontier training slower still. Continual learning becomes an orchestration problem rather than “continually retrain the neural network.” Then it gets nuts. With millions or billions of local agents, there’s no reason they communicate only vertically with frontier models. They can communicate horizontally. Your agent hits an obscure Kubernetes problem. Instead of waking the trillion-parameter god-model, it finds a trusted specialist that has solved that class of problem 40k times. Tasks begin moving through an agent economy according to capability, cost, latency, privacy, reputation, and trust. Now you have a market/network of cognition. We tend to imagine small models as inferior versions of frontier models. A small model trained or distilled for one narrow task may outperform a huge general model at that task for a fraction of the cost. Instead of building one intelligence, you dynamically assemble the intelligence required for the problem. Frontier model’s role changes from knowing everything to architecting cognition. The frontier becomes a metacognitive layer. Then strangeness happens: frontier systems can begin pushing their own workload downward. If the frontier solves the same class of problem millions of times, eventually the system can say: I shouldn’t be doing this anymore. Distill the capability. Deploy a specialist. Move it closer to the edge. Loop becomes: novel problem → frontier solves it → pattern becomes common → specialist is created → capability moves outward → frontier is freed for harder novelty Almost the inverse of how we currently imagine scaling. Successful frontier AI now continuously compresses mature capabilities into cheaper intelligence farther down the stack. It becomes a competency generator. Which brings ISPs back into the picture: Work routes to the cheapest competent layer. With sufficient security, attestation, sandboxing, and privacy, even some idle compute across billions of devices could become part of the substrate. AI infrastructure looks more like a mesh. Wild implication: Every person has a persistent local intelligence that is genuinely theirs. Not “GPT with memory.” Something inside their trust boundary. Knows their context, permissions, projects, preferences, history, and goals, and represents them to other machine systems. Humans stop directly operating much of the digital world. Our agents negotiate with the world’s agents. At that point AI becomes a coordination layer for civilization. Once intelligence is distributed enough, the meaningful unit becomes the network. This rhymes strongly with biological intelligence: specialization, communication, hierarchy, distributed processing. Evolution discovered the same basic architectural principle. Intelligence scales through organization, versus solely by making individual neurons more powerful.