For the past five years, the AI industry has pursued a remarkably simple strategy: build a bigger model. More parameters. More data. More GPUs.
The approach has worked spectacularly well. Take ChatGPT, GPT-2 became GPT-3. GPT-3 became GPT-4. Reasoning models have grown dramatically more capable, and there is every reason to believe frontier models will continue improving.
Yet the next leap in AI may come from a very different direction. Not from making one model smarter, but rather from learning how to organise thousands of them.
That possibility brings to mind Douglas Hofstadter’s Gödel, Escher, Bach. His central insight was that intelligence emerges not from isolated components but from systems whose interactions create capabilities that none possess individually. His examples came from mathematics, music and human consciousness.
For decades, many readers interpreted this as a vision of a single intelligent machine that could perform the recursiveness within itself. Perhaps we misunderstood the lesson as a quiet shift is already underway.
The leading AI labs are no longer building only larger models. They are building coding agents, research agents, browser agents, memory systems, planning systems and above all, and orchestration layer. The question is gradually shifting from “How intelligent is the model?” to “How effectively do the components cooperate to create an intelligent system?”
That is a fundamentally different engineering challenge.
Consider a software engineering team. One AI agent plans the work. Another writes the code. A third generates tests. A fourth reviews security. A human engineer, possibly an agent in the future, makes the final architectural decisions.
None of these agents are dramatically smarter than today’s frontier models. Yet together they accomplish far more than any one of them could alone. The intelligence resides in the organi sation.
Human organi sations have always worked this way. Companies outperform individuals not because they employ superhuman people, but because they divide work into specialised roles, operate in parallel, verify one another’s work, accumulate institutional memory and improve through continuous feedback. AI is beginning to follow the same pattern.
This changes how intelligence scales. Frontier models will continue to improve, but every new generation requires dramatically more compute, capital and energy. Scaling remains essential, but it is becoming increasingly expensive.
Organisation scales differently. Specialisation, verification, memory and coordination can improve largely through software. Even if future models become vastly more capable, there will still be compelling economic reasons to organise thousands of specialised agents rather than ask one superintelligence to perform every task sequentially. Organisations exist because coordinated intelligence is often more productive than individual intelligence.
Critically, that shifts where value may be created.
Many of AI’s next breakthroughs are likely to come not from raw intelligence but from better coordination. Which agent should perform which task? Which model should verify the answer? What should be remembered? And how does a system operate effectively with thousands of rapidly learning components. In a 100,000-person corporation there are dozens of management layers. For 100,000 agents the orchestration is no less challenging.
If that is right, today’s AI race may be missing part of the picture. Frontier models will remain indispensable. They are the foundation of the entire ecosystem. But foundations rarely capture all the value.
Processors transformed computing, yet operating systems became equally indispensable. Cloud infrastructure transformed enterprise software, yet orchestration platforms became some of the industry’s most valuable businesses.
AI may follow the same pattern. The defining companies of the next decade may not simply build the smartest models. They may build the systems that organise intelligence itself. For the past decade, AI researchers have pursued one question: How do we build a smarter model? The next decade may revolve around a different one: How do we build a smarter organisation?
The first chapter of AI was about building artificial intelligence. The second may be about building artificial organisations.
Judah Taub is founder and managing partner of Hetz Ventures, former Israeli intelligence officer, and adviser to governments on AI, cybersecurity and defence strategy.
