From Pre-training to Teams of AI Agents: Why Jensen Huang's Four Scaling Loops Change the Computing Question
Introduction: What grows when artificial intelligence becomes more capable?
The discussion about artificial intelligence [AI] often begins with the size of a model. How many parameters does it contain? How much information was used to train it? How large was the computing system behind it?
Those questions describe only part of the system. Jensen Huang's broader argument is that AI can grow at several stages: while learning its foundations, while refining its behaviour, while working through a particular problem and while distributing work among agents. Understanding the connection between those stages changes how we think about computing demand.
The important distinction is between building a model and using an intelligent system. They are related activities, but they are not the same workload.
1. Pre-training builds a foundation, not a finished worker
Pre-training [Pre-training] develops a model's broad capabilities by exposing it to large amounts of information. The model learns patterns and relationships that can support later tasks.
More data and more computing can improve those foundations. However, a larger training run is not a complete explanation of useful intelligence. A model can possess extensive general knowledge and still need better judgement, task-specific experience or access to information it did not encounter during training.
This is why the question cannot stop at whether the supply of human-written material will eventually become a constraint. The next stages can change both the way a system learns and the work it performs.
2. Post-training changes how the foundation is used
Post-training [Post-training] refines a model after its initial learning. The emphasis shifts towards behaviour, practical tasks and the kinds of responses people need.
Synthetic data [Synthetic Data] can contribute to this process. A system can generate examples, explore variations and produce material that becomes useful for further learning. The value lies in the quality of the resulting experience, rather than in the fact that a machine produced it.
The distinction matters. Generating more material does not automatically establish that the material is correct. The useful loop selects experiences worth retaining and uses them to improve later behaviour. Volume without selection can repeat weaknesses instead of correcting them.
3. Test-time computing makes inference a harder problem
Inference [Inference] is sometimes imagined as a simple final step: the model has already learned, so producing an answer should require little further work.
Test-time scaling [Test-time Scaling] challenges that picture. A difficult task may require the system to break a problem into smaller parts, compare possible approaches, search for information and reconsider an unsuccessful attempt.
The computing requirement then depends on what the system must do before returning a useful answer. A short response can still follow substantial work. Conversely, a long response is not proof of careful reasoning.
For a business, this makes response time [Latency], answer quality and operating cost connected choices. The cheapest immediate response is not always the cheapest successfully completed task.
4. Agentic scaling adds coordination to capability
Agentic scaling [Agentic Scaling] introduces another route: a system can distribute work among several agents [Agents] rather than asking one model to perform everything in sequence.
One agent might inspect records while another compares options. Their findings can support a shared result. This is a different form of growth from making one model larger or allowing it more time.
The comparison with an organisation is useful. More employees can increase capacity, but only when responsibilities, communication and the final decision are clear. More agents likewise create work in coordination, reconciliation and checking. Their number is an input, not a measure of success.
5. The four stages form a learning loop
The most important connection is what happens after a system works. Useful experiences from agents can become material for later training. Refined models can then support better reasoning and more effective agents.
Pre-training, post-training, test-time work and agentic activity therefore need not remain separate boxes. They can feed one another.
This also changes the infrastructure question. Training a model, searching through files and coordinating several workers place different demands on processors, memory, storage and networking. A system designed around one narrow benchmark may be poorly matched to the complete workflow.
Conclusion: Measure the work, not only the model
Huang's four-part framework moves the discussion beyond the idea that AI progress means only a larger model trained once on a larger dataset.
The practical lesson is to identify where additional computing produces a better result. Does the task need better foundations, more useful refinement, longer problem-solving or a coordinated division of work?
An enterprise should judge the answer through completed tasks, quality, time and total cost. That is a more useful starting point than counting parameters or agents alone. The future computing question is about the whole learning-and-working system.
Powered By K7