Noah Zhou Learned to Post-Train AI on $20 at a Time

When Noah Zhou began experimenting with artificial intelligence, a failed training run could exhaust the ten or twenty dollars of cloud credit he had available. There was no research budget to absorb the mistake and no cluster of GPUs waiting for another attempt. If a run failed, Zhou might have to wait until he could put together enough money to try again. That forced him to decide what an experiment was supposed to prove before starting it, then watch closely enough to understand why it succeeded or failed. "I had to plan every run before I touched a GPU," Zhou said.

Growing up in Surrey, British Columbia, outside Vancouver, Zhou did not have a local machine learning community or his own graphics card. He began with computer vision projects, training small models to perform tasks such as detecting a banana in real time and identifying potholes in dashcam-style footage. Working small was initially a necessity because those were the models he could afford to run. The experiments gave him a practical way to learn how model behavior changed as he adjusted the training process rather than simply reading about those changes.

Large language models eventually shifted his attention toward post-training. Pre-training gives a model broad knowledge by exposing it to enormous amounts of information, while post-training shapes how that underlying capability is expressed in practical use. It is where a model can be trained to reason through a problem, follow instructions more reliably, write code, or behave in a particular way. Zhou became interested in understanding that second stage closely enough to manipulate it himself rather than remaining only a user of the resulting systems.

Limited compute affected how he learned the work. Zhou planned workloads in advance, watched gradients for signs that training was going wrong, and studied reward structures because a poorly designed reward could make an entire run useless. He could not afford to test dozens of configurations and simply choose the best result, so he had to form a stronger hypothesis before running an experiment. "I do not reach for more compute, because I never had it," he said. "I reach for the detail that is wrong."

That discipline became especially important when Zhou encountered DeepSeek's research on reinforcement learning for reasoning. He read the paper and related resources, then attempted to replicate DeepSeek-style reinforcement learning himself at a much smaller scale. His experiments included reward design intended to encourage step-by-step reasoning and repeated testing to understand how different choices changed the model's behavior. The objective was not to reproduce the resources or results of a major AI laboratory, but to make the underlying method work at a scale he could personally afford and examine closely.

Zhou later joined a group focused on distilling frontier open-weight coding models into smaller models capable of running on a laptop. That work required deciding which behaviors mattered enough to preserve as capability moved from a larger model into a more compact system. It also reinforced a principle that had already emerged from Zhou's own experiments: model size alone does not answer whether a system is suitable for a particular job. The more useful question is whether the model can deliver the capability a task requires under the conditions where people will actually use it.

Those early constraints now influence Zhou's professional work at a small company, where he is responsible for post-training and for evaluating which models make sense for particular customer use cases. That means weighing quality against cost, speed, and latency rather than assuming that the largest or most expensive model is automatically the correct choice. A system can perform well on a technical evaluation and still be impractical if it costs too much to operate or responds too slowly for the environment in which customers need it. For Zhou, those considerations are part of the engineering problem rather than secondary infrastructure concerns.

The broader AI market is moving closer to the conditions under which Zhou learned. Capable open-weight models give smaller teams more control over what they run and modify, while training infrastructure is increasingly available as a hosted service instead of something every organization must build internally. That shift can give more teams access to tools that were once concentrated inside large laboratories, but Zhou's experience suggests access alone is not enough. Decisions about data, training environments, reward design, and evaluation become more important when teams can no longer rely on scale to compensate for weak choices.

Zhou's advice reflects the way he learned. He recommends proving ideas on small models before scaling them and treating each training run as something that should answer a defined question. "Start small and stay small for as long as possible," he said. "If it does not work at small scale, do not scale it." He also treats compute as something that should be spent deliberately, even when the immediate budget is less restrictive, because the discipline forces a clear hypothesis and a reason for continuing or stopping.

The lack of abundant compute did more than limit what Zhou could run. It forced him to build judgment about which experiments deserved resources, which signals mattered, and which failures pointed to a deeper problem in the training setup. He bought his first real graphics card only after he was employed, after learning through papers, cloud credits, small models, and repeated attempts to reproduce methods himself. Those habits continue to shape his approach to post-training now: understand the problem first, prove what can be proved at small scale, and spend additional compute only when the evidence supports it.