The development of AI to date has required three major laws to describe how different utilization of computing resources affects model performance. These laws include pretraining scaling, post-training scaling, and test-time scaling. These laws reflect how the AI field is evolving by applying additional computing technologies to increasingly complex AI use cases.
The recent rise of test-phase scaling has enabled new types of large language models (LLMs), such as AI reasoning models, to perform multiple inferences to handle complex problems while describing the steps required to solve the task. Test-phase scaling requires significant computing resources to support AI reasoning, further driving the demand for accelerated computing.
What is pre-training extension?
Pre-training scaling is an original law of AI development. It demonstrates that by increasing the size of the training dataset, the number of model parameters, and the amount of computing resources, developers can expect predictable improvements in model intelligence and accuracy.
Data, model size, and computation are all closely related. Research indicates that when larger models are fed more data, their overall performance improves. To achieve this, developers must scale up computation, which requires powerful accelerated computing resources to run larger training workloads.
This principle of pre-training scaling has enabled groundbreaking capabilities in large models and spurred significant innovation in model architectures, including the rise of transformer models with billions and trillions of parameters, mixtures of experts models, and new distributed training techniques, all of which require massive computational resources. Meanwhile, the relevance of the pre-training scaling principle continues to grow as humanity continues to generate more and more multimodal data—text, images, audio, video, and sensor information—that will be used to train future powerful AI models.
What is Post-Training Scaling?
Pre-training large, basic models isn't suitable for everyone and requires significant investment, skilled experts, and datasets. However, once organizations have pre-trained and released models, they can allow others to use them as a foundation for their own applications, lowering the barrier to AI adoption.
This post-training process drives additional cumulative demand from businesses and the broader developer community for accelerated computation, and improves the specificity of models and their relevance to the use cases their organizations require. Post-training scaling primarily enhances the skills of a model for its intended work. For example, a large language model can be post-trained and scaled to handle tasks like sentiment analysis or translation, or to understand terminology in fields like healthcare or law.
The post-training scaling law posits that the performance of pre-trained models can be further improved in terms of computational efficiency, accuracy, or domain specificity using techniques such as fine-tuning, pruning, quantization, distillation, reinforcement learning, and synthetic data augmentation.
Reinforcement learning (RL) is a machine learning technique that uses reward models to train agents to make decisions tailored to specific use cases. The agent's goal is to make decisions that maximize cumulative rewards over time as it interacts with its environment. For example, a large chatbot language model might be positively reinforced by users' "like" responses. This technique is called reinforcement learning with human feedback (RLHF). Another newer technique, reinforcement learning with AI feedback (RLAIF), uses feedback from AI models to guide the learning process and streamline post-training tasks.
To support post-training expansion, developers can augment or supplement fine-tuning datasets with synthetic data. Supplementing real-world datasets with AI-generated data helps models improve their ability to handle edge cases that were underrepresented or omitted in the original training data.
What is a beta extension?
Extended test phase, also known as long thinking, occurs during the inference process. While traditional AI models quickly generate a one-time answer to a user prompt, models using this technology allocate additional computational work during the inference process, allowing the model to reason through multiple possible responses before arriving at the optimal answer.
This AI reasoning process can take minutes or even hours to generate complex custom code for developers, and difficult queries can require more than 100 times the amount of computing power compared to a single inference of a traditional large language model, because traditional large language models are unlikely to produce the correct answer to complex questions on the first try.
This test-phase computing capability allows AI models to explore different solutions to problems and break down complex requirements into multiple steps, in many cases showing users their work during reasoning. Research has found that when AI models are given open-ended prompts that require multiple reasoning and planning steps, the expanded test-phase approach can produce higher-quality responses.
There are various computational methods used during the testing phase, including "chain-of-thought prompts" that break down complex problems into a series of simpler steps, "majority sampling" that generates multiple responses to the same prompt and selects the most frequently occurring answer as the final output, and "search" processes that explore and evaluate multiple paths within a response tree. Furthermore, post-training extension methods similar to best-solution search sampling can be used to optimize long-term thinking during inference to optimize responses that align with human preferences or other goals.
How to perform AI reasoning during the testing phase
The rise of computational technologies in the testing phase is enabling AI to provide informed, helpful, and increasingly accurate responses to complex, open-ended user queries. These capabilities are crucial for the detailed, multi-faceted reasoning tasks expected of autonomous agent-based AI and physical AI applications. They can provide users across various industries with powerful assistants to accelerate their work, thereby improving efficiency and productivity.
In healthcare, a model can use test-phase expansion techniques to analyze large amounts of data, infer the progression of a disease, and predict potential complications of new treatments based on the chemical structure of the drug molecule. Alternatively, it can comb through a database of clinical trial data to recommend treatment options that suit an individual's condition and share its reasoning about the pros and cons of different studies.
In retail and supply chain logistics, long-term thinking helps solve complex decisions required for both near-term operational challenges and long-term strategic objectives. Reasoning technology can simultaneously predict and evaluate multiple scenarios, helping companies mitigate risk and address scalability challenges. This enables more accurate demand forecasting, streamlined supply chain routing, and procurement decisions that align with organizational sustainability initiatives.
For global companies, the technology can be used to draw up detailed business plans, generate complex code to debug software, or optimize the routes of trucks, warehouse robots, and self-driving taxis.
AI reasoning models are developing rapidly. OpenAI o1-mini and o3-mini, DeepSeek R1, and Google DeepMind's Gemini 2.0 Flash Thinking were all released in the past few weeks, and more new models are expected to be released soon.
These models require a large amount of computing during the inference process to reason about complex problems and produce correct answers. This means that companies need to expand accelerated computing resources to provide next-generation AI inference tools that can solve complex problems, write code, and plan multi-step processes.
