Just as there are widely understood empirical laws of nature - for example, what goes up must come down, or every action has an equal and opposite reaction - the field of AI was long defined by a single idea: that more compute, more training data and more parameters makes a better AI model.However, AI has since grown to need three distinct laws that describe how applying compute resources in different ways impacts model performance. Together, these AI scaling laws - pretraining scaling, post-training scaling and test-time scaling, also called long thinking - reflect how the field has evolved with techniques to use additional compute in a wide variety of increasingly complex AI use cases.
The recent rise of test-time scaling - applying more compute at inference time to improve accuracy - has enabled AI reasoning models, a new class of large language models (LLMs) that perform multiple inference passes to work through complex problems, while describing the steps required to solve a task. Test-time scaling requires intensive amounts of computational resources to support AI reasoning, which will drive further demand for accelerated computing.
What Is Pretraining Scaling? Pretraining scaling is the original law of AI development. It demonstrated that by increasing training dataset size, model parameter count and computational resources, developers could expect predictable improvements in model intelligence and accuracy.
Each of these three elements - data, model size, compute - is interrelated. Per the pretraining scaling law, outlined in this research paper, when larger models are fed with more data, the overall performance of the models improves. To make this feasible, developers must scale up their compute - creating the need for powerful accelerated computing resources to run those larger training workloads.
This principle of pretraining scaling led to large models that achieved groundbreaking capabilities. It also spurred major innovations in model architecture, including the rise of billion- and trillion-parameter transformer models, mixture of experts models and new distributed training techniques - all demanding significant compute.
And the relevance of the pretraining scaling law continues - as humans continue to produce growing amounts of multimodal data, this trove of text, images, audio, video and sensor information will be used to train powerful future AI models.
Pretraining scaling is the foundational principle of AI development, linking the size of models, datasets and compute to AI gains. Mixture of experts, depicted above, is a popular model architecture for AI training. What Is Post-Training Scaling? Pretraining a large foundation model isn't for everyone - it takes significant investment, skilled experts and datasets. But once an organization pretrains and releases a model, they lower the barrier to AI adoption by enabling others to use their pretrained model as a foundation to adapt for their own applications.
This post-training process drives additional cumulative demand for accelerated computing across enterprises and the broader developer community. Popular open-source models can have hundreds or thousands of derivative models, trained across numerous domains.
Developing this ecosystem of derivative models for a variety of use cases could take around 30x more compute than pretraining the original foundation model.
Developing this ecosystem of derivative models for a variety of use cases could take around 30x more compute than pretraining the original foundation model.
Post-training techniques can further improve a model's specificity and relevance for an organization's desired use case. While pretraining is like sending an AI model to school to learn foundational skills, post-training enhances the model with skills applicable to its intended job. An LLM, for example, could be post-trained to tackle a task like sentiment analysis or translation - or understand the jargon of a specific domain, like healthcare or law.
The post-training scaling law posits that a pretrained model's performance can further improve - in computational efficiency, accuracy or domain specificity - using techniques including fine-tuning, pruning, quantization, distillation, reinforcement learning and synthetic data augmentation.
Fine-tuning uses additional training data to tailor an AI model for specific domains and applications. This can be done using an organization's internal datasets, or with pairs of sample model input and outputs.
Distillation requires a pair of AI models: a large, complex teacher model and a lightweight student model. In the most common distillation technique, called offline distillation, the student model learns to mimic the outputs of a pretrained teacher model.
Reinforcement learning, or RL, is a machine learning technique that uses a reward model to train an agent to make decisions that align with a specific use case. The agent aims to make decisions that maximize cumulative rewards over time as it interacts with an environment - for example, a chatbot LLM that is positively reinforced by thumbs up reactions from users. This technique is known as reinforcement learning from human feedback (RLHF). Another, newer technique, reinforcement learning from AI feedback (RLAIF), instead uses feedback from AI models to guide the learning process, streamlining post-training efforts.
Best-of-n sampling generates multiple outputs from a language model and selects the one with the highest reward score based on a reward model. It's often used to improve an AI's outputs without modifying model parameters, offering an alternative to fine-tuning with reinforcement learning.
Search methods explore a range of potential decision paths before selecting a final output. This post-training technique can iteratively improve the model's responses
