AI Model Training

AI model training is the computational process of teaching neural networks to perform tasks by iteratively adjusting their parameters (weights) based on data. For modern large language models, training is a multi-stage pipeline that consumes extraordinary resources and has become one of the defining engineering challenges of the era.

The training pipeline for frontier models typically has three stages. Pre-training exposes the model to trillions of tokens of text (and increasingly images, audio, and video), teaching it to predict the next token. This is the most compute-intensive phase—frontier models require thousands of GPUs running for months, connected by high-speed networks, consuming megawatts of power. Fine-tuning adapts the pre-trained model for specific tasks or behaviors using smaller, curated datasets. Alignment (via RLHF, DPO, or Constitutional AI) shapes the model's outputs to be helpful, harmless, and honest.

The economics of training define the AI industry's structure. Pre-training a frontier model costs $100 million to $1 billion+ in compute alone. This creates a natural oligopoly of organizations that can afford frontier training: Anthropic, OpenAI, Google, Meta, and a handful of others. But the cost of fine-tuning and reinforcement fine-tuning is orders of magnitude lower, enabling a long tail of specialized models built on open-weight foundations.

Training is also where the megascale datacenter challenge originates. The exponential growth in training compute—roughly 4x per year for frontier models—drives demand for HBM, custom silicon, advanced cooling, and increasingly, dedicated power generation including nuclear. Training is the furnace that forges AI capability, and its resource requirements are reshaping energy infrastructure worldwide.

Pretraining versus post-training

Model developers increasingly describe the pipeline as two phases rather than three. Pretraining is the single, expensive pass that produces a general next-token predictor. Everything afterwards—supervised fine-tuning, preference alignment, reinforcement learning against graders, distillation—is grouped as post-training. The DeepSeek-V4 technical report (April 2026) uses exactly this framing: both models were pre-trained on more than 32 trillion tokens, "followed by a comprehensive post-training pipeline that unlocks and further enhances their capabilities".

The two phases differ by orders of magnitude in cost, which is why post-training is where most organisations can participate. The Recursive Language Models paper (v3, May 2026) post-trained an 8B open model for a new skill with about 1,000 examples and 48 H100 GPU-hours. Methods at this end of the pipeline include adapter training with LoRA and other parameter-efficient fine-tuning, reinforcement fine-tuning with GRPO inside purpose-built RL environments, and model merging, which combines already-trained weights without further gradient steps. Post-training has its own failure modes, covered separately: forgetting earlier abilities (continual learning), degradation from training on a model's own output (model collapse), and reward hacking when the grader can be gamed.