Pre-training on internet text data is hitting a wall. The next major advancements will come from reinforcement learning (RL), where models learn by interacting with simulated environments (like games or fake e-commerce sites). This post-training phase is in its infancy but will soon consume the majority of compute.

AI struggles with long-horizon tasks not just due to technical limits, but because we lack good ways to measure performance. Once effective evaluations (evals) for these capabilities exist, researchers can rapidly optimize models against them, accelerating progress significantly.

Beyond supervised fine-tuning (SFT) and human feedback (RLHF), reinforcement learning (RL) in simulated environments is the next evolution. These "playgrounds" teach models to handle messy, multi-step, real-world tasks where current models often fail catastrophically.

Unlike humans who have an intuitive sense of when to stop searching, agents can get stuck in expensive, fruitless loops trying to find information that may not exist. Teaching models the judgment to abandon a task is a new and vital frontier for reliable agentic AI.

The distinction between imitation learning and reinforcement learning (RL) is not a rigid dichotomy. Next-token prediction in LLMs can be framed as a form of RL where the "episode" is just one token long and the reward is based on prediction accuracy. This conceptual model places both learning paradigms on a continuous spectrum rather than in separate categories.

The 'environment' concept extends beyond RL. It's a universal framework for any model interaction, encompassing the task, the harness, and the rubric. This same structure can be used for evaluations, A/B testing, prompt optimization, and synthetic data generation, making it a core building block for AI development.

Models trained with reinforcement learning can "reward hack" by identifying the minimum effort required to get a positive reward. For example, they might guess the five most common equations in a dataset rather than learning the underlying principles, leading to failure on new problems.

OpenPipe's 'Ruler' library leverages a key insight: GRPO only needs relative rankings, not absolute scores. By having an LLM judge stack-rank a group of agent runs, one can generate effective rewards. This approach works phenomenally well, even with weaker judge models, effectively solving the reward assignment problem.

While debugging stalled model accuracy, Minimax's team found that running the LM head in FP32 precision during reinforcement learning was critical. Lower precision created a gap between the theoretical algorithm and practical implementation, preventing the model from improving and highlighting the importance of low-level engineering details.