Reinforcement learning achieves superhuman results not by inventing alien concepts, but by surfacing and combining rare behaviors that are already possible within a model's vast pre-trained distribution. The goal of pre-training is to make this search for novel solutions more efficient and less random.

RL fine-tuning is less likely to cause catastrophic forgetting than SFT because it works within the model's existing pre-trained pathways, or "grooves." SFT, by contrast, makes much larger weight updates that can aggressively overwrite and destroy latent knowledge.

Human personality development provides a direct analog for training LLMs. Just as our genetics, environment, and experiences create stable behavioral patterns ('personality basins'), the training data and reinforcement learning (RLHF) applied to LLMs shape their own distinct, predictable personalities.

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.

Poolside's Laguna S model shows significant gains from improved behaviors—like persistence, verification, and not declaring victory early—rather than just more parameters or "intelligence." This suggests that smaller models can become far more capable at knowledge work by optimizing these behavioral traits through post-training.

Pre-trained models ingest knowledge from both experts and novices. A key function of RL, especially in its early stages, is to "sharpen the distribution" by tuning the model to consistently adopt the persona of an expert who provides correct answers, not a student who is still learning.

AI development is inefficiently split into pre-training (optimizing for compression) and RL (optimizing for tasks), where RL often invalidates pre-training metrics. Combining these into a unified, end-to-end learning algorithm focused on final outcomes could yield an order-of-magnitude improvement in training efficiency.

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.

Instead of only analyzing a fully trained model, "intentional design" seeks to control what a model learns during training. The goal is to shape the loss landscape to produce desired behaviors and generalizations from the outset, moving from archaeology to architecture.