A Rice PhD showed that training a vision model on a game like Snake, while prompting it to see the game as a math problem (a Cartesian grid), improved its math abilities more than training on math data directly. This highlights how abstract, game-based training can foster more generalizable reasoning.

The path to a general-purpose AI model is not to tackle the entire problem at once. A more effective strategy is to start with a highly constrained domain, like generating only Minecraft videos. Once the model works reliably in that narrow distribution, incrementally expand the training data and complexity, using each step as a foundation for the next.

Demis Hassabis describes an innovative training method combining two AI projects: Genie, which generates interactive worlds, and Simmer, an AI agent. By placing a Simmer agent inside a world created by Genie, they can create a dynamic feedback loop with virtually infinite, increasingly complex training scenarios.

Modern LLMs use a simple form of reinforcement learning that directly rewards successful outcomes. This contrasts with more sophisticated methods, like those in AlphaGo or the brain, which use "value functions" to estimate long-term consequences. It's a mystery why the simpler approach is so effective.

In domains like coding and math where correctness is automatically verifiable, AI can move beyond imitating humans (RLHF). Using pure reinforcement learning, or "experiential learning," models learn via self-play and can discover novel, superhuman strategies similar to AlphaGo's Move 37.

The transition from supervised learning (copying internet text) to reinforcement learning (rewarding a model for achieving a goal) marks a fundamental breakthrough. This method, used in Anthropic's Opus 3 model, allows AI to develop novel problem-solving capabilities beyond simple data emulation.

Karpathy identifies the AI community's 2010s focus on reinforcement learning in games (like Atari) as a misstep. These environments were too sparse and disconnected from real-world knowledge work. Progress required first building powerful representations through large language models, a step that was skipped in early attempts to create agents.

Just as crawling is a vital developmental step for babies even though adults don't crawl, some learning processes that AI can automate might be essential for cognitive development. We shouldn't skip steps without understanding their underlying neurological purpose.

Unlike traditional software, large language models are not programmed with specific instructions. They evolve through a process where different strategies are tried, and those that receive positive rewards are repeated, making their behaviors emergent and sometimes unpredictable.