We are in the "gentleman scientist" era of AI research (www.seangoedecke.com)

The piece argues that AI research is in a "gentleman scientist" phase: many high-impact discoveries are surprisingly simple ideas or engineering tricks applied to large language models (LLMs), so productive contributions don’t always require decades of domain immersion. Rather than being inscrutable, much of the math-heavy literature masks concepts that are easy to implement (or at least experiment with) in code. That democratization matters because it accelerates iteration and discovery—informal experimentation by engineers, hobbyists, and small teams is producing real capability advances and new avenues for research that professional labs then refine. Concrete examples show why: group-relative policy optimization (GRPO) replaces costly, hard-to-train critic models by using the model’s own average performance as a baseline—a simple idea with big RL payoff for hard prompts. Anthropic’s “skills” (on-disk scripts/markdown) and Recursive Language Models (agents with code access to their full prompt) are likewise low-complexity, high-leverage innovations that let LLMs act as richer agents and tools. The implication for the community is twofold: encourage more informal, exploratory work (which finds surprising capabilities and practical tricks), and recognize that while many hard, formal problems remain, a long tail of accessible questions will continue to yield outsized returns.