To map the space of AI research, you would: Map the space of all possible architectures, all possible architecture details, all possible learning algorithms, all possible objectives, all possible learning algorithm details, all possible data, all possible data details, all possible RL environments, all possible RL environment details, all possible search methods, all possible additions on top, all possible agentic harnesses, in all possible modalities, all possible tasks, all possible open problems, all possible philosophy ideas, etc. The space of all possible science of AI analyzing methods, all possible algorithms to analyze neural networks, all possible evaluations of the output of the neural networks, evals of the internals, reverse engineering of internals, all possible causal tests, from correlation to causality to counterfactual causality to simulations, all possible theory, etc. All possible variants, all possible SoTA details, etc. Map the math, physics, neuroscience, biology, chemistry, medicine, linguistics, complexity science, engineering, software engineering, hardware engineering, electrical engineering, STEM, philosophy, psychology, social science etc. space that generate modalities and tasks for AI, etc. Go though the space of all possible good science rigor standards, all possible failure modes, all possible ways to do a good experiment, all possible ways to write a good codebase and a paper, etc. Map the space of all possible open problems for all the above. Map the space of research ideas from the most incremental to the most transformational, from the most known to the most novel (combinations or novel structures), from the most in distribution to the most out of distribution, simplest to the most complex, etc.. Find combinations of these lego blocks. Building on top of something existing in the literature by mutating it or starting more from scratch.