Imagine a world where AI could solve complex puzzles and find answers to questions with the same efficiency and accuracy of a computer algorithm. That's the vision behind a new research project that aims to automate the "thought" process of search, a fundamental aspect of many AI reasoning tasks. Search, in its essence, is about exploring a vast space of possibilities to find a solution that meets specific criteria. Whether it's a chess computer evaluating moves or a route planner finding the shortest path, search algorithms are the backbone of decision-making in many AI applications. But these algorithms typically rely on rigidly defined rules and data structures. What if the search process itself could be more flexible and adaptable? This is where "Thought of Search" (ToS) comes in. ToS combines the power of large language models (LLMs), like those behind ChatGPT, with the rigor of traditional search algorithms. LLMs excel at understanding natural language and generating code, allowing them to define the search space in a more flexible and adaptable way. However, existing approaches often require human intervention to ensure the search process is both sound (it only finds valid solutions) and complete (it finds all possible solutions). The new research takes ToS a step further by automating the entire process. Called AutoToS, it takes the human out of the loop by using automated feedback and unit tests to guide the LLM. This is a significant advancement because it removes a major bottleneck in using LLMs for complex search tasks. AutoToS works by giving the LLM a problem described in natural language. The LLM then generates code that defines the search space, including the rules for moving from one state to another and the conditions for reaching a goal state. The clever part is that AutoToS uses automated tests to check whether the LLM's generated code is sound and complete. If errors are found, feedback is provided to the LLM, which then refines its code. This iterative process continues until the LLM produces code that consistently generates correct solutions. The result? AutoToS achieves 100% accuracy on a range of challenging search problems, including classic AI puzzles like BlocksWorld, games like the 24 Game, and even complex tasks like logical reasoning. The implications of this research are far-reaching. Automating the "thought" process of search opens up new possibilities for using LLMs in a wider array of applications, from solving complex scientific problems to designing more efficient algorithms. While challenges remain in generalizing this approach to even more complex scenarios, this work represents a significant leap towards more autonomous and intelligent AI systems.