Genie

What is Genie

Genie is launched by Cosine AIAI programming assistant, capable of solving software programming problems as quickly as humans, completing code writing and execution in just 84 seconds. Genie can mimic the thinking process of human programmers, automatically parsing problems, iteratively analyzing, and writing code. Genie leads by a wide margin on the SWE-Bench benchmark with a 30.07% resolution rate. Genie was trained using a large amount of real human programming data, including methods such as outcome analysis and static analysis, as well as self-improvement mechanisms.

Genie's main features

• Automatic problem parsing: Genie can understand and parse problem descriptions from sources such as GitHub Issues to determine the software problem that needs to be solved.
• Iteratively thinking about required files: It can automatically determine the code files needed to solve the problem and iteratively think to find the best solution.
• Automatic iterative analysis: Genie conducts in-depth analysis of the problem and continuously iterates to optimize the solution.
• Automatic code writing: Based on the analysis results, Genie can automatically write code to solve the problem.
• Automatic code execution: After writing is completed, Genie can automatically run the code and debug and correct it when errors occur.
• Self-improvement mechanism: Genie uses erroneous code generated by older versions as training data to self-learn and improve in this way.
• Multiple ways to trigger work: Genie's workflow can be started in multiple ways, such as prompts, GitHub Issues, Linear Tickets, or APIs.

Genie's technical principles

• Cognitive simulation: Genie simulates the cognitive process of human engineers, including problem decomposition, information retrieval, logical reasoning, and decision-making.
• Proprietary dataset: Genie is trained on data generated with Cosine AI's proprietary technology, including methods such as outcome analysis and static analysis, as well as self-improvement mechanisms.
• Progressive learning: Genie's training emphasizes incremental knowledge discovery and step-by-step decision-making, enabling it to gradually build solutions when facing new problems.
• Context understanding: Genie can understand and process long-context information, achieved through a large-context model that can handle large amounts of token data.
• Agentic characteristics: Genie is autonomous, able to independently respond to input information and operate in a logical manner.
• Iterative optimization: Genie is trained through self-improvement, using synthetic data to guide the model toward the target state and improve its problem-solving ability.

Genie's official website address

Genie's application scenarios

• Software development: Genie can be applied to all stages of software development, including requirements analysis, design, coding, testing, and maintenance.
• Code generation: Genie can automatically generate code based on given specifications or problem descriptions, improving development efficiency.
• Defect repair: Genie supports analyzing existing code, identifying and fixing potential errors and defects.
• Code review: Based on logical reasoning ability, Genie supports code review to ensure code quality and consistency.
• System refactoring: When an existing system architecture or codebase needs improvement, Genie can provide refactoring suggestions and implementation strategies.