(The white code comes from the AlphaCode artificial intelligence system, while the purple code is written by humans to solve similar problems.) Worldwide, there is a shortage of programmers. If AI can write code - as long as human beings tell AI what they want to do through the program, it will bring a lot of convenience. According to the research published in the latest issue of Science, a new AI system called "Alpha Code" is bringing mankind closer to this vision. The system comes from the "deep thinking" research laboratory, and its overall performance is equivalent to that of a novice programmer who has been trained for several months to a year. In the test, "alpha deduction" can achieve "performance close to human level" by predicting code segments and creating millions of potential solutions, and can also solve previously undetected natural language problems in the competition. After generating too many schemes, "alpha deduction" will filter them into up to 10 schemes, all of which are generated "without any built-in knowledge about computer code structure". In the recent coding simulation evaluation of Codeforce competitive coding platform, "alpha deduction" ranks in the top 54.3% of the 10th generation solutions for each problem on average, and 66% of the problems are solved the first time they are submitted. Ziko Colt, a professor at the Bosch Artificial Intelligence Center of Carnegie Mellon University in the United States, said: "No matter how much 'alpha deduction' truly understands this task, it performs very well in the coding challenge that has never been seen before." "Alpha deduction" is not the only AI model that can code, but its performance exceeds that of the "predecessor" Codex - the system released by OpenAI Research Laboratory in 2021. GitHub also has its own popular AI programming tool Copilot. However, in terms of solving complex competitive problems, neither AI has shown the ability to compete with humans. Deep Thinking said that after training, "Alpha deduction" solved about 34% of the designated problems. To further test its capabilities, the company let it participate in the online coding competition. In the competition with at least 5000 participants, the performance of "alpha deduction" exceeded 45.7% of programmers. The researchers also compared it with the program in the training database and found that it did not copy large pieces of code or logic, which indicates that it has amazing creativity. In addition to improving overall productivity, "alpha deduction" makes programming easier for a new generation of developers. Its future may one day change people's programming philosophy, that is, human beings are mainly to formulate problems, and then AI will solve problems. (Liu Xinshe)