A new tool for generating artificial intelligence-based code, alphacodium , inspired by Google DeepMind Alphacode (and recently recently neglected alphacode 2 based on gemini), now they have changed them , causing a stormy discussion on Twitter this week.
“We are one step closer to the fact that AI begins to generate the code better than people!” – wrote Santiago Valdarram. “The results demonstrate that Alphacodium is the best approach to generation of the code that we have seen. It is ahead of Alphacode from Deepmind and their new Alphacode2, without requiring additional settings of the model!”
Andrey Karpati, former director of artificial intelligence in Tesla and Openai employee, focused on the method of “engineering of the flow” in the method The framework of the new tool for optimizing code generation. He described this approach as evolution from the “question-answer” base to the more advanced “flow paradigm”, where the answers are developed and specified through an iterative process.
To improve the work of large language models (LLMS) in the tasks related to encoding, AlphaCode engineering is expanding the technology of thought chains, including Gan architecture elements (developed by Yan Goodfellow in 2014), including the model that generates the code, and The opposing model that ensures the integrity of the code through testing, reflection and compliance with specifications.
The process begins with data entry and includes a number of preliminary steps, where Alphacodium reflects on the task and eventually comes to the first solution in the form of code. Then, additional tests are generated that help to clarify the solution, and the final, really working solution is achieved.
mission companies , according to its web The site is “to help developers create faster and without errors.” Alphacodium was tested at the Codecontestts data set containing about 10,000 programming tasks. Its performance at this benchmark showed an improvement in the accuracy of the GPT-4 from 19 to 44%. “This result is not just a numerical improvement; it is a breakthrough in the capabilities of LLMS in the generation of code, which sets a new standard in this area,” said Codiumai.
codiumi, founded in 2022 and attracted $ 10.6 million in March 2023, shared the alphacodium repository on github and