Google Introduces Alphacode 2 for Generating Code
Google has unveiled Alphacode 2, an upgraded version of its code generation model developed by the Deepmind laboratory. Alphacode 2 is built on the Gemini model, specifically the Gemini Pro edition which is designed for programming contest data. According to Google, Alphacode 2 surpasses its predecessor, particularly in a specific test.
During software competitions on the CodeForces Alphacode 2 platform, Alphacode 2 outperformed more than 85% of participants in Python, Java, C++, and GO languages. In comparison, the previous version, Alphacode, only surpassed 50% of participants.
The model underwent further refinement by utilizing an additional data set of “higher quality”. However, details regarding the specific data used and its volume remain unclear. Researchers selected 12 recent competitions containing over 8,000 participants from the more difficult sections 2 or above, which included 77 tasks. Alphacode 2 successfully solved 43% of the tasks in 10 attempts, almost double the success rate of the original Alphacode (25%).
Alphacode 2 is equipped to handle intricate mathematical and theoretical problems in computer science. It also possesses the ability to apply dynamic programming, a method that posed challenges for the original Alphacode.
To solve problems, Alphacode 2 utilizes “Policy models” that generate code for each task. It eliminates code that does not align with the problem description and employs a clustering algorithm to group “semantically similar code samples”. The evaluation model in Alphacode 2 then selects the best solution from the top 10 code clusters.
However, like all AI models, Alphacode 2 has its limitations. It requires a significant number of samples and is susceptible to errors, making it costly to operate. Additionally, it heavily relies on the ability to filter out poor code samples. Transitioning to a more advanced version of Gemini, such as Gemini Ultra, could help alleviate some of these challenges.
While Alphacode was never publicly released, Eli Collins, Vice President of Deepmind, suggests that Alphacode 2 may be used as a tool to enhance the software development process in the future.