Brother
Professional
- Messages
- 2,590
- Reaction score
- 539
- Points
- 113
Artificial intelligence that can replace professionals is a reality of the near future.
Google introduced AlphaCode 2, an improved version of its code generation model developed by the DeepMind lab about a year ago. AlphaCode 2 works on the basis of the Gemini model, or rather on its version, Gemini Pro, configured for data from programming contests. Google claims that AlphaCode 2 is significantly superior to its predecessor, especially in one particular benchmark.
AlphaCode 2, which uses Python, Java, C++ , and Go, performed better than 85% of participants in the Codeforces programming competition. This is in comparison with the 50% of participants who were surpassed by the previous AlphaCode.
The model was refined on an additional set of "higher quality" data, but judging by the scant details in the technical report, it is not entirely clear which data was used and to what extent. The researchers selected 12 recent competitions with more than 8,000 participants from sections 2 and more complex 1+2, which included 77 tasks. AlphaCode 2 solved 43% of the problems in 10 attempts, which is almost twice as much as the original AlphaCode (25%).
AlphaCode 2 is capable of solving problems with "complex" mathematics and theoretical computer science questions. It also knows how to apply dynamic programming, which is especially important because the tasks that require this method were difficult for the original AlphaCode.
To solve problems, AlphaCode 2 uses "policy models" that generate code for each task. It filters out code that doesn't match the task description and applies a clustering algorithm to group "semantically similar code samples". The evaluation model in AlphaCode 2 then selects the best solution for each of the 10 largest code clusters.
However, all AI models have drawbacks. AlphaCode 2 requires a lot of trial and error, is too expensive to operate, and relies heavily on the ability to filter out obviously bad code samples. Upgrading to a more powerful version of Gemini, such as the Gemini Ultra, can reduce some of these issues.
AlphaCode was never released initially, but according to DeepMind's vice President of product, Eli Collins, it is possible that AlphaCode 2 will be used in the future as a tool to facilitate the entire software development process.
Google introduced AlphaCode 2, an improved version of its code generation model developed by the DeepMind lab about a year ago. AlphaCode 2 works on the basis of the Gemini model, or rather on its version, Gemini Pro, configured for data from programming contests. Google claims that AlphaCode 2 is significantly superior to its predecessor, especially in one particular benchmark.
AlphaCode 2, which uses Python, Java, C++ , and Go, performed better than 85% of participants in the Codeforces programming competition. This is in comparison with the 50% of participants who were surpassed by the previous AlphaCode.
The model was refined on an additional set of "higher quality" data, but judging by the scant details in the technical report, it is not entirely clear which data was used and to what extent. The researchers selected 12 recent competitions with more than 8,000 participants from sections 2 and more complex 1+2, which included 77 tasks. AlphaCode 2 solved 43% of the problems in 10 attempts, which is almost twice as much as the original AlphaCode (25%).
AlphaCode 2 is capable of solving problems with "complex" mathematics and theoretical computer science questions. It also knows how to apply dynamic programming, which is especially important because the tasks that require this method were difficult for the original AlphaCode.
To solve problems, AlphaCode 2 uses "policy models" that generate code for each task. It filters out code that doesn't match the task description and applies a clustering algorithm to group "semantically similar code samples". The evaluation model in AlphaCode 2 then selects the best solution for each of the 10 largest code clusters.
However, all AI models have drawbacks. AlphaCode 2 requires a lot of trial and error, is too expensive to operate, and relies heavily on the ability to filter out obviously bad code samples. Upgrading to a more powerful version of Gemini, such as the Gemini Ultra, can reduce some of these issues.
AlphaCode was never released initially, but according to DeepMind's vice President of product, Eli Collins, it is possible that AlphaCode 2 will be used in the future as a tool to facilitate the entire software development process.
