It has been argued that AI coding assistant, which has exceeded Antropic’s ‘Claude’, has emerged. The producers have heard of the mixture of Open AI’s reasoning model O1 as a secret.
American startup agement code recently launched a model called ‘Augment Agent’. Code repository GitHubIt was released through.
The official name of this model is ‘AUGMENT SWE-Bench Agent’. In other words, it’s an optimized model for ‘SWE-Bench’, which evaluates AI’s software engineering ability.
To this end, Claude, which is understood to have the most effective coding performance, was used because the default model. As well as, O1 combined with the very best success rate of 65.4%on the SWE-bench.
This exceeds 62.3%recorded by the newest model Claude 3.7 Sonnet.
The model is understood to concentrate on solving the complexity of huge -scale software engineering projects reasonably than easy code creation. Unlike other AI coding tools, it’s explained that developers have focused on helping developers to explore and modify large -scale existing codebases consisting of thousands and thousands of codes across several storage.
Scott Di Chen Ugment Code CEO said in an interview with Enterprise Beat, “The recent popular AI code assistant is concentrated on writing code called Vive Code.” Software engineering. ”
The corporate cited the power to grasp the context within the vast codebase. This agent boasts a 200,000 token context window larger than other corporations.
As well as, when the corporate collaborated on team units, it was also a strength to synchronize code changes in real time. Due to this, he said he had a 70%win rate within the competition with GitHub Copilot.
This includes the ‘memory’ function that reflects individual coding style and preference over time through interaction with developers through interaction with developers.
Integration of Claude and O1 is the result of assorted attempts. The corporate said that advanced technologies reminiscent of Claude’s ‘pondering mode’ didn’t make much improvement. As well as, the tactic of determining the output by multiple models is inexpensive. Due to this fact, it was simply found that attaching just one O1 was essentially the most efficient.
Founded in 2022, the agent code has been investing $ 270 million. Through Series B, which was closed in April last yr, the corporate’s value soared to $ 977 million. It also has corporations reminiscent of web flow, pigments, lemonade and DDN.
By Dae -jun Lim, reporter ydj@aitimes.com
