Baidu has unveiled ERNIE Evolve, an AI system designed to automate the research and improvement of AI code, which has already accelerated the performance of Baidu’s in-progress ERNIE-5.5-Flash model by over 30% on GPUs and more than sixfold on its domestic Kunlunxin chips.
- ERNIE Evolve sped up ERNIE-5.5-Flash by 33.7% on GPUs and 6x on Kunlunxin chips
- System autonomously edits AI code, runs experiments, and filters results
- Baidu plans to expand use to model training and automated test design
What happened
Baidu announced ERNIE Evolve, an autonomous AI research agent capable of running experiments to improve AI codebases independently. This new system evaluates and edits AI model code, retaining only changes that pass rigorous, separate quality checks. Using ERNIE Evolve, Baidu accelerated the development of its upcoming ERNIE-5.5-Flash model, achieving a 33.7% throughput increase on GPUs in about three days of optimization.
Even more strikingly, on the Kunlunxin P800 chips developed by Baidu's chip unit, the system improved the model's text generation speed by over six times. While part of this gain came from enabling previously unused chip features, the results showcase the potential of automating AI model refinement processes. Additionally, Baidu benchmarked ERNIE Evolve on three public AI research challenges, surpassing best public records on two of them.
Why it matters
The introduction of ERNIE Evolve underscores a broader industry shift toward self-improving AI, which can accelerate development cycles and optimize performance without constant human intervention. This approach is competitive with similar efforts from other AI leaders; for instance, startups and companies like Anthropic and Google report AI systems contributing extensively to their own model code and successor development.
By leveraging autonomous AI-driven research, Baidu aims to refine its models faster and more efficiently, potentially lowering costs and gaining speed advantages in the competitive field of large language models. The integration of such systems on domestic hardware like Kunlunxin chips carries additional significance in China's strategic pursuit of AI innovation and semiconductor autonomy.
What to watch next
Baidu plans to extend ERNIE Evolve’s capabilities beyond code improvements to include model training enhancements and the AI-driven design of test cases. Monitoring how Baidu manages control and oversight of these increasingly autonomous systems will be important, since the company emphasizes the necessity of human supervision to ensure transparency and safety.
Further developments could reveal how well Baidu's AI research agent scales to complex tasks and whether it can maintain or improve quality standards as it expands its scope. The interplay between hardware and software improvements, especially on Kunlunxin chips, will also be critical in assessing the overall impact on China’s AI ecosystem and global AI research competition.