Baidu has introduced ERNIE Evolve, an artificial intelligence system capable of conducting its own research cycles to enhance AI model performance. The company announced on Monday that the system has already delivered measurable improvements to ERNIE-5.5-Flash, its forthcoming model. According to Baidu's measurements, the system achieved a 33.7% throughput increase on GPUs and more than a sixfold acceleration on Kunlunxin processors. (All performance figures come from Baidu's own testing.)
ERNIE Evolve operates by identifying research directions, modifying code, and executing validation tests. The system retains only those modifications that successfully pass independent verification processes that the AI itself cannot access. During trials on ERNIE-5.5-Flash, the model generated text six times faster when running on P800 chips from Kunlunxin, Baidu's chip manufacturing subsidiary that is preparing for a public listing. On GPUs that engineers had already optimized, the system achieved its 33.7% improvement in roughly three days.
The Kunlunxin performance gains warrant additional perspective. According to China AI Dispatch, a significant portion of the speedup resulted from enabling chip features that the system had not previously activated, making the comparison with GPU results less direct than a straightforward performance benchmark.
Baidu also evaluated ERNIE Evolve against three publicly available AI research benchmarks. The company reports that it achieved top results on two of these challenges, including a 38.9-second completion time on the NanoGPT Speedrun test, surpassing the previous record of 39.9 seconds. Performance on the third benchmark came close to existing records, though China AI Dispatch noted that Baidu has not yet submitted this result to the official leaderboard.
Baidu's performance metrics also exceed those of Recursive, a startup whose AI system reportedly contributed substantially to building its Flint coding model. The broader competition to develop AI systems that can improve other AI systems is intensifying. Anthropic has stated that Claude writes more than 80% of its code, and Sergey Brin has indicated that Gemini now assists in developing its successor.
Baidu intends to extend ERNIE Evolve's application to model training workflows next, with plans eventually to enable the system to design its own test cases. The company emphasizes that this form of self-directed improvement must remain subject to human oversight. "Its scope should grow only while people can still understand, supervise, and stop the system," Baidu wrote.
Source: The Next Web
