Baidu Unveils AI That Runs Its Own Experiments
Baidu's ERNIE Evolve system picks research ideas, edits code and tests changes to improve AI models.
Updated on Nov. 9, 2026 in Artificial Intelligence
Live Poll
Do you think AI systems should be allowed to improve themselves?
Baidu unveiled ERNIE Evolve, an AI system that selects research ideas, edits code and runs tests to improve AI code. The company says the system sped up its in-development ERNIE-5.5-Flash model.
Why it matters
The system keeps only changes that pass separate checks the AI cannot see, which points to AI tools that help build and refine other AI models. Baidu reports the results in its own announcement, so the gains have not been independently verified.
Baidu reports ERNIE-5.5-Flash text generation was more than six times faster on Kunlunxin P800 chips and throughput rose 33.7% in about three days on tuned GPUs. Much of the Kunlunxin figure came from enabling features the chip setup was not using, and the two results are not directly comparable.
The players
Baidu
Baidu is a major Chinese technology company that operates a large search engine and develops the ERNIE family of AI models.
The details
ERNIE Evolve works by choosing research directions, modifying code and running experiments, then retaining only changes that pass evaluation checks hidden from the system itself. Baidu says it beat the best public results on two of three public AI research challenges.
Timeline
Baidu unveiled ERNIE Evolve on November 9, 2026.
GPU throughput gains were measured over about three days.
The Tech Race
ERNIE Evolve extends the automated model-design approach pioneered in Google's AutoML neural architecture search research, moving from searching architectures to letting an AI edit code and run experiments. The move positions Baidu in a race with rival labs to automate their own AI development pipelines.
Developers using Baidu's ERNIE models could eventually see faster model updates if this approach reaches commercial products. Speed claims are company-reported, so users should wait for independent benchmarks before assuming the gains apply to their own workloads.
The takeaway
Company-reported speed gains from self-improving AI systems are promising but need independent, like-for-like verification before they change how models are built.
Further reading
Explore more on Artificial Intelligence to see how automated research systems are reshaping model development.
Live Poll
Do you think AI systems should be allowed to improve themselves?
