Does the end that Hua Weihai considers automatic road-sense arrive how do carrying large model train?

Hua Weihai thinks of the end of automatic road-sense to be based on WEWA framework to train to end large model, create fictitious setting through engine of world of high in the clouds, combine aggrandizement study mechanism, use in Che Duanli be born formerly base model and much expert mix a framework, be based on much mode to feel data to undertake training, specific as follows:

• Setting of high in the clouds is generated learn with aggrandizement: Engine of world of Hua Weiqian elder brother (WE) use diffuse make model technology, the ground accuses to generate huge extreme to drive in high accuracy of high in the clouds, Gao Ke setting, of setting of its key bad precedent generate density more traditional pattern promoted 1000 times. In the meantime, compose is built ” safe and preferential ” aggrandizement study mechanism, the rewards and punishment that makes clear for set of AI intelligence system is regular. Make those who accord with highest and safe level is decision-making obtain to drive, if behavior appears,safe hidden trouble gets puissant penalty. Plant through this ” AI trains AI ” process, the model is turned inside the viewpoint of value that drives safety decision-making core. Up to April 2025, the setting of bad precedent of class of high speed L3 that already finished more than 600 million kilometer in WE is emulated with safe test and verify.
• Mode data trains Che Duanquan: Model of action of Che Duanshi group (WA) be wisdom drive be born formerly base model, be based on confluence of car much sensor (vision, hearing) information of collection complete mode perception undertakes direct training. These primitive perception data are passed ” Token is changed ” processing, compose is built go out to drive for intelligence the model of original unripe base with custom-built body of core task quantity, ensure the model is dynamic to environment of car inside and outside have deep understanding, reduce demand of Shi Yanhe resource.
• MoE framework expert is decision-making train: WA uses much expert to mix (MoE) framework, face complex road besides when, can feel information according to real time, millisecond class essence activations definitely and call most those who be good at treating current setting is specific ” expert ” child the model comes dominant is decision-making. These ” expert ” the model is respective essence of life at specific domain, be like nightly drive, wet drives etc, can improve decision-making efficiency, prevent consideration redundant.
• Output orbit and intent training: The output of WA is car travel contrail not only, still include setting intent information, be like ” preparation avoids let on the right side of bicycle ” etc, enhance a driver to feel to systematic accredit with this. In training a process, the model can be optimized ceaselessly, make the contrail of output fluenter, safe, the intent is clearer and reasonable, implementation from the model copy mankind drives to break through human reaction limit and the qualitative change that sentence ability beforehand.