Tao Zhang
Technical Leader of Little Raccoon Family, SenseTime Technology
The technical leader of the Little Raccoon family of SenseTime Technology, leading the whole process of the application landing of the Little Raccoon family, has a deep understanding of the AI application landing and rich practical experience. With the code model as the product AI capability base, the Raccoon family on the one hand provides AI programming assistant products for the native developer scenario, Code Raccoon; on the other hand, it also breaks the circle of the code capability as an AI tool, expands the boundary of the product capability, and provides AI productivity applications for the office crowd, Office Raccoon. This also makes the Little Raccoon family become a highly distinctive AI productivity series products in China.
Topic
Prepare for an explosion of AI-native apps
Introduction: AI native applications are experiencing a systematic change from technological breakthroughs to industrial landing, we, as early developers of AI applications, have formed quite a lot of in-depth thinking from our past practices along the way to share and discuss with you as this change is approaching. Technical breakthrough dimension, large model capabilities through multi-modal fusion and continuous pre-training to achieve qualitative leap, reasoning costs due to sparse activation, model distillation and other technological breakthroughs down two orders of magnitude, combined with scenario-based data targeted tuning, AI began to penetrate the complex business logic. In the dimension of architectural evolution, the system design has shifted from a single model center to a distributed network of intelligences, the development paradigm has been reconfigured into a dual-cycle model of “hint engineering - feedback reinforcement”, and the hardware level has formed a deep synergy between the integrated architecture of storage and computation and algorithm lightweighting. On the ground practice path, the model continues to build “home field advantage” in the vertical field, the engineering architecture breaks through the scale bottleneck through dynamic load balancing, incremental learning and other mechanisms, and the cross-system interconnection gives rise to a new ecology of human-machine symbiosis. Outline: I. Technology breakthrough: the core driving force of AI native applications 1. qualitative leap in model capability 2. Sharp drop in reasoning cost 3. Effective use of scenario-based data II. Architecture Trends: Evolutionary Direction of AI Native Systems 1. From “Model Center” to “Network of Intelligent Bodies 2. Development Paradigm Reconfiguration 3. Hardware-algorithm co-evolution III. Landing Practice: Technology to Scene Penetration Paths 1. Continuous refinement of model home capabilities 2. Engineering architecture exploration and optimization 3. System Interconnection Expanding Infinite Possibilities VI. Risks and Challenges: Technology Evolution Reefs 1. Model reliability cliff 2. Lagging behind in credible technology
