Summary of Conference Call on Robotics Domain Controller Industry Trends Industry Overview - The conference discusses the robotics domain controller industry, highlighting the transition of automotive intelligent driving domain control solutions to robotics platforms, aiming to create an efficient, safe, and open robotics controller, referred to as the "brain" platform [1][2] - Rapid advancements in artificial intelligence, particularly large language models (LLMs), are driving the growth of the robotics industry, similar to the developments seen in the autonomous driving sector a decade ago [1] Key Insights and Arguments - The acquisition of an integrated joint company by the company lays the foundation for testing and exploring new directions in robotics [1][2] - The introduction of high-performance chips like NVIDIA's Orin chip is being adopted by domestic companies such as Desay SV and Joyson Electronics, indicating strong market recognition and expectations for NVIDIA's technological solutions [1][5] - Domestic intelligent driving chip systems, including Horizon Robotics and Huawei, are reshaping the edge chip market landscape, suggesting that the robotics sector is becoming a new ecological soil for competition [1][6] - Horizon Robotics has shown promising performance in edge testing, with significant computational output efficiency, indicating substantial future potential [7] - The current market for robotics domain controllers is expected to be competitive, with multiple companies driving industry development rather than being dominated by a few players [6][10] Future Business Models and Competitive Landscape - The future business model and competitive landscape for robotics domain controllers are still uncertain, with significant differences from the automotive industry, which has a long history [4] - The rapid iteration of AI technology is expected to influence the robotics industry, but whether the controller segment will mirror the automotive sector remains to be seen [4] - Companies are encouraged to be pioneers in the industry to secure strategic advantages, as the fast pace of AI development necessitates adaptability to avoid obsolescence [4] Technological Developments - The demand for high-performance platforms is increasing due to advancements in technologies like Google DeepMind's Gemini 5.0, which showcases complex brain functions based on natural language processing [3][8] - The robotics domain controller's complexity is heightened by the lack of complete standardization, with ongoing changes in hardware and parameters [9] Domestic Chip Companies and Market Dynamics - Optimism is expressed regarding the capabilities of domestic chip companies, with Horizon Robotics expected to gain more opportunities amid U.S.-China tech competition [7] - The market is anticipated to be characterized by multi-party competition rather than a one-sided dominance [7] Collaborations and Partnerships - The company maintains close collaborations with leading humanoid robot enterprises in China, demonstrating proactive market engagement and significant progress in joint efforts [13] - The partnership with Diguang Robotics is highlighted, showcasing successful deployment of open-source models and the development of generalized kits based on Diguang chips [12] Conclusion - The robotics domain controller industry is poised for significant growth driven by technological advancements and competitive dynamics, with a focus on collaboration and innovation to navigate the evolving landscape [1][4][12]

Core Insights - The article introduces Mixture-of-Lookup-Experts (MoLE), a new architecture designed to optimize the deployment of Mixture-of-Experts (MoE) models, particularly in resource-constrained environments [1][28] - MoLE addresses the challenges of high memory usage and transmission delays associated with traditional MoE during inference by replacing matrix operations with lookup tables [28] Group 1: MoLE Architecture - MoLE activates only a small subset of experts needed for each token during inference, significantly reducing computational load while maintaining a large parameter scale [1] - The architecture allows for the pre-computation of input-output mappings stored as lookup tables, enabling efficient retrieval during inference [5][6] Group 2: Training Phase Differences - In the training phase, MoLE modifies the input to routed experts from the previous layer's output to shallow embedding tokens, facilitating the pre-computation and storage of lookup tables [8] - MoLE employs an activation strategy that activates all routed experts during training, eliminating the need for sparse activation to control computational load [9] - The loss design in MoLE focuses solely on language modeling loss, without additional load balancing loss terms [10] Group 3: Inference Phase Process - During inference, MoLE constructs lookup tables from the embedding layer's weight matrix, allowing for direct retrieval of expert outputs based on token IDs [15] - The lookup table is stored in lower storage devices, and during inference, the corresponding expert outputs are retrieved and loaded into memory for computation [16] Group 4: Performance and Efficiency - MoLE's computational complexity during inference is comparable to dense models and traditional MoE models, while significantly reducing transmission overhead [17] - Experimental results indicate that MoLE achieves performance on par with MoE while drastically reducing transmission costs by over a thousand times [20][28] Group 5: Experimental Results - The experiments conducted on the Pile dataset show that MoLE maintains performance equivalent to MoE while using the same training parameters and inference activation parameters [20] - MoLE demonstrates lower inference latency compared to MoE, especially in batch decoding scenarios, highlighting its advantages in high-throughput tasks [28]