[Tencent Hyra 2.0 Moves Towards Autonomous AI Research: Training Models from Scratch, Designing Robots, and Even Participating in the Development of Next-Generation AI] Beating AI News reports that Tencent's Hyra 2.0 research agent expands its research focus from repeatedly testing and failing on single problems to developing scientific models, algorithms, and tools that can be reused long-term. To enable AI research to continue for days or even weeks, Hyra 2.0 distributes work among three layers of agents: Overall, Mission, and Worker, responsible for resource coordination, research planning, and experiment execution, respectively. Tasks are persistently saved, allowing other research to proceed while awaiting training results. Each conclusion requires supporting evidence such as code and logs, and is reviewed by an independent agent when necessary. Effective experiences are also accumulated into Skills for use in subsequent tasks. Tencent disclosed that Hyra 2.0 has already trained two protein models from scratch. Among them, HyInv, with 1.96 million parameters, can design protein sequences based on a target backbone, generating them 57 times faster than the commonly used ProteinMPNN. In preheating tests on a single H20 GPU, generating a sequence of approximately 400 residues (128 sequences) took only about 50 milliseconds. Another model, HyPLM-Flash, has 650 million parameters and uses 726 billion training tokens. The weights, inference code, and technical reports for both models are publicly available. Hyra 2.0 also participated in optimizing the open-source RISC-V processor "Xiangshan," improving SPEC CPU2006 performance per GHz by 5.79% with only a 0.3% increase in chip area. It also designed a "spiral ladder hand" with five fingers of increasing length. Initially about 11% slower than the original optimal design, after training with the same budgeted control algorithm, the simulated rotation speed was actually 28.6% faster. Currently, Hyra 2.0 is also involved in the pre-training, post-training, and inference system optimization of next-generation mixed-mode models, attempting to improve AI through enhanced R&D tools to help train stronger next-generation AI. [Original Link]