Japan's Supreme Court Rules AI Cannot Be Listed as Patent Inventor
Japan's top court has ruled that artificial intelligence cannot be credited as an inventor on patent applications, aligning with similar legal precedents in the US and EU. This decision underscores the global judicial consensus that inventorship is a human-centric legal status, even as AI increasingly assists in the generative phase of research and development. The ruling maintains a clear legal boundary between human intellectual property and machine-generated outputs.
Using DSPy to Optimize Specialized SQL Agents
Simon Willison details a practical implementation of DSPy to optimize the Datasette Agent, focusing on improving SQL system prompts. By leveraging programmatic prompt engineering rather than manual tweaking, this approach demonstrates how developers can use evaluation-driven frameworks to create more reliable specialized agents for data querying and analysis. The use of DSPy allows for systematic optimization of prompts, resulting in significantly higher accuracy for complex SQL generation tasks.
The Case Against One-Shot AI Design: Embracing Skill Engineering
Industry experts are advocating for a shift from one-shot AI design—the idea that a single model invocation can solve complex tasks—toward skill engineering. This approach treats AI capabilities as modular, iteratively improved skills that require constant human steering and judgment. By moving away from monolithic agent architectures toward granular, steerable skills, developers can better manage the inherent unpredictability of Large Language Models while building more robust and reliable agentic systems.
Autoresearch and the Software Factory: The Rise of Self-Improving Agents
The concept of autoresearch is emerging as a cornerstone of the next generation of AI development, focusing on feedback loops that allow agents to autonomously refine their internal recipes and reasoning steps. While these autonomous loops enable a software factory vision where agents handle more of the development lifecycle, the tension between automated agency and human control remains a central challenge. This movement emphasizes loopmaxxing—maximizing the iterative capacity of AI—while insisting that human understanding must remain the final arbiter of quality and intent.
ELDR: Optimizing Mixture-of-Experts Serving via Locality-Aware Routing
ELDR (Expert-Locality-Aware Decode Routing) introduces a significant optimization for Mixture-of-Experts (MoE) models serving in prefill-decode disaggregated environments. By predicting expert activations and routing requests accordingly, the framework reduces communication overhead and improves the efficiency of serving large-scale MoE models. This reduction in cross-node communication significantly improves throughput and reduces latency for large-scale MoE deployments, a critical advancement for enterprise-grade inference.
TurboServe: Efficient Infrastructure for Streaming Video Generation
TurboServe is a specialized serving system designed to handle the unique demands of streaming video generation. It addresses session state management and dynamic resource allocation challenges that traditional LLM serving systems are not equipped for. By integrating scheduling, autoscaling, and migration mechanisms, TurboServe offers an efficient and economical solution for real-time video tasks, which typically suffer from high computational overhead and stateful session requirements.
Seed2.0: Tackling Long-Tail Knowledge and Complex Reasoning
Seed2.0 introduces an upgraded model framework designed to push the intelligence frontier for complex, real-world tasks. The model focuses on mitigating long-tail knowledge gaps and improving instruction following, specifically targeting reasoning and visual understanding capabilities that are often lost in generic benchmarks but critical for actual user applications. Unlike standard benchmarks, Seed2.0 focuses on reasoning, search capabilities, and visual understanding tailored to user-centric needs, marking a shift toward models optimized for production environments.
AutoTrainess: Enabling Autonomous Agent-Driven Model Training
AutoTrainess presents a framework for LLMs training LLMs by using structured agent-computer interfaces (ACIs). By providing agents with better tools for planning, data preparation, and logging, the system outperforms traditional command-line scripts, moving the industry closer to fully autonomous model fine-tuning and optimization pipelines. The system enables agents to handle the entire training lifecycle more effectively than previous methods, suggesting a future where AI models are predominantly trained and tuned by other specialized agents.