Alibaba Damo Academy's AI Agent Discovers Four New Superconductors
According to SCMP, Alibaba's research division, Damo Academy, has discovered four new superconductors using an AI agent called 'Elements Claw.' The news is drawing attention as a case in which AI has directly contributed to the discovery of new materials in materials science.
Superconductors are materials that can carry electrical current with zero resistance, and they hold promise for a wide range of applications, including power transmission, medical MRI machines, and quantum computers. Discovering new superconductors, however, has traditionally been a time-consuming and costly process. By narrowing down candidate materials and identifying which ones should be tested, AI agents could dramatically accelerate this process.
While this discovery won't immediately translate into industrial applications, it is an important milestone showing that AI is becoming a powerful tool in the hypothesis generation and candidate exploration stages of scientific research.
Nvidia Emerges as the 'Bank' Powering the AI Boom
According to Startup Fortune, Nvidia has effectively taken on the role of extending financing to AI companies through the provision of GPUs, making it something of a 'bank for the AI boom.'
Nvidia invests in AI startups by offering priority access to GPU clusters — a structure quite different from the traditional venture capital model of investing cash. Leveraging its position as the supplier of GPUs, the 'oil of the AI era,' Nvidia is betting directly on the growth of the AI industry as a whole.
This strategy feeds directly into Nvidia's own revenue growth, but it also carries a structural risk: the supply of AI infrastructure is becoming heavily dependent on a single company. From a competition standpoint, voices are increasingly calling for close scrutiny of the influence Nvidia's market power exerts.
An 'Application' Design That Makes Small Models Efficient for Agent Use
An interesting approach for optimizing small, locally-run LLMs for agent use cases was shared in Reddit's LocalLLaMA community.
The technique groups the tools given to an agent into independent views called 'applications,' where each application holds only a limited toolset and a small context. For example, the web browsing app gets just three tools, while a PC control app provides 20 tools through a dedicated interface. Each time the agent switches apps, the previous context is swapped out and restored to its original state when it returns.
According to the poster, when they tried this setup with Gemma 4 E4B (a small model), it actually performed better than Gemma 4 26B (a large model). The larger model reportedly tended to shy away from the planning tools. By reducing the number of tools and simplifying menu actions into a 'verb + number' format, even the small model could accurately handle URLs and text.
The setup reportedly runs locally on an RX6600XT (Vulkan) at 70-85 tok/s with a context of around 10k tokens. For developers looking to run agents on edge devices or low-spec hardware, this looks like a practical reference case.
Authentication and Authorization Challenges in the AI Era
An article from FusionAuth takes on the design challenges of authentication and authorization in AI systems.
As AI agents increasingly access various APIs and services on users' behalf, new problems are emerging that traditional OAuth flows cannot handle. Mechanisms for managing under whose authority and to what extent an agent can act are one of the key infrastructure challenges that come with the spread of AI.
The article organizes practical implementation considerations, including identity management for AI agents, dynamic scope control, and audit log design. For teams running AI in production, authentication and authorization design tends to get postponed, but it is an area that cannot be avoided if you want to prevent security incidents.
---