Large Language Models

Today's Stories

1. 1

   Storage tech gets new role in AI agent era as context memory demand soars

   Storage has taken on a new dedicated function in AI clusters—storing context memory and KV cache for agentic AI systems. Nvidia's BlueField-4 STX architecture, announced in March, introduced Context Memory Storage (CMX) to expand GPU memory across the rack and enable large language models to handle million-token context windows. As AI workloads shift from single prompts to extended agentic sessions, the volume of data to store is growing into petabytes—far exceeding what standard GPU and DRAM can handle. Storage is now a strategic differentiator rather than a supporting component. For enterprises, this means the ability to feed models with high-quality data and sustain throughput across distributed environments has become equally important to GPU performance alone.

   A Deloitte forecast notes that inference will account for approximately two-thirds of AI compute in 2026. The emerging context graph tool—an accumulated structure of decision traces woven among entities and time—is being positioned as critical infrastructure for agentic AI to access the right context across fragmented data sources.

2. 2

   Visa (V) Is Testing AI Agents That Can Actually Complete Purchases

   Visa (V) Is Testing AI Agents That Can Actually Complete Purchases

3. 3

   Even if Fujikura Plummets, Is It Okay? The Truth About 5x Explosive Growth Stock, Sumitomo Electric, and Nidec | The "Shadow Favorite" of AI Hegemon OpenAI and Gemini? [5803] Carol Vorderman (etDOCplIVa)

   Even if Fujikura Plummets, Is It Okay? The Truth About 5x Explosive Growth Stock, Sumitomo Electric, and Nidec | The "Shadow Favorite" of AI Hegemon OpenAI and Gemini? [5803] Carol Vorderman (etDOCplIVa)

4. 4

   SoftBank builds custom AI gateway to manage 100 agents per person

   SoftBank has developed Cloud Proxy, an internal AI gateway system designed to support its vision of "100 agents per person." The gateway, which began development in 2023, manages security, identity verification, and token usage across multiple AI models including Azure OpenAI Service, Gemini, and Sarashina. The system solves a core security and operational challenge in deploying many autonomous AI agents (software that makes decisions and takes actions) simultaneously. By handling API key management, preventing unauthorized access, and enabling fine-grained monitoring of token usage—the basic units measured for AI processing—Cloud Proxy allows SoftBank to scale agent deployment while maintaining security and controlling costs.

   SoftBank reports that its scaling tests showed the system can support "100 agents per person" architecture in 14 weeks, with request volumes doubling while maintaining infrastructure stability. The company is adding support for new AI models through a weekly update process and is aiming to demonstrate that integration with different identity providers can reduce enterprise AI adoption friction by 70 percentage points.

5. 5

   Anthropic shifts Claude pitch to eliminating busywork, expands to mobile

   Anthropic is expanding its Claude Cowork agentic AI platform (software that can independently perform tasks) to mobile devices and refocusing its messaging around eliminating routine work rather than replacing workers. The shift in narrative reflects how AI companies are addressing concerns about job displacement by positioning their tools as assistants that handle tedious tasks—a framing that may resonate more strongly with businesses and regulators as the company prepares for potential public market activity.

   The expansion to mobile makes the agentic AI capability accessible beyond desktop environments, potentially widening the user base for task automation.

6. 6

   Researcher proposes constraining AI model fine-tuning to trusted adapters

   A researcher published a paper exploring a defense against fine-tuning poisoning by constraining model adaptation to a subspace learned from trusted LoRA adapters (a technique for efficiently modifying AI models). The approach prevents the model from learning certain malicious updates while preserving useful adaptation. Companies and on-device assistants that fine-tune models on user data or external sources face the risk of hidden behaviors introduced by poisoned data. This method offers a way to restrict what the model can learn without needing to detect every possible attack, potentially reducing the vulnerability surface when deploying continuously-learning systems.

   The paper frames this as one approach to a specific problem—constraining the geometry of fine-tuning rather than detecting poison itself—so organizations relying on model adaptation may want to evaluate whether this or similar geometric constraints fit their deployment model.