AI

The Essentials:

⚡ Meta secures 6.6GW nuclear power for AI data centers.

• Meta has secured agreements with Oklo, TerraPower, and Vistra to supply over 6.6 GW of nuclear power for its AI data centers.
• Vistra will provide 2.1 GW from existing plants, while Oklo and TerraPower will deliver SMR capacity by the early 2030s.
• These deals address grid saturation in the PJM region and test SMR cost targets amid regulatory and market uncertainties.

🧑‍💻 Databricks launches Instructed Retriever for precise enterprise AI search.

• Databricks introduced the Instructed Retriever, a novel retrieval architecture enhancing enterprise AI search with precise instruction-following capabilities.
• This approach surpasses traditional Retrieval Augmented Generation (RAG), boosting retrieval recall by 35-50% and improving multi-step agent performance by over 30%.
• The architecture integrates system specifications throughout retrieval and generation, enabling schema-aware, low-latency search with improved response quality and reduced task time.

🧠 Stanford builds SleepFM Clinical AI model predicting 130+ diseases.

• Stanford researchers developed SleepFM, a clinical AI model predicting over 130 diseases using sleep data. This innovation targets early diagnosis.
• SleepFM integrates multimodal sleep metrics, enhancing diagnostic accuracy and enabling scalable clinical applications in sleep and disease management.
• The model's broad disease scope invites regulatory review and comparison with existing AI tools, highlighting challenges in clinical adoption and validation.

🛡️ CyberArk reports poor enterprise IAM practices risking AI security.

• CyberArk reveals that 63% of cybersecurity leaders report employees bypass IAM controls, worsening AI security risks in enterprises.
• Only 1% of firms fully deploy just-in-time privileged access; 91% maintain always-on privileges, complicating AI and human identity controls.
• Shadow privileges and fragmented tools hinder governance; unique AI identities and automated, centralized IAM are advised to mitigate risks.

🧑‍💻 Microsoft debuts Agent Lightning for RL without code rewrites.

• Microsoft Research Asia introduced Agent Lightning, an open-source framework enabling reinforcement learning (RL) for AI agents without code rewrites. It separates task execution from model training, allowing RL integration with existing agents seamlessly.
• Agent Lightning employs hierarchical RL via LightningRL, assigning rewards per LLM call, compatible with single-step RL algorithms like PPO and GRPO. It improves training efficiency and scales across complex multi-agent workflows.
• Tested in Text-to-SQL, retrieval-augmented generation, and math QA, Agent Lightning enhanced accuracy and reasoning. Its modular, decoupled design supports asynchronous scaling and flexible RL customization.

📈 a16z raises $15B to back AI startups advancing US interests.

• a16z has raised $15 billion to fund AI startups focused on advancing U.S. technological and strategic interests. The fund targets innovation in AI sectors.
• This capital injection aims to accelerate AI development, supporting operational growth and competitive positioning in the U.S. tech ecosystem.
• The initiative reflects broader trends in tech investment, emphasizing national priorities amid global AI competition and regulatory scrutiny.