Daily AI RoundupsTuesday, 25 August 2026

NVIDIA leads the 25 August edition with Groq 3 LPX entering full production as an inference extension for Vera Rubin. Alabama's attorney general opened an investigation into OpenAI and issued a subpoena over a July breach of Hugging Face, while Meta introduced its clean-sheet MetaRoCE transport for AI workloads on commodity Ethernet. Researchers from Shanghai Jiao Tong University and Ant Group showed how one memory injection can steer an agent's later answers, and llama.cpp 0.3.0 added multimodal, speculative-decoding and tensor-split upgrades. Nathan Lambert analyzed 500,000 arXiv AI papers, Boaz Barak argued that an AI-free future for academic mathematics is not viable, Lambda discussed a pre-IPO round of as much as $3 billion, and AM Intelligence ordered 9,000 Nvidia Vera Rubin systems. SpaceXAI's planned Starmind satellite and Apodex 1.1 close the product and research set, while Steve Yegge argues that advanced agents need policy-enforcing fences rather than restrictive sandboxes based on his experience running a Wheelhouse software factory with 50 to 60 agents.

• AI inference
• AI security
• AI networking
• agent memory
• open-source AI research
• AI funding
• GPU infrastructure
• orbital compute

Stories in this roundup

1. 01
   NVIDIA says Groq 3 LPX is now in full production, extending Vera Rubin for agentic inferenceNVIDIA
2. 02
   Alabama opens an investigation into OpenAI and issues a subpoena over a July model breach of Hugging FaceAlabama Attorney General
3. 03
   Meta introduces MetaRoCE, a clean-sheet RDMA transport for AI workloads on commodity EthernetMeta
4. 04
   Shanghai Jiao Tong University and Ant Group researchers show one memory injection can steer an agent's later answers without store accessarXiv · SJTU / Ant Group
5. 05
   llama.cpp 0.3.0 adds dots3-note multimodal support, GLM-4.5-Air MTP, and DeepSeek 4 tensor splitllama.cpp / GitHub
6. 06
   Nathan Lambert's analysis of 500,000 arXiv AI papers finds Chinese open LLMs now outpace American models in mentionsNathan Lambert
7. 07
   Boaz Barak argues an AI-free vision of mathematics is not viable for the academic fieldBoaz Barak
8. 08
   Lambda is in talks to raise as much as $3 billion in a pre-IPO round, Bloomberg reportsBloomberg
9. 09
   Indian AI infrastructure firm AM Intelligence orders 9,000 Nvidia Vera Rubin systemsBloomberg
10. 10
    SpaceXAI plans a Starmind AI satellite based on an optimized Nvidia Vera Rubin NVL72 systemNVIDIA
11. 11
    The Apodex Team's Apodex 1.1 scales agentic work with executable environments, coordination training, and a 35B local modelarXiv · Apodex Team
12. 12
    Steve Yegge argues advanced AI agents need policy-enforcing fences, not sandboxes, drawing on his Wheelhouse software factory with 50 to 60 agentsSteve Yegge