8am☕Coffee

MIT updates its SEAL technique to advance self-improving language models, enabling models to refine performance through iterative learning with reduced external supervision. The method focuses on letting large language models learn from their own outputs and feedback to improve reliability and capabilities.

Researchers find that retraining only small parts of AI models can reduce fine-tuning costs while preventing or lessening forgetting. The approach concentrates updates on selected parameters instead of the full network to preserve prior knowledge and improve efficiency.

Salesforce bets on AI agents to address what it calls a $7 billion problem in enterprise software, focusing on automating complex workflows across business functions. The strategy centers on agentic systems that coordinate tasks, reduce manual steps, and streamline outcomes for enterprise users.

Meta and Oracle choose NVIDIA Spectrum‑X for AI data centers to enhance networking performance and efficiency. The move indicates alignment around specialized networking designed to scale AI infrastructure across demanding workloads.

NVIDIA delivers the just-launched DGX Spark petaflop AI supercomputer to SpaceX. Elon Musk receives the system to accelerate AI workloads at the company with petaflop-class compute.