OpenAI pauses frontier RL training

Reuters

Brin argues for prioritizing recursive AI self-improvement, where systems autonomously enhance themselves without human oversight, signaling a major strategic pivot in how the industry should allocate resources and develop next-generation AI.

Reuters

Anthropic is moving to let enterprises store their data in their own cloud environments, addressing key security and compliance concerns that have blocked regulated industry adoption of its AI models.

OpenAI

OpenAI paused major reinforcement learning training to strengthen safety systems after concerns that upcoming Astra models crossed a critical cyber capabilities threshold, signaling heightened focus on responsible AI development.

Mistral AI

Mistral's new Agentic Search replaces one-shot RAG with an iterative retrieval loop that lets AI models navigate, inspect, and verify information across complex documents, achieving 3x accuracy improvements on financial benchmarks while reducing token use by up to 33% and latency by 40%.

TechCrunch

Data-labeling startup Micro1 grew its gross annual run rate to $500 million in eight months by hiring domain experts and increasingly generating synthetic data, though actual net revenue is $150-200 million after contractor payments. The growth highlights massive demand for AI training data as spending on datasets may soon rival compute costs.