[AINews] Everything is CLI

AI News: Everything Is CLI

Model Releases and Updates

• Cohere Transcribe: Cohere has released a new audio transcription model, which is open-source and comes with permissive licensing. This release continues the trend of making production-oriented audio models more accessible.

• Mistral Voxtral: Mistral AI also announced its own audio transcription model, further expanding the options for developers looking to integrate speech-to-text capabilities into their applications.

Agent Training and RL

• NVIDIA's ProRL Agent: A significant systems thread highlighted that agentic Reinforcement Learning (RL) frameworks have been incorrectly architected by coupling rollout and optimization in the same process. NVIDIA’s ProRL Agent claims that fully decoupling the rollout into a standalone service nearly doubled Qwen 8B on SWE-Bench Verified from 9.6% to 18.0%, with similar gains for other variants, alongside much higher GPU utilization.

• Cursor's Real-Time RL: Cursor announced it can ship improved Composer 2 checkpoints every five hours, presenting this as a productized RL feedback loop rather than a static model-release cadence. This is seen as an early sign of continual learning in production,

Read the full article at Latent Space

---

Want to create content about this topic? Use Nemati AI tools to generate articles, social posts, and more.