LocalLLaMA user finds slower inference beats raw tok/sec for coding

CUDA 13.5 moves NVIDIA Vera Rubin support from early developer access to general availability for enterprise production, and extends Windows on Arm support to DGX Station for Windows. The release adds locality-domain SM partitioning controls and MPS v3 scheduling options.

TechCrunch Equity podcast interviews Hugging Face Chief Policy Officer Irene Solaiman on building on open models, including when cheaper open models beat the most powerful ones. She argues the U.S. may need deep-pocketed players like Nvidia to build an open frontier model to compete with China's.

User reports paying $94.46 upfront in April for an annual Suno Pro plan running to April 2027, then losing Pro access in October. They had contacted Suno on October 17 about a new 20-download limit on a plan they say was sold as unlimited.

CEO Carl Pei said Nothing's first "AI-native" product ships next month and has no new screen, adding it "couldn't have been an app." Pei dismissed the industry's hunt for an "iPhone of AI" as "a dead end."

Google Research's Agile Deliberation turns subjective classifier training into an iterative conversation, using an LLM-driven optimization loop to build classifiers for fuzzy categories like "healthy food" or "inspiring artwork." It targets the gap where labelers lack a stable definition before labeling begins.

Render dropped its previous 30-day continuous runtime limit on paid workspace plans, letting AI sandboxes run indefinitely. It also added tools for long-running agent workflows and shared storage; Vercel still caps individual sandboxes at 24 hours.

ChatGPT can now generate GIFs from a prompt or uploaded photo, with controls for animation speed and frame-by-frame editing. The Images 2.6 update also speeds up image generation; OpenAI says users create over 4 billion images weekly, up from 3 billion in September.