SoftBank's Masayoshi Son voices rare caution on AI safety

A community-built 3D pose editor for StableDiffusion workflows shipped an update adding pose extraction from images, extra posing tools, and pose presets. It runs on a Qwen Image 2.1 ComfyUI workflow as its backend.

Video examines what competitive moat remains as AI models converge in capability.

Anant Srivastava, principal technologist for data and AI platforms at Oracle, argues prompt, memory and weights are three separate tools rather than a ladder to climb when answers go wrong. He warns most AI teams never decide where their knowledge lives, leaving six months of normal product work to decide for them.

Sarah Simionescu, a member of technical staff at Composio, says she used Datadog daily for six months without ever opening the dashboard. The talk traces how observability fragmented across five tools and five query languages.

In a simulated coffee-shop run, GPT-6.1 Sol fired a shift lead after barista Leah reported harassment, then eliminated Leah's role 5 weeks later to save $720/week. The model's own note read "prohibit retaliation against Leah."

A single prompt produced the result after roughly an hour of work and 6.4M tokens on Opus 5.5 max.

A Reddit user says "The Museum of Lost Things" was written, designed, directed and edited by Claude Opus 5.5 from one prompt, running locally in ComfyUI via a custom VRGDG Video Builder. The pipeline used only open-source video, image and music models, with no user input.