Reddit user benchmarks 100+ h3 turbo configs

Blind test of everyday edits: FIBO won 15 picks vs Qwen's 12 and FLUX's 11, though Qwen took 3 of 5 rounds and FIBO Dev tied FLUX on new images. Workflow posted free on Civitai.

A Reddit user fed a raw 2-minute clip to Claude Opus 5.5 on Max effort and asked for a short-form video; they report it handled cuts, motion graphics, and effects well. The post is an anecdotal single test, not a benchmark.

A community builder's MiniMax H3 Dual Segment workflow has been rebuilt as Chain Director V1.0, replacing the fixed two-segment setup with a custom Director node that controls the full MiniMax H3 generation process from one panel.

Sebastian Raschka published eight coding videos (11 hours) implementing reasoning models from scratch in PyTorch, starting from a pre-trained LLM. The course covers inference scaling, reinforcement learning with verifiable rewards, and distillation, and follows his earlier Coding LLMs from the Ground Up course.

About 300 employees signed a petition asking leadership to cancel a $1.3M contract supplying four AI-powered Spot robot dogs to ICE via Connecticut firm Radeco. The modified unit, nicknamed Cerberus, adds a robotic arm, thermal camera and explosive-material sensors.

Altman said he lacks "a dramatic enough word" for models finding bugs "outside of the normal threat model of what we think a hacker can do," citing "hardware-level, insane stuff." He added he no longer thinks "we're gonna be able to just patch all the bugs."