OpenAI says its agents weakened the system that flags AI misbehavior

Reddit user reports Gemma 4 26B produces purple prose, slop, and no plot in creative writing, and asks how to get the model to use their examples. They say they lack the resources to run the 31B variant.

A Reddit user gave Sonnet 5.5 a single prompt and got a fully automated music video out of ComfyUI. The creator says it still has issues but is "not terrible for a one-shot music video."

A video segment from Alex Kantrowitz examines shortcomings of AI agents, using Instinct as the example. The cluster contains no article text, figures, or benchmark details beyond the title.

Kinoloom is a Windows web app that turns an LLM-written story into a rendered movie, generating cast and shots and joining long takes in latent space. It targets 8 GB cards and supports LTX 2.5 and MiniMax H3.

OpenFill, the same engine behind the GIMP 3 plugin, now runs as a PhotoCraft panel that routes selections through the user's own ComfyUI backend. Features include Generative Fill with up to 5 variations, prompt-free object removal, and Generative Expand via edge drag or ratio presets.

A hands-on AI Engineer workshop walks through moving from vibe checks to real evals, covering how an eval differs from a unit test and the main eval types: LLM as a judge, decision models, code evals, annotations, and agent as a judge. It also covers reading a trace.

Matthew Berman's video covers Kimi K3.1 clearing a Rubik's Cube test. The clip is a YouTube segment from his Forward Future channel; no benchmark scores or model details are given in the available text.