Reddit thread asks how much developers rely on Claude for coding

A ComfyUI user reports that an existing prompt enhancer for MiniMax H3 broke image-to-video continuity: the subject appeared in a completely different setting after roughly two seconds of a near-frozen start image.

HarnessDev tests whether LLMs can build and evolve their own agent harnesses, finding self-built harnesses transfer poorly across models. GPT-5 solves 35.2% of Terminal-Bench 2.1 tasks in Terminus 2 but 49.6% in Codex CLI with identical weights.

Inherent co-founder and Chief Scientist Edward Hughes tells Machine Learning Street Talk that creativity is not optimisation, and that AI's missing capability is choosing which questions are worth asking. He argues scientific judgement must be learned through practice rather than specified.

Cohere-hosted talk by Raymond Chua addresses continual learning in deep reinforcement learning, framing it as a major unsolved challenge for AI agents. The work centers on predictive representations and memory to let agents adapt in complex, dynamic environments.