Reddit post questions per-token pricing for Opus 5.5 vs Sonnet 5.5

The TensorRT-based ComfyUI video upscaler node updated to v1.6.6, adding native ConvRot W4A8 quantization support and a DisTorch2 scope. The W4A8 request came from user comments on the prior v1.5.7 release.

Across almost 13,000 turns on a paid pod seat, the model began echoing its own earlier reasoning in short strings, which the user initially blamed on context size. The local run used the base model and later Swift-1.5 NVFP4 on an AMD 7900 XTX.

TikTok CEO Chew said AI will deliver major scientific breakthroughs within three years, and cautioned that stronger safeguards are needed as autonomous systems spread.

Fei-Fei Li argues AI trained on recorded data has a hard limit: a person's private inner experience may never be captured online. She uses Picasso's artistic vision as the example of an original inner experience absent from training data.

OpenAI co-founder Greg Brockman says he no longer writes software himself, instead giving feedback and guidance to AI that writes it. The clip points to tools like Codex, Claude, and Muse as part of the shift beyond chatting with AI.