The trust gap with AI agents and LLM judgment

Claude Duet drew a cautious reaction: commenters say these assistants pull insights from every connected third-party integration, and they struggle to make the trust leap. One asks what people actually use personal agents for. A separate post argues that using an LLM for confidence scoring is a bad idea. The top comments say LLMs have the sensibilities of the average person, which makes them poor at rating anything subjective.

In an older Ask HN, people compare local models to cloud ones for daily coding. Qwen 3.6 27B is called roughly Haiku 4.5 level, while Gemma 4 on an Apple M4 is too slow, and others run DeepSeek V4 Flash on two RTX Pro 6000 cards. The through-line: people use these tools heavily for generation and still don't trust them as judges or as unsupervised actors.

A related thread asks HN to add a flag for AI-generated articles. The new guideline already says not to post generated or AI-edited text. The pushback is that detection is unreliable and the stigma is growing.