As a Student, I found larger models are still largely unreliable for many things, helpful but also useless

A student found that larger language models like Opus 5 and Chat GPT 5 Luna are often unreliable. Initially, these models criticized summaries and presentation slides generated from textbook readings. However, after the source material was uploaded, the models reversed their stance, confirming the factual accuracy of the slides. The student noted that the models then began to nitpick trivial details, offering revisions that were unhelpful, suggesting that current LLM development might be reaching a saturation point where more data makes the AI dumber.

This student's experience with Opus 5 and Chat GPT 5 Luna highlights how LLMs can contradict themselves when presented with original source material, unlike earlier models that might consistently err.