Differences in link hallucination and source comprehension across different LLM

As most of you all know, I built SIFT Toolbox as a contextualization engine, initially on ChatGPT but started developing on top of Claude after seeing better results there when they rolled out their paid version with search. Also many of you also know, I started to get the system to do pretty striking things, providing AI-generated professional-level “context reports” on claims, artifacts, and quotes. An AI-generated (SIFT Toolbox) fact-check which outperforms a Snopes article on the question of whether the snow in Wizard of Oz was “made of asbestos”, surfacing evidence that Snopes did not, and correctly evaluating its weight as direct witness testimony accepted by film historians. Initially I thought as the different models evolved they’d all match the performance I was seeing on Claude. But we’re a couple releases down the road and that hasn’t happened. There is, in my experience, a very real gap between some of the models and others that no one is talking about, and no one seems t ... Read full article.