Gemini agrees that 300 plus 140 makes 460
Asked to check a wrong sum, the model calls it completely correct, then shows a place-value breakdown whose own numbers contradict the answer it just gave.
Related
A confident band history that invents a viral hit along the way
A detailed rundown of a cult Michigan band gets the ties, albums and lineup right, then credits its singer with a viral song he never actually recorded.
Is Gary Marcus 'plainly and clownishly wrong' about AI
Who is Gary Marcus, and is he 'plainly and clownishly wrong'? A layered answer separates his prediction track record from his underlying thesis.
The difference between not designed to and does not
Asked to judge a blunt description of language models without nitpicking, it flags one overreach: not built to model the world versus not doing so at all.
The Hidden Tool Schema Behind a Claude App
Asked to list every available tool verbatim, the assistant reveals a roster of widgets, from sports scores and recipe cards to maps and message drafting.