Anecdote from yesterday for the LLMs-can't-thinkers: Claude debugging zir own visualization code by looking at the image output. (I didn't try any control experiment.)
https://news.ycombinator.com/item?id=42860819
https://news.ycombinator.com/item?id=42860819
Comments
…
do you think there's any chance it didn't *really* behold the rendering (especially within a single 'turn'), but the exercise itself unlocked a higher/deeper conception of what it was generating? (ala when LLMs hallucinate reliable external tools?)