thinking hurts design
check out https://blog.kekepower.com/ai/ zero-shot webpage generation to compare frontend capabilities
note: models generate better, more cohesive design without thinking
this holds true for models across families and sizes (small sample, as of 5/7/25) compare thinking vs non-thinking for Claude 3.7 sonnet, Qwen 2.5 max, and GLM 0414 (4 vs z1) thinking seems to introduce weird quirks or glitches
observations: qwq-32b is somewhat of an exception, but there is no non-thinking version for comparison. deepsite punches above its size. gemini theme seems to be 'aged corporate'
guess: pre-training contains lots of webpages, most of which have gone through several rounds of refinement. models may have developed an instinct for good design that works out better than reasoning over the details.
this feels obvious to some extent. good design is intuitive. overanalyzing minor details leads to subpar results. don't get "lost in the sauce" (Gucci Mane 2013).
art requires a level of 'zooming out'. reasoning (during creation) might as well be a dirty word for many artists.
hobbling alert: vercel is like a bunch of parts of stitched together. which makes sense because they basically ask the model to use pre-built/existing components. it looks nice at first glance due to high quality of underlying components, but the design cohesion suffers. also, glitches.