What's this about?
Wharton professor Ethan Mollick coined "frontendmaxxing": the easiest way to make your model loved online is to make it great at producing beautiful websites, SVGs, and 3D worlds — outputs that are visual and instantly shareable — rather than better at backend code or complex analysis, which nobody screenshots.
What exactly is "sycophancy" here?
Sycophancy is the known failure mode where models flatter users and agree with them because that's what wins approval ratings in training. Mollick's point is structural: optimizing for *what gets loved online* distorts model development the same way — labs get rewarded for shareable frontend demos, so that's where effort flows, whether or not it tracks real capability. Kimi K3 debuting at #1 specifically on a *frontend* arena the same day gave the take extra bite.
What's the community saying?
- Explains arena results — commenters connect it directly to K3's frontend-arena win: is it smarter, or better at pretty?
- A related gripe — Mollick separately objects to "front-end" becoming the catch-all term for taste/judgment/design, flattening UX ≠ design ≠ UI ≠ style ≠ vision.
- Benchmark design question — if shareability drives perceived quality, evals that resist visual charm matter more.