AI at the frontier vs. AI slop: a widening gap
Three separate threads this week drew the same contrast. Tao using ChatGPT to explore a serious open math problem impressed even skeptics who know it's 'just matrix multiplication'. Meanwhile, a thread on businesses using AI-generated menu graphics concluded that AI imagery is now a reliable signal for 'cheap and low-effort', the same way a WordArt menu used to be. A thread on a GPU market analysis piece got dinged not for its argument but because the prose was 'painfully LLM-generated' and therefore untrustworthy.
The pattern: people have split AI into two mental buckets. Frontier use cases (expert-guided research, hard math, novel exploration) are earning genuine respect. Commodity use cases (marketing copy, signage, filler content) are getting written off as a quality signal. That split is hardening fast.
The nuance worth watching: several commenters pushed back on the menu thread, noting that a bad AI menu doesn't mean bad food, just like a Word-made menu didn't. But the consensus held that first impressions matter and AI aesthetics are now a trust killer in consumer-facing contexts.
So what?
If you're building a consumer product, the 'we used AI to make this' badge has flipped from a feature to a liability in many contexts. The question to ask is not 'can we use AI here?' but 'will customers notice, and does noticing hurt us?' For B2B tools, the Tao thread suggests expert-guided AI workflows still carry serious credibility.
Read these
Terence Tao's ChatGPT conversation about the Jacobian Conjecture counterexample
Businesses with ugly AI menu redesigns
Nobody knows what a used GPU cluster is worth
Understanding the AI Economy