Yupp AI

» see yupp.ai as it lived, on the wayback machine

~ the eulogy ~

i did not use yupp. i heard the name a couple of times in the 2025 ai-industry chatter, filed it in the "another model comparison site" mental folder, moved on. by then i was 42 in ankara doing edge ai on drones and my model-picker habit was pinning one or two providers for real work rather than swirling through 800 on a leaderboard. yupp was in the part of the ai stack that was interesting to industry watchers more than to anyone actually shipping.

the site launched in late 2024. pankaj gupta and gilad mishne, both ex-linkedin, built a consumer-facing comparison tool where you typed a prompt and got side-by-side outputs from a pool that eventually reached about 800 models. every major closed lab, plus meta, mistral, chinese labs, and a long tail of open-weight releases. you voted on which output you liked. the votes were meant to become an anonymized preference dataset that model makers would then pay to see. rlhf for the rest of us, kind of, but sold as a service.

chris dixon at a16z crypto led a $33 million seed round in 2024. more than 45 angels signed the same doc, including jeff dean from google deepmind and biz stone from twitter. aravind srinivas of perplexity and evan sharp of pinterest were on the list too. the cap table looked like a class photo of the ai and consumer-web establishment. the exact roster, but it was that kind of round.

~ the dossier ~

~ the unraveling ~

the pitch had a subtle timing problem baked into it. yupp was betting that pairwise human preference data would remain a scarce and valuable asset. the labs were betting the opposite: that synthetic feedback loops and internal eval harnesses would eat most of the demand for external preference signal within a year or two. by early 2026 the labs looked like they had won that side of the argument.

the other cracks were the usual consumer-web ones. free comparison tools have thin loyalty. the users came for the leaderboard novelty and left when the next benchmark drop reordered the rankings. paying customers on the model-maker side did not materialize at the scale the round required. through the industry press coverage of the closure.

~ the day it died ~

on 31 march 2026 the co-founders posted the shutdown notice. the statement was honest in a way not every startup closure is. they said the product had not reached strong enough market fit and that ai models were improving in leaps so large that the preference-data thesis kept slipping out from under them. no acquihire. no soft-landing at one of the labs whose founders sat on the cap table. clean wind-down.

~ what it left ~

yupp becomes a footnote in the second wave of ai consumer plays, the wave that assumed the model layer would fragment into a marketplace where a neutral picker had leverage. the actual 2025-2026 shape was closer to two or three labs pulling ahead and everyone else consolidating or exiting. a16z crypto took the loss. the 45 angels took smaller ones. the anonymized preference vote history probably still sits on a backup drive somewhere, worth whatever preference data is worth in the agentic era, which is less every month.

@}-->--

              )
             (_)
            .-'-.
            |   |
            |   |
            |   |
            |   |
          __|   |__   .-.
       .-'  |   |  `-:   :
      :     `---'     :-'
       `-._       _.-'
           '""""""

« previous  ·  random  ·  next »