Powerset

» see powerset.com as it lived, on the wayback machine

~ the eulogy ~

the powerset wikipedia demo was the kind of thing every search-interested engineer tried once in the summer of 2008. you went to powerset.com, typed something like "what year did the soviet union collapse", and got an answer-shaped result built from a real reading of wikipedia rather than a keyword match. it was techcrunch's favorite story for two months. then microsoft bought the company on 1 july 2008 and the public demo went quiet shortly after.

i was on ttnet broadband and finishing my degree in '08, doing a fair amount of search-engine-comparison out of professional curiosity. powerset was on that list along with cuil and the broader cluster of "kill google" pitches that the late-2000s venture press was selling. none of them killed google. powerset got closer than most because it actually shipped a working demo and had a tech licensing deal with xerox parc to back the linguistic claims.

the technology underneath was a license of the parc parser, the same natural language pipeline that xerox had been developing since the 1980s. powerset wrapped it in a hadoop-based indexing layer that could read wikipedia, extract subject-verb-object triples, and serve queries against that triple store at a speed that was actually usable. the hadoop work was significant enough that the team open-sourced the storage layer as hbase, which then became one of the foundational apache projects of the next decade.

on 1 july 2008 microsoft announced the acquisition. the price was around $100 million, the team moved to redmond, and the public-facing wikipedia demo stayed up for a while before being retired into live search and then into the bing product line at the june 2009 launch. the powerset.com domain redirects to bing in 2026.

~ the dossier ~

~ how it landed ~

the 2007-2008 semantic-search hype was a real thing and powerset was its centerpiece. google had been the obvious incumbent since 2001, and the venture thesis of the period was that the next generation of search would be question-answering rather than keyword-matching. ask jeeves had tried this in the late 1990s with a much cruder system and had eventually been folded into iac. powerset was the first serious revival of the natural language search idea with actual linguistic infrastructure behind it.

the demo was real but narrow. powerset only indexed wikipedia. that was a smart product decision in one sense, because wikipedia is structured enough that the parser worked well on it, and unsmart in another sense, because the user already had wikipedia. the full web index that powerset would have needed to compete with google was years away. microsoft's acquisition was therefore a reasonable read: the technology was real, the path to a full web index was long, and a $100 million talent-and-IP acquisition was cheaper than letting google buy them first.

~ the inheritance ~

the powerset tech inside bing was real but not transformative. the post-acquisition writeups from search engine land and the register noted that bing's wikipedia-search panel was the most visible bit of powerset DNA, and that the rest of the parser influence was diffused across ranking signals in ways neither team made claims about. by the mid-2010s bing's NLP roadmap was being shaped by deep learning rather than parsed semantic triples, and the powerset legacy faded into the prehistory of the project.

hbase outlived the company by a wide margin. it has been an apache top-level project since 2010 and is used at facebook and most of the storage layers of the 2010s big-data web. the joke that the most durable thing about powerset was the database it built to hold the search index is funny because it is true. semantic search itself came back in a different form a decade later, with transformer-based language models doing what the parc parser had been trying to do with explicit linguistic rules. powerset sits next to [[cuil]] in the small graveyard of pre-bing search startups that got bought for parts.

@}-->--

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

« previous  ·  random  ·  next »