BrainBoost

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

~ the obit ~

Assaf Rozenblatt registered brainboost.com in June 2004 and built, by himself, a search engine that tried to answer the question you had actually asked. You typed a sentence in plain English. It went out across other engines' results and pulled candidate answers out of the text it found. The one it was most confident in got printed as a short statement rather than as a list of pages that happened to contain your words.

It was good enough to be measured. One 2005 comparison had it producing an answer somewhere on the first page about 87 percent of the time, against barely 50 percent for the major search engines, and putting the answer in the first result 75 percent of the time against their 23 percent. In March 2005 its language engine was picked up for a NASA project. The reviews from ordinary users were cooler, and more than one person noted that a confident sentence pulled out of a web page inherits whatever that page got wrong.

Answers Corporation bought it on a purchase agreement dated 1 December 2005, for $4 million in cash plus 439,000 shares of restricted stock. Rozenblatt went with the deal and became the company's Director of Natural Language Research. His site stopped being a destination soon after: the last archived copy of brainboost.com serving its own page is from 2 February 2007, and by 27 September 2007 the address had become a redirect into Answers.com.

Then the owner changed its mind about the whole idea. In the second quarter of 2008 Answers.com decided to put its weight behind questions and answers written by its own users, and abandoned the BrainBoost Answer Engine outright. The engine was carried on the books at $3.138 million as of 25 May 2008, and that entire figure was written off in the quarter ending 30 June.

~ the rap sheet ~

~ how it answered ~

Extraction was the design and also the ceiling. The engine had no knowledge of its own, so every answer it gave was a fragment it had found and decided to trust, which made it excellent on questions with one settled fact behind them and useless on anything contested. It also meant the whole apparatus ran on the English-language web. A question typed in Turkish in 2005 had almost nothing underneath it to extract from, which is why this entire generation of answer engines stayed a story I read about in English rather than a tool anyone around me opened.

~ the humans won ~

The thing that replaced it inside the same company was WikiAnswers, where people wrote the answers themselves for free and other people corrected them. Answers.com was running both at once and watching them compete for the same job. The engine did not lose to a smarter engine. It lost to volunteers with keyboards, who were cheaper than $3.138 million and better at knowing when a question was a real question.

Rozenblatt went on to build SuperCook. The domain went its own way and now serves a large, unrelated site under the same name, several hundred kilobytes of it, with nothing to do with answer extraction. The clearest surviving trace of the product is in an SEC filing, where it appears as an asset being removed from a balance sheet.

@}-->--

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

« previous  ·  random  ·  next »