Mindpixel

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

~ the last words ~

Mindpixel asked a question I now get paid to have opinions about, and it asked it in the last years when the answer was still open. That is the honest distance here. In 2005 I was a computer engineering student in Ankara and machine learning was a course you sat through, not something you did. Twenty years later I train vision models for a living, which lets me read what Mindpixel was betting on and say plainly that the bet lost.

Chris McKinstry founded it in July 2000, but the idea was older. In 1996 he had published something he called the Minimum Intelligent Signal Test, arguing that you could replace the open conversation of a Turing test with a very large number of binary questions. If a machine answered enough true-or-false statements the way a human population answers them, the difference would stop mattering. Mindpixel was the attempt to collect the population's answers first.

The mechanism was strict. Before you could submit a statement of your own you had to validate twenty existing ones. Each new item went out to the crowd to be voted on, and the majority answer became the stored value. Participants whose answers were consistently out of step with everyone else were downgraded and eventually excluded. By January 2004 the database held 1.4 million of these items from thousands of contributors, and the software holding them was called GAC, for Generic Artificial Consciousness, pronounced jak.

What the project was really claiming is that common sense is a finite list and somebody has to sit down and write it out. Everything that came afterward went the other direction. The systems that ended up working did not get their picture of the world from a curated table of validated truths. They got it from an enormous pile of unlabeled text that nobody voted on. Mindpixel was five years of careful manual work aimed at a problem that turned out to yield to volume.

~ the record ~

~ the shares nobody could sell ~

Contributors were not paid. They were promised shares in whatever the database turned out to be worth, in proportion to how many accepted mindpixels they had written. McKinstry believed a validated corpus of human common sense would have real commercial value once a neural network was pointed at it, and that belief was the entire compensation structure.

There was never a buyer or a valuation that anyone has been able to point to since. The volunteers carried all of the risk and the project carried none of the cost, which is a fine arrangement while the work is going well and an unanswerable one the moment it stops.

~ the day it stopped answering ~

On 20 September 2005 Mindpixel lost the free server it had been running on and went offline. There was no announcement and no wind-down, because there was nothing to wind down. The whole operation had been living on hardware somebody else was paying for.

McKinstry described the project as closed by that December and said a second version would appear on a new machine in France. It never did. mindpixel.com today loads a domain parking page.

~ what happened after ~

McKinstry was not a full-time researcher. From 1999 he worked in Chile as a Very Large Telescope operator for the European Southern Observatory, running somebody else's instruments through the night in the Atacama and building his corpus around that. He was found dead in Santiago on 23 January 2006, four months after the server went, having posted a note online first.

A paper written with two cognitive scientists, using Mindpixel data to study the timing of thought, appeared in Psychological Science in January 2008 after his death. That is the last clearly documented use of the corpus. Wikipedia's own line on it is that the integrity of the data is uncertain, which is a careful way of saying nobody knows what survived or who has it.

@}-->--

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

« previous  ·  random  ·  next »