Blogotariat

Oz Blog News Commentary

Revisiting Open Source Ventures

July 14, 2017 - 01:31 -- Admin

Six years ago, I floated a business plan, for building a new type of company to building extremely valuable databases.  Apparently, it was far more insightful than anticipated.  What changed?  AIs (more specifically, deep learning/ConvNet).  We now have a method for creating artificial intelligence (more accurately, artificial understanding) that works.  AIs that solve real world problems.  In fact, these AIs are getting so good, so fast, that all of the big tech companies are now AI companies -- from Google to Facebook to Uber to Tesla.  The reason I bring this up?  The way AIs are built creates an opportunity for open source ventures.  Currently, AI's are built using three things:

  • Scalable models with high capacity (we have that now deep learning/ConvNet now).
  • Lots of processing power (to train AIs).  Moore's law cracked the initial barrier to using these methods. Cloud based training systems filled with GPUs is making this less expensive.
  • Big data sets (specifically, human curated data) that are used by the AI's to learn.

Given this, it appears that an open source venture (a company that can scale to millions of worker/owners creating a new economic ecosystem) that builds massive human curated databases and decentralizes the processing load of training these AIs could become extremely competitive.  Here's why:

  • A recent study by Google showed that the size of the training database correlated to the quality of the AI.  The bigger the better.  That's a bit of a problem for the big tech companies.  Currently, most of the of the databases currently available were built by very low cost workers for a fraction of a penny a label, using places like Amazon's Mechanical Turk.  That makes them clunky to build and difficult to scale.  Further, the labelled, human curated data they do have (on Facebook, Google, etc.), while voluminous, isn't complete enough to do high quality training.  Any company that could build these massive databases would be set to train the best AI's in the world.
  • The processing power needed to train these AIs, as they become more sophisticated, could outstrip current corporate capabilities (of even Google).  This suggests that training systems that break apart the computation tasks among tens of millions of participants (cell phone/desktop) could surmount this barrier.  In practice, it would be a combination of folding@home and bitcoin mining.  Already, deep learning AIs can "run" on cellphones.  Configuring them to contribute to the training load and/or actively gather data for database construction isn't that much of a stretch.
  • The people participating (earning equity instead of fee for service for their contributions) could become very, very large.  They would be familiar with the platform and become increasingly sophisticated at finding amazing ways to capture new data and new AIs to build.  Over time, a platform like this could be the source of many (if not most) of the best AIs ever built.  A source of immense wealth (seen as dividends) for hundreds of millions of owners, earning equity with each contribution.  

Worth thinking about.

John Robb

PS:  Given our experience with bitcoin, this isn't impossible to do.  It's also a much better future than a world that one built on Turking.  

PPS:  Lots of readers don't like the tech stuff.  I think you are missing the point.  It's coming and it can't be avoided.  It's better to profit from it than be a victim of it.