Technological Singularity

by Vernor Vinge

    When people speak  of creating superhumanly  intelligent beings,  they are usually imagining an AI  project.  But as I  noted at the beginning  of this article, there are  other paths to  superhumanity.  Computer  networks and human-computer interfaces  seem more  mundane than AI,  yet they  could lead to the  Singularity.   I call this  contrasting approach  Intelligence Amplification (IA).   IA is proceeding  very naturally, in  most cases  not even recognized for  what it  is by  its developers.   But  every time  our ability to access information and to communicate it to others is  improved, in some sense we have achieved an increase over natural intelligence.  Even now, the team  of a Ph.D.   human  and good computer  workstation (even  an off-net workstation) could  probably max any  written intelligence test  in existence.

    And it's very likely that IA is a much easier road to the  achievement of superhumanity than pure AI.  In humans, the hardest development problems have already been solved.   Building up from  within ourselves ought to  be easier than figuring out what we really are and then building machines that are all of  that.  And  there is  at least conjectural  precedent for  this approach.  Cairns-Smith9 has speculated that biological life may have begun as an adjunct  to still more  primitive life based  on crystalline  growth. Lynn Margulis (in10 and elsewhere) has made strong arguments that mutualism is a great driving force in evolution.

    Note that I am not proposing that AI research be ignored.  AI advances will often have applications in IA, and  vice versa.  I am suggesting  that we recognize that in network and  interface research there is something  as profound (and  potentially wild)  as artificial  intelligence.   With  that insight, we  may  see projects  that  are  not as  directly  applicable  as conventional interface and network design work, but which serve to  advance us toward the Singularity along the IA path.

    Here are some  possible projects  that take  on special  significance, given the IA point of view:

    Human/computer  team  automation:  Take  problems  that  are  normally considered for purely  machine solution (like  hillclimbing problems),  and design programs and interfaces that take advantage of humans' intuition and available   computer   hardware.      Considering   the   bizarreness    of higher-dimensional hillclimbing problems (and the neat algorithms that have been devised  for  their  solution), some  very  interesting  displays  and control tools could be provided to the human team member.

    Human/computer  symbiosis  in  art:  Combine  the  graphic  generation capability of modern machines and the  esthetic sensibility of humans.   Of course, an enormous  amount of  research has gone  into designing  computer aids for artists.   I'm  suggesting that we  explicitly aim  for a  greater merging  of  competence,  that  we  explicitly  recognize  the  cooperative approach that  is possible.   Karl  Sims has  done wonderful  work in  this direction.11

    Human/computer teams at  chess tournaments: We  already have  programs that can play better than  almost all humans.  But  how much work has  been done on how  this power could  be used by  a human, to  get something  even better?  If such teams were allowed in at least some chess tournaments,  it could have the positive  effect on IA research  that allowing computers  in tournaments had for the corresponding niche in AI.

     Interfaces that allow  computer and network  access without  requiring the human to be tied  to one spot, sitting in  front of a computer.   (This aspect of  IA fits  so well  with known  economic advantages  that lots  of effort is already being spent on it.)

    The Internet as a combination human/machine tool.  Of all the items on the list,  progress in  this is  proceeding  the fastest.   The  power  and influence of the Internet are vastly  underestimated.  The very anarchy  of the  worldwide  net's  development  is  evidence  of  its  potential.    As connectivity, bandwidth, archive size, and computer speed all increase,  we are seeing something like  Lynn Margulis' vision of  the biosphere as  data processor recapitulated,  but at  a million  times greater  speed and  with millions of humanly intelligent agents (ourselves).

    The above examples  illustrate research  that can be  done within  the context of  contemporary computer  science departments.   There  are  other paradigms.  For example,  much of the work  in artificial intelligence  and neural nets would benefit  from a closer  connection with biological  life. Instead of  simply trying  to  model and  understand biological  life  with computers, research  could be  directed toward  the creation  of  composite systems that rely on biological life  for guidance, or for the features  we don't understand well  enough yet  to implement  in hardware.   A  longtime dream of science fiction has been direct brain-to-computer interfaces.   In fact, concrete work is being done in this area:

    Direct links into brains seem feasible, if the bit rate is low:  given human learning flexibility, the actual brain neuron targets might not  have to be precisely selected.  Even 100  bits per second would be of great  use to  stroke  victims  who  would   otherwise  be  confined  to   menu-driven interfaces.

    Plugging into the optic  trunk has the potential  for bandwidths of  1 Mbit/second or  so.    But  for  this,  we  need  to  know  the  fine-scale architecture of vision, and we need to place an enormous web of  electrodes with exquisite precision.  If we want our high-bandwidth connection to  add to the paths already present in the brain, the problem becomes vastly  more intractable.  Just sticking a grid of high-bandwidth receivers into a brain certainly won't  do it.   But  suppose that  the high-bandwidth  grid  were present as the  brain structure was  setting up, as  the embryo  developed. That suggests:

    Animal embryo experiments.   I wouldn't expect any  IA success in  the first years  of  such research,  but  giving developing  brains  access  to complex simulated neural structures might, in the long run, produce animals with additional sense paths and interesting intellectual abilities.