Entropica Claims “Powerful New Kind of AI”

A new start-up called Entropica is claiming to have discovered mathematical equations that allow an autonomous system to select and achieve its own goals.

Entropica is a powerful new kind of artificial intelligence that can reproduce complex human behaviors, including the ability to autonomously set and implement its own goals.  In this video, we will see how Entropica can walk upright, use tools, cooperate, play games, make useful social introductions, globally deploy a fleet, and even earn money trading stocks, all without being told to do so.

Here’s the full pitch video:

Sounds great, doesn’t it?  But unfortunately, there’s a lot here that causes my hype detector to peg the red zone.  Let’s take the first claim: “Entropica can walk upright.”  As humanoid robot hobbyists, we know what an impressive feat that is; you have to coordinate a large number of actuators, maintain balance while moving forward, shift weight from side to side, deal with possible slopes or small obstacles, and so on.  Entropica can do all that?  Well, no.  What’s actually shown is Entropica solving the classic pole-balancing problem.  That’s a control problem that’s easy to solve with any of a dozen different techniques, and has almost no relationship to walking at all.

OK, how about making useful social introductions?  A robot that can say hello, introduce itself, and exchange a bit of personal information would indeed be pretty neat, especially if it did so without being specifically programmed to do that.  But again, no, that’s not what’s shown here.  Instead they show Entropica adding links in an acyclic graph, while other links are randomly removed.  Again, there are other fairly trivial solutions to that same problem.

Now, not to be a complete curmudgeon, it is normal and expected for new approaches to first cut their teeth on toy problems.  (It’s not all that normal to overhype the results of such toy-problem tests this much, at least on this side of the pond, however.)  It’s entirely possible that the authors (A. D. Wissner-Gross of Harvard and MIT, and C. E. Freer of Hawaii) really have found a useful new approach to solving problems.  The basic idea — that an agent should make choices that maximize its future options — makes some sense as a general principle.  Whether it scales up to nontrivial problems remains to be seen.

The full paper is behind a paywall at Physical Review Letters; and for more curmudgeonly analysis, see this New Yorker article.