Drop Everything and Work on AGI?

A good question, and one we explored at a recent Foresight Institute event.

My reply was driven by two question-answer pairs:

  • Is human-level general AI (AGI) possible in the near term? Yes, it is.
  • Is it desirable? (i.e. good outcome) Yes, in fact, it is probably necessary!

In my view, both points are best addressed by realizing that firstly, current AI research is badly off course for AGI; and secondly, that there is a much more direct path which can significantly accelerate achieving advanced AI.

To make my point I need to return to the roots of AI, and ask: What are the core properties and dynamics of actual generally intelligent system? What does true intelligence demand? Some answers are here.

Furthermore, I ask: Do current mainstream approaches to AI meet these requirements?

The answer is ‘no’ — current machine learning system are narrow, not general; they are static, not real-time adaptive; they don’t directly learn from their mistakes; and they can’t explain their actions or be reasoned with. In short, they aren’t actually very smart. Also, most importantly, these limitations are not incidental — they are largely inherent.

However if we follow the road less traveled — if we focus on real-time adaptive, self-learning AI systems capable of contextual, conceptual reasoning — we will not only make much faster progress towards AGI, but we will also mitigate much of the risk.

Specifically, truly intelligent systems will not suffer most of the problems outlined in the book Superintelligence: the various ‘alignment’ problems such as utility function, rouge instrumental goals, orthogonality issues, and ‘paper-clipping’ issues. The reason that these problems largely go away, is that these AIs are inherently designed to deeply understand knowledge, motivation, values, etc. in order to achieve their general-purpose cognitive competence.

Given the right approach, it seems that not only can we have safer AI and have it sooner, but it will also be much smarter and thus offer all the advantages of advanced AGI. It will help us solve problems that we are struggling with:

  • To comprehensively conquer disease and ageing
  • To help develop molecular nano-technology
  • To help provide cheap and clean energy
  • To improve human morality, and governance
  • Prevent us from wiping ourselves out

So then finally: Why aren’t more people working on real AGI? I will address this question in another post .