Hacker News Reader

Comment by anon1253 | Hacker News Reader

There's a larger truth hidden in here: experts disagree. It was one of the things that took me by surprise when building Machine Learning systems, but it makes sense if you think about it. For all but the most trivial matters, it's the job of the expert to exercise personal judgement. It's why they're experts, not because they have a solid understanding of the "knowns", but because they have developed methods and intuitions about the "unknowns". Ask 10 different experts a complex question in their field of interest, and you'll likely get widely varying answers. My guess is that it also played a role in the early expert systems AI bust (AI winter). There are interesting papers published about this as well, e.g. http://www.sciencedirect.com/science/article/pii/S0951832007... the money shot is in this chart http://imgur.com/a/shaii but you'll see similar distributions everywhere where expert judgement is warranted.

View on HN

There is a corollary prediction: AIs will disagree.

Too many people imagine AIs are somehow "perfectly rational" and therefore will never disagree, they'll just tap into a unified decision making engine and automatically integrate all of their knowledge.

In order to integrate knowledge you need more knowledge. And in order to integrate that you need more still. It's turtles all the way down. If anything, AIs are at a disadvantage to humans, because we at least share a body, an emotional palette, and to some extent a cultural timeline. AIs have nothing to ground the beliefs of the other in.

If AI is to agree they need all the facts. And as we keep adding facts almost daily, there is simply no chance that the AI will have all the facts.

We like to think of logic as 1+1=2. But logic is more like A+B-C*D/E=F. Where any number of those being either completely unknown (we sometimes do not even know we are missing them), or some wide range of potential values.

You need a third model to map facts from one model into facts in another model. Doesn't really solve the problem.

Have you heard of Aumann's agreement theorem? https://en.wikipedia.org/wiki/Aumann's_agreement_theorem

Theoretically it means that two AI should be able to reach a consensus pretty quickly without needing to share a lot of information.

Unfortunately you need "honest, rational Bayesian agents with common priors" for this to work. Given that humans rarely agree with each other, it's interesting to think about where we fall short of that criteria.

The issue with the medical profession seems to be when you ask them about something not in their specialty.

Most doctors I've spoken to without a personal relationship refuse to comment and instead refer to a specialist.

... Which forces the patient to make his or her own choice, even though they most information they have is a summary from hurried medical professionals.

I understand why they do it, but it's a suboptimal result compared to their weighing the relative merits with their expert knowledge in at least one and offering advice.

Reply on news.ycombinator.com