Hacker News Reader

Comment by RyanZAG | Hacker News Reader

I'd guess the correct form of that statement should be:

"If a human can't theoretically perform the task with a lot of time and a pocket calculator, then neither can an AI."

But it's not really saying much as that applies to almost everything.

As for the core idea of designing AI UX better to get people to trust it - maybe put a list of all the data being collected and stored up front, and make it possible for the user to easily and permanently delete specific data while still allowing them to use the system. Also an explanation of actions taken: "the system took this photograph because 76% of testers found photographs with these particular features (high light level, human looking at camera, action) to be desired." You could probably trace your dataset back to these features by collecting all the photos from training that matched to those features and manually labeling them.

In short: un-black-boxify your UX.

View on HN

> un-black-boxify your UX

I thought the whole point of AI was that the program itself is effectively a black box, even to the programmers.

They ask a neural network to solve a problem; it tries repeatedly until it succeeds. The algorithm it arrives at is not necessarily sufficiently simple that a human could understand it, even given the source code.

The problem isn't just that Google won't tell the user how their software works; that's nothing new. The scary thing is that even Google doesn't understand how their software works; they just know that it does seem to work most of the time.

Hopefully I've misunderstood completely, and we're not entering an age of “the algorithm (that we don't fully understand but is extremely useful most of the time) says so, and we shouldn't question its authority”.

> I thought the whole point of AI was that the program itself is effectively a black box, even to the programmers.

No, it isn't. In fact, it's the opposite of a responsible approach to AI, and one of the reason why you have smart people screaming about dangers of uncontrolled intelligence explosion.

Being a black box is a feature of a particular method people now choose to do AI - namely, neural networks. Neural networks are basically chains of multiplications, additions and thresholding; you start throwing tons of data at them and have it tweak its coefficients, until the result sorta looks like what you want. The knowledge hidden in those coefficients is almost completely opaque to us ("an open research problem").

People use deep learning because it's effective, but I hope that we go back to methods we're able to trace and analyze before we attempt to bootstrap a general AI.

I think you have it backwards there. Computers can already perform all sorts of calculations that a human can't. AI aims to make computers solve problems that currently only humans can.

Reply on news.ycombinator.com