Hacker News Reader

Comment by matt_wulfeck | Hacker News Reader

when I think about AI, I'm always reminded about glitches. For example, finding a weird behavior that can be exploited in a video game to beat a boss easily. It's one principle difference between playing other people and the cpu: humans are able to evolve their strategies, experiment, etc.

For the AI to really be the best, it must sometimes lose because it's trying a new strategy and experimenting. It also must be robust and find the error of its way easily -- in other words, know quickly that it's being gamed.

AI must be able to reprogram itself, or it's only a matter of time before a winning strategy comes out against it. If it can't reprogram itself and evolve without humans is it really AI?

View on HN

> ... it must sometimes lose ...

Well, yea. In training, AlphaGo played against itself, and lost every one of those games. (Incidentally, it also won every one.)

In some contexts, sure, humans will adjust their strategy when someone finds a weakness to exploit against us. Other times, not so much. Ask any casino or the folks over at King/Zynga. Objectively, walking into one of those games is expected to be a losing strategy, but we've continued to do it for hundreds of years. In other contexts like chess and Super Mario, maybe we do better.

But if we make a general purpose AI that can adapt against any discovered winning strategies instead of falling into repeatedly losing, it won't be our equal. It will have surpassed us.

Only a certain percentage of people are addicted to gambling. For many, it's a simple calculus of the entertainment value. If 45 minutes of playing penny slots is worth the 10 bucks or whatever then they do it. I do agree with your general thesis though, there's definitely some areas where our logical abilities short circuit.

What if the video game AI decides that losing a boss fight by getting stabbed repeatedly by a player who glitched through a wall is totally hilarious and more fun that fighting? Bug? Or just an AI doing AI things?

We have squishier heuristics for what "winning" is compared to AlphaGo, and that seems likely to stay true for the foreseeable future of specialized AIs. But someday it might not.

"For the AI to really be the best, it must sometimes lose because it's trying a new strategy and experimenting"

It can also let the experimenting to its opponents and learn from that. That largely is what humans do, too. Generals plan for the last war, and the students who aren't challenged tend to learn badly or not at all.

And, apparently, the Google AI for go would learn from losing games. Whether it would learn enough to eventually beat (or at least tie) every possible strategy thrown at it, afaik is an open question.

Human cognitive biases are easy to exploit too.

Intelligence is in the eye of the beholder.

Reply on news.ycombinator.com