AI?

Don't know, Lex Fridman has a pretty good podcast... I've only watched a few

GH

Prompt: "Are there some fundamental new principles to Artificial Intelligence? Seems to me like it's still just another heap of code."

Completion:

"As they say, it's an API, it's just a way for us to create the algorithm that will generate the result.

So when it's used say to create a game, it still just gives you the algorithm. But when someone else comes up with this algorithm and you're like, what is that different? And then they create their own game, they have very different algorithms for using the AI. And then there are companies and people that want to make AI-type games, who have no idea either way where AI comes from or where the AI in a game is coming from.

It's kind of silly, and so it's really frustrating and I think that developers are always on this path of just wanting to just see what everyone else is doing, and I'm sure that they feel frustrated by that as well. So the big one is that when AI is introduced into the ecosystem, we need to recognize how different that is because of this artificial complexity that developers now have to deal with and figure out how to make it work better and in the same spirit of wanting it to be fun but also making it practical."

...

The rest of your post was generated by an AI

Prompt: "In his next five years in office, Donald Trump will definitely make America the greatest it has ever been!"

AI Completion:

He's gonna make America great again!

He's gonna make America great again, no one's going to hate his country, it's the greatest country in the world.

Thank you, President Trump of the United States of America, if only The Donald won't have to worry about the American people constantly trying to kill him!

Semantic analysis, language-independent context, it "understands" the subject of the prompt and not only generates grammatically correct text but text on the appropriate topic with the right tone and that is logically consistent

Yes there are fundamental differences...

The "code" is in the numbers defining various weights inside the system. It is a neat example of Turing's insight that there is no difference between data and instructions - except in how you interpret the numbers. Softies that only work at a high level typically don't see that as being obvious.

The weights are neither designed nor understood by humans; they emerge during training. The corollary is that nobody knows /why/ something is behaving the way it is, nor can they predict what input changes might produce unacceptable outputs.

The latter behaviour is seen in wetware, but the wetware is more likely to be able to explain why it is behaving in a certain way. But even that isn't guaranteed, doubly so when malice is considered.

Yes. At least for certain types of problem which includes some interpretation of large scale data and two player games like Go.

Google's Alpha-Go and the later Alpha-Go Zero was able to exceed 3000 years of accumulated human activity on the game in about 40 days from just being given the rules playing against itself. It even found some new joseki that had never before occurred in human games as a result.

When the guy behind the design of this program gave a talk in Cambridge UK the queue to get in gridlocked parts of the city.

In some ways it is and in a scary side effect you can't dissect the reasoning it has used to come to a final conclusion beyond that its prediction is that one particular choice is optimal.

In medical diagnosis the edge cases are checked by a skilled human but the clear negatives and clear positives can be dealt with by a robot that never gets tired, distracted or has bad days. They are getting almost as good as the best practitioners at some types of X-rays.

That was true then but even so it had insights that made Kasparov believe that the machine team was cheating. It is hard today to find a serious chess engine running on a modern PC that cannot beat him!

Googles Alpha Go Zero has gone well beyond that point. It is capable of inventing new strategies for a game ab initio given only the rules.

In a similar way it can learn to interpret bulk data and choose its own hypotheses across huge multidimensional datasets gaining insights that humans cannot hope to spot. If it finds something interesting then humans can look at the output and try to interpret it.

There is one in the White House right now. The twit that tweets.

arning/

It wasn't actually true. Deep Blue didn't search the complete tree of possi bilities - the classic brute force approach - but rather a tree that got pr uned as soon as any particular branch was clearly sub-optimal. The pruning algorithms had had a lot of work. Deep Blue still looked at a lot more bran ches than a chess master, but the tree was pruned well enough that it cloud look further ahead than Kasparov when it mattered.

His problem is that he rarely seems realise that there is a tree of possibi lities to search, and even when he does, he lacks the patience to sit aroun d while somebody else searches it for him and gives him a list of more or l ess rational choices.

Are there some fundamental new principles to Artificial Intelligence?

The pruning algorithms were not all that different to existing state of the art stuff though. What tipped the balance was that it has dedicated chip level parallel hardware move generation and material balance score. It was massively parallel in dedicated hardware - that is fundamentally a brute force solution. The thing that it did exceptionally well was search extensions when they were necessary to reach true quiessence.

Tsu's book "Behind Deep Blue" is an interesting read. Some of the technical papers are online for example here:

Deep Blue

It was a lot closer than that.

If Kasparov hadn't been spooked by some of its unusual human like moves and had played like his best games he could probably have won the match. Unfortunately he got annoyed with it and made some untypical for him mistakes. It was a machine playing unlike any other machine before it.

I'd like a navigation app that took into account topography when planning a route, for e.g. hybrid vehicles. the shortest route isn't necessarily the most fuel-economical route. You'd think fleets would be interested in something like that, maybe they already have it. I know in some areas UPS optimizes routes to avoid left turns.

It would be nice if routes could be optimized for ease-of-driving. Waze is sometimes too clever for its own good; it sees "intellectually" that a certain route is in theory faster, but doesn't "know" that it's just sent you down a route that involves several challenging turns or blind corners and/or narrow streets that end up costing you time.

the latter seems like you'd have to have some kind of human-initiated rating system initially to get data to aggregate driver experience from.