Is AI an existential threat to humanity?

Is AI an existential threat to humanity?

First…Stop it.

Artificial intelligence is not "intelligence". And it's not "artificial consciousness".

Everyone is afraid that AI will suddenly wake up, get upset, and take over the world.

Or that AI will wake up and take all of our jobs. This will happen. But without the "wake up" part.

Below I describe what real AI is.

If we want to understand the “existential threat” we first need to know what AI is.

Then, if you are at a cocktail party and someone says, "but what if robots are intelligent?" you can argue with facts, mixed with a little bit of alcohol.

---------

A) STATISTICS

Statistics is at the heart of most AI programs.

Just like statistics is at the heart of a lot of human decision making.

For instance, if you see clouds in the sky, your brain thinks: "Hmmm, the last 100 times I saw clouds this dark, it usually meant it was about to rain".

When you think like that, you are using statistics to make the decision: "I should _probably_ go inside now."

I'll give an AI example: Siri or Alexa. How does Alexa understand the words you just said?

In 1989 I was visiting Carnegie Mellon to decide if I would go to graduate school there.

One of the graduate students, Kai-Fu-Lee (now one of the most famous investors in the world and I would check out his excellent recent TED talk on AI) showed me what he was working on:

It was speech recognition for the 60 or so commands that might happen on a Navy battleship (ten guesses as to who was funding his project).

When you say the word "Fire!" a sound wave is created. When you say the word "hello" a sound wave that looks different is created.

If 100 people say "Fire" and 100 people say "hello", all of those sounds waves are stored in a database.

Now, if a brand new person says "Hello" the computer program needs to determine if that person said "Hello" or "fire".

There might be 10 different attributes of every sound wave. It breaks the new person's sound wave into those 10 attributes.

Then it compares that "vector" of 10 attributes with all of the vectors in its database for "Hello" and "Fire!"

It uses a statistical technique called "Hidden Markov Analysis" to determine if the sound wave is more like the "hello"s in the database or more like the "Fire!" in the database.

Then it says to itself, "This guy said "Hello". "

It then has a line of code that says, "If someone says "Hello" Then say "Hello" back".

Additionally, it adds your "Hello" to its database.

Your "Hello" might be slightly different than the other 100 "Hello"s so it just learned a new way to say "Hello". That gives it greater ability in the future to recognize the word "hello".

In other words, it "learned".

So it used Statistics to hear you, code to respond to you, and database technology to learn. There's no real intelligence there but it feels like it's intelligence.

Multiply that by 30 years and millions of patterns and computers a million times faster and you have Alexa and Siri in today's kitchens.

Ask "Siri" what gender it is.

-----

B) EVALUATION FUNCTION

I just mentioned about language recognition. But how does a self-driving car work?

Every second it has to make a decision. Does it move forward? Does it brake? Does it swerve to avoid an accident? Does it turn left?

How does it get from point A to point B?

1) Google Maps. - Using GPS it knows where it is. And it puts itself on Google Maps.

2) List all of the possible routes. This is a "hard" problem in the mathematical sense (there's no way for it to guess the fastest route. It has to list each route and then sort by the shortest. )

But now computers are so fast what would normally be a slow decision (drive me from this corner in Piscataway, New Jersey to the capital building of Sacramento, California) now just takes seconds.

3) Waze. Use Waze to eliminate the routes with too much traffic.

4) Start driving.

5) Statistics: Every microsecond it uses statistics to see if there is blank space or an object that must be avoided or a traffic sign that must be followed.

6) Decide what to do according to the code. For each traffic sign, it has code that tells it what to do (if a sign says "Stop" it Stops for a second, uses Statistics to see if any traffic is happening on its sides (with radar and cameras to provide the images). )

If there is a person standing in front of it, it might just stop.

If there's traffic it didn't expect, it might trigger the program to re-route.

If it's blank space it will just keep going.

If there's a baby crossing the street and it has to swerve to avoid hitting it, but if swerving will cause the car to hit a truck, killing the passenger in the car, then the "AI" of the car is dependent on the ethical decisions of the programmer of the car.

In other words, in every situation, it determines it's options, then uses an "evaluation function" programmed by a coder, to determine which option has the most successful outcome (move the trip forward, don't kill anyone).

Eventually the evaluation function will NOT be programmed by a human coder.

Instead, through thousands of experiences of other self-driving cars, the experiences plus the outcomes will all be put into a central database.

When a new experience is encountered, the code will look up that experience in the database and the database will spit back the best possible outcome.

The code will learn statistically what the best outcomes are of each possibly decision and change the code accordingly and send updates to all self-driving cars.

-----

C) TREES

The hardest game in the world is a board game called GO. With chess, if a computer can evaluate a billion possibilities a second, it can be a world champion level player.

But a Go game can involve trillions of possibilities. How did Google make a program, Deep Go, to beat the world's best Go player. This was thought to be impossible.

And yet Google did it.

For any game, a computer program first builds a tree of possibilities. Much like a human would.

A human thinks: "If I make this move in checkers, my opponent might respond with A, B, or C and then I can do D, E, or F and then my opponent can do G, H, I if I do D or it can do J, K, L if I do E and I'm never going to do F.

A computer doesn't select as well as a human so it builds the FULL tree. Meaning, what are ALL of the possible moves it can do, what are ALL of the possible responses of my opponent, etc.

And then it uses a programmed evaluation function to look at the leaves of the tree it built.

Whichever move results in the best leaf of the tree (as determined by the evaluation function) that is the move it makes.

That's how computer chess worked for decades. I'll get to the secret sauce in a second for how computers conquered chess.

And then after that I'll describe how computers miraculously conquered Go.

It's only a miracle until science can explain it. It's only "intelligence" until it can be coded by a programmer.

D) HARDWARE

Everybody thought for decades (including many Nobel Prize winners) that the best computer chess programs would be developed when scientists encoded the knowledge of the best chess players in the world into the evaluation function.

How does the world champion value a position instead of a weak player?

This turned out to be wrong.

The MORE code in the evaluation function (i.e. the "smarter" the evaluation function was from a human perspective) the SLOWER the program.

Which meant a smaller tree would be built, which meant less possibilities would be analyzed.

What really allowed the programmers at IBM to build "Deep Blue" which beat Garry Kasparov in 1997 were two things.

Both related to hardware.

a. Computers got faster.
b. First the creators of Deep Blue developed software. But then they made the software into hardware, building the logic right into the hardware infrastructure of the computer. Making the program 100x faster than it would have been.

And finally, they made the evaluation function STUPID in order to use less code so the hardware could value more positions.

Then, before anyone caught on to their "artificial intelligence" they retired Deep Blue right after it beat the World Champion of chess.

As hardware gets faster, artificial intelligence gets "smarter".

[as an aside, I once gave a date a chip that was the initial chip for “Chip Test” - the “ancestor” of what became the best chess computer, Deep Blue. She was weirded out.]

----

INTERLUDE

What I just described is all the basics. You can stop now.

The rest of artificial intelligence is simply combining the basics to make more advanced techniques.

-----

E) STATISTICS + TREE

Remember the TREE from computer gaming. And STATISTICS from speech recognition.

Now let's go to the impossible game of Go. Google developed the program "AlphaGo" to win at Go when everyone else thought it would take another 20 to 50 years.

First, remember Kai-Fu Lee who worked on speech recognition. And later developed Apple's first attempts at speech recognition in the 90s?

At one point in his grad student days, he was getting tired of navy battleship commands (as one does) and decided to focus on building a program to play Othello.

He ended up building the world champion of Othello.

He took a lot of games, let's say a million, and put them in a database. And each position from each game, he would label, "winning" (if it was a position on the winning side) or "losing" in a massive database.

He would identify several attributes of each position (how many white pieces, versus black pieces, how many corners were controlled, how many pieces were on the sides, etc).

Now, if the computer was playing a brand new game, it would determine all the attributes of that position, then use Hidden Markov Analysis (remember: speech recognition) to match that position to the database.

If the position pattern-matched a "winning position" then it would make the move that would lead to that winning position. If it matched a "losing position" it would not make that move.

That program became the world champion of Othello.

AlphaGo took it one step further.

It put in the positions of millions of Go positions and did the same sort of breakdown.

It used faster hardware to speed up the process.

Then, once it became pretty good at GO, it played BILLIONS of games against ITSELF to put many BILLIONS of new positions into the database. In other words, it "learned".

Now it was ready to play Go. It crushed the world champion

------

That's basically it. That's all of artificial intelligence.

Let's say a bank wanted to fire all of the employees in charge of lending. And replace them by artificial intelligence.

How would the bank lend money?

Well, there's 100s of millions of loans already out there. And for each person who has ever borrowed money I know:

- their age
- where they grew up
- what their job is, are they married?
- are they divorced? do they have kids?
- How often do they move? how have they done on prior loans like this? and I even know what they buy on Amazon and how often they fly to Las Vegas.

I can put all these vectors in a database and divide them into people "most likely to pay back the loan" and people "most likely to default".

Then, just like speech recognition or the Othello program above, I can use statistics to determine who I should loan money to.

And if I say "no", I don't have to explain. On to the next one!

---

Let's say I want to fight terrorists.

I already have examples of many terrorists who trained in the US and then went on to perform or attempt acts of terror.

I know everything about their bank accounts. How often they transferred money. How often they traveled. How often they took out cash versus using a debit card.

And so on.

I can build a vector of attributes of what a terrorist bank account looks like. Then I can match new people against that database of vectors of terrorists.

Believe me, every time you do a bank transfer, some AI program is out there trying to determine if you are a terrorist.

----

This is all that AI is.

It is nothing more. It's not "intelligent" from a human sense. It's not conscious, nor will it ever be.

Here's how AI has improved in the past forty years (and how it will improve the next 40):

- statistics has gotten better
- methods of building the trees have gotten better (this was the subject of some of my research when I was in graduate school)
- hardware has gotten faster
- more data is available about everything.

What is changing the fastest is data. The land grab of modern society is not land, or gold, or oil.

It's data.

I have been invested in many companies that collect and sell data. I was an early investor (and on the board of) bit.ly, as an example. bit.ly accounts for about 2-5% of all Internet traffic.

Believe me when I say, data-driven companies know how many strawberries you ate last summer.

And right now that data is used mostly to target you for ads about sneakers. Or politics.

But this is AI 1.0. Soon that data will be used to target your every movement, your every want, your every need.

Amazon Prime won't be about delivering you what you want tomorrow. Amazon Prime Plus will be about delivering you what you want yesterday.

Police 2.0 will be like the movie "Minority Report".

Even art and music will be driven by AI that studies the neurochemical responses to music you like to music you don't like. And then compose accordingly.

Where will humans still be unique?

I don't know. Ask the humans with AI implants that enhance their brains so when they look at you they know exactly what answers will make you happy.

---

BUT… will AI replace jobs?

The answer (at least in the next decade or so…) is NO.

Look at recent examples:

A) Many people were worried ATM machines would replace bank tellers.

Instead, the banks made so much in profits they opened up more branches than ever, creating new jobs.

B) Will autonomous delivery services cost jobs.

Right now there are millions of truck drivers involved in delivering goods. With autonomous delivery, less people will go shopping, more people will be required to shop in the aisles, finding products for people.

Obviously this is not a high-end job. But this replaces the fact that less cashiers and drivers will be needed.

Meanwhile, there will be more high-end jobs. More maintenance engineers for the cars, customer service, marketing, etc.

C) Ecommerce. Branding will become less important (branding is VERY important when everyone is shopping at the big box store but advertising will have to become more clever and digital) so the millions in profits that are generated from AI will filter down to more people starting e-commerce ventures and the ancillary businesses associated with that.)

Final conclusion:

• AI will probably create a “have” and “have not” situation, particularly as humans start to use AI to increase mental and physical capacity

This is probably a net NEGATIVE for society as the higher classes will be able to afford “super AI” capabilities, making them demi-gods to lower-classes.

• AI will not destroy as many jobs

Instead, massive profits will be generated, which will be soaked into the economy through a rising stock market, increase in opportunities, etc.

• We can’t predict. ATMs didn’t destroy bank tellers. VCRs didn’t destroy movie theaters. Spotify/Pandora/etc did destroy music stores and record sales but that was replaced by growing revenues in music tours.
• Education needs to scale up. AI, programming, and the higher-end jobs that will be created need to be studied. College is not the place to study these opportunities. Instead: Khan Academy, Lynda, CodeAcademy, Coursera, etc should become accredited and get people ready for the opportunities that will arise.