The Limits of AI: Generative AI, NLP, AGI, & What’s Next?

Category: AI Overview

Tags: AICapabilitiesCollaborationKnowledgeLimitations

Entities: Artificial IntelligenceDeep BlueGary KasparovIBMWatson

Transcript

00:00

Artificial intelligence is everywhere right  now. In your phone, in your car, even writing   emails for you.

You may be wondering if there  are actually any limits to what AI can do. I've   heard many people over the last few decades  confidently assert AI can do certain things,  

00:16

but it's never going to be able to do, and  then you fill in the blank. Guess what most   of those predictions have in common?

They were  wrong. The past few years have shown exponential   growth in AI capabilities, bringing it from the  research lab to everyday life.

And it's doing   most of those things that so many thought  it never would or even could do. Of course,  

00:35

many limitations still exist, but my advice would  be this. Don't bet against AI, unless of course   you want to be wrong.

In this video, we're going  to start with a look at what knowledge really is,   how it differs from data and information, and  this will help set the context. Then we'll take  

00:52

a look at what have been considered to be the  limits of AI and see which ones of those things   have actually been accomplished and what's still  left to do. Then we'll conclude with some ideas   about the role of AI and humans and where each one  excels with the hope of learning how to use this  

01:09

amazing technology to our best advantage. Let's  start off with looking at the relationship that   exists among data, information, knowledge, and  wisdom.

and we'll use this pyramid to spell it   out. So, we'll start with data.

Okay, this is just  basically raw facts. If I give you data that looks  

01:30

like this, I say 10 six uh 42 and 8. Okay, that's  raw facts.

So, what you don't know really what to   do with that, but that's data for you. Okay, now  if I add some context to this data, now we have  

01:47

information. So this is where we sort of processed  it a little more and now I'm going to tell you   that this data actually represents the ages of  people in a room.

So now we have more context.   This has more meaning to us. Now if I take that  and say okay but let's apply some interpretation  

02:08

to the information that we just had. Now we end up  with knowledge.

Now knowledge tells us yet more.   So for instance in this case we might say okay  I've observed that most of the people in this   room are under the age of 21. So now we've done  yet more processing with this.

And now finally the  

02:30

last piece of this is applied knowledge. Applied  knowledge now gives us wisdom.

and wisdom might   look at this all of this information, all of this  data, all of this knowledge and say, you know   what, we've got these people in a room. Let's  do something like uh do age appropriate games  

02:50

to keep them occupied. So, uh the 42-year-old  probably won't mind too much playing a game that   a 10-year-old and a and an 8-year-old would play,  but you know, they they can go along with that   for a little while.

So this is an example very  trivial example but you can see what I've done  

03:06

here data information knowledge and wisdom each  one of these adds more context more interpretation   and all of these then lead to the ultimate of  wisdom. So another way to look at this pyramid   is data.

Well, that's a database. For instance,  you know, we can store a lot of stuff in there,  

03:25

but that's all it is, just a collection of raw  facts. Information, okay, we have an application   running on a computer.

That's now information  technology. That's why we call it that.

We've   added context to all of that data. Knowledge, this  is where AI really starts to come in.

Now, we're  

03:41

adding more interpretation to the information  that we've just processed. But here is where we're   still trying to get.

And that's wisdom. Back when  I was an undergrad riding my dinosaur to class and   studying AI in its earliest days, there were a lot  of things that people said, "These are the limits  

03:57

of AI. Maybe one day we'll have a system that's  able to do these, but they won't be anywhere,   maybe even in our lifetimes." For instance, one of  the things that was talked about was the ability   to reason.

We needed a system. If we really  consider it intelligence, then then reasoning  

04:12

is a part of that. So the ability to figure out  and do problem solving, complex problem solving,   uh this was beyond our capability uh certainly in  those days.

But since then we've come out with a   computer that can play chess. IBM in 1997 came  out with a computer called Deep Blue that played  

04:33

Gary Kasparov, the best chess player in the world,  a grandmaster. That's a lot of reasoning.

That's   a lot of problem solving. People thought you'd  never have a computer that would be able to beat   a grandmaster.

Again, that's already happened. So,  what seemed to be a limitation wasn't.

Another one  

04:49

that was really difficult for a long time was  natural language processing. Uh, human language   has a lot of nuance, a lot of idioms, things  where we say things that we don't mean literally.   And sometimes you're supposed to interpret it  literally, sometimes it is figurative speech.  

05:06

Uh for instance, as I've given examples before, if  we say it's raining cats and dogs, we know that it   doesn't mean that there are small animals falling  out of the sky. That's an idiom.

So we have if a   system is going to really be intelligent in  the way that we are. It needs to be able to   understand those things.

It needs to be able  to understand things like humor and understand  

05:24

when you're cracking a joke and when you're not.  Well, sometimes people can't tell that either,   and sometimes it's because it's a bad dad joke.  But be that as it may, in general, we're able to   tell the difference between what is humor and what  is not. And we've actually made some advancements  

05:39

here. In 1965, there came about a first the first  of what really is the modern chat bots, but this   was not using modern technology called Eliza.

And  it was able to have conversations with you. Now,   it wasn't very great conversations, but it would  ask you questions and and answer questions.  

05:59

how are you feeling today? How does that make you  feel?

Uh this kind of thing almost like you feel   like you're talking to one of these very passive  psychologists. Uh but IBM advanced this a lot in   2011 when we came out with Watson which played  Jeopardy the uh TV game show and was able to win  

06:18

and beat champions at that because Jeopardy is  full of natural language and play on words, puns   and things like that. You can't program all of  those into the system and have it know those.

It   really has to understand the meanings behind  those things in order to do it. And in fact,  

06:34

as I say, we've already accomplished that. And  look at today's modern chat bots.

They're able   to understand a lot of this nuance and they're  able to take the instructions you give it in   natural language and understand what you mean in  a surprising way. In fact, I think that's maybe  

06:50

one of the most remarkable aspects of generative  AI technology is that it's able to do that for   the first time. We feel like a computer really  understands us.

It's able to infer what we're   asking for. In some cases, even anticipate the  next thing that we need, just like a person  

07:05

would. We consider that to be intelligent.  How about creativity?

The ability to create.   I remember hearing a lot of people say, you know,  computers can't really create information. Well,   they actually do.

Uh, we've got where with  generative AI, we can create art. We can create  

07:22

new works of music. And you can say, well, but  those are really just mashups of existing.

Well,   guess what? When people compose a new song or draw  a new picture, we're influenced by the things that   we've heard as well.

Listen to all the top musical  artists that you know, and they'll tell you, "Oh,  

07:39

yeah. Here are my musical influences." So, those  things all went into the back of their heads and   influenced the way that they create.

So, we are  creating new things and they are variations on   the old. But that doesn't mean just because a  computer did it, it wasn't creative because in   fact it is.

They're coming up with new ideas and  will continue to do that. We base our learning and  

07:58

our creativity on certain things that have  been done in the past and so does AI. Now,   here's another one.

Real time perception.  Things like robots. Well, that was the stuff   of science fiction at one point, but we have them  today.

And you might not think of it as a robot,  

08:16

but a self-driving car is one of those where it's  having to in real time perceive its environment,   see what's going on, anticipate where the next  car is going to move, and where it's going to   be at a specific point in time and do all of  those calculations in real time, and make real  

08:34

uh decisions about that. Robots are having to do  the same thing in order to navigate around a room.   So all of these things that basically we used to  consider to be limits of AI, I'm going to say,   you know what, we've done all of those.

Now, let's  take a look at some other areas where we've made  

08:51

progress, but I don't know if we would say, you  know, it's sort of mission accomplished yet. And   one of those would be uh the area of you've  heard of an IQ, how about an EQ, an emotional   intelligence uh and an index for that?

Well, these  systems are able to simulate that. And honestly,  

09:10

I feel like some people are just able to  simulate emotional intelligence as well,   but that's a whole other subject. But an EQ in  a system, you can see in the modern chat bots   the ability for them to understand your moods  and the way that you're expressing yourself.   So there is some level of awareness in terms of  the way that you're describing things.

I mean,  

09:28

we have the stories about people who felt an  emotional relationship to a chatbot. Well,   some people feel emotional relationship to their  shoe, but that's a whole other thing.

The fact   that these systems can talk to us and understand  at least give the appearance of understanding  

09:45

moods and things like that is certainly in the  area of okay, I it looks like we're doing this   at least in some cases. Now, another area that's a  limitation though that we still have is this area   of hallucinations.

Hallucinations are a difficult  problem. and they're a a byproduct of generative  

10:03

AI where the system basically confidently asserts  something that just isn't true. So it's trying to   predict what the right answer would be and many  many times it's right.

It's shockingly right.   But when it's wrong, it is shockingly wrong in  these cases. Now we've got technologies that  

10:20

are making hallucinations less and less likely.  Uh things like retrieval, augmented generation   uh helps with this where we feed additional  information to give more context so that the   model doesn't just use its own imagination to come  up with answers. Uh things like mixture of experts  

10:37

helps as well where we have different models  used for different areas. Chaining of models.   Uh so there are things that we can do in order  to reduce the hallucination problem and we're   doing that.

So this is one of those I wouldn't say  uh is a solved problem but we can certainly see  

10:55

that we're moving into it. So this one's somewhat  solved.

Okay. So, those are the things that we've   kind of already done or are still working on  and maybe be able to see an end in sight.

Let's   move those out of the way. And now, let's take a  look at the future.

In other words, what are the  

11:11

current limits? What are the problems that we're  still having to to work on these days?

Well, one   of the limits of AI is a thing called artificial  general intelligence. Right now, we see AIs that   are super smart in a specific area, in a specific  knowledge area.

Now again with some of these chat  

11:28

bots that we have today, they seem to know a lot  about pretty much everything, but they also have   limitations. For instance, they don't do real-time  perception.

Uh they can't tie their own shoes,   for instance. So artificial general intelligence  would be something that was as smart as a person  

11:43

doing all the things that we consider to be  intelligent and at least on par with what a   person would do across all the different domains.  That's something that we haven't really fully   achieved in a single system yet. The next level  beyond that would be artificial super intelligence   where we have something that is better than  humans in every domain and that's the right  

12:03

now again the stuff of science fiction. Not saying  that we won't do it but we haven't really done it   yet.

Another problem that's still to be solved is  with sustainability. So right now we have systems   that can do amazing stuff but boy do they suck  up the gas.

They take up all the electricity.  

12:22

They need lots of cooling. They're very expensive  to run.

This is not something that's going to be   able to scale if we just keep throwing more and  more processors at this situation. Uh that's not   going to work.

We're going to end up using all the  electricity that's on the planet just in order to  

12:37

uh to to run some of these queries. So, we're  going to have to be able to make better, smarter   decisions with sustainability.

Use models that are  the right size, not just the biggest model, but   the right size model. In some cases, a small model  might be more efficient and do a better job and  

12:52

might even hallucinate less if we've got the right  use case. So, this is still work that we're that   we're doing that is not yet, I would say, a solved  problem, but there's a lot of things we can do   about it.

Another one that is really the area that  is is science fiction today is self-awareness. So,  

13:10

is a system self-aware? Does it know it exists?  Does it have consciousness?

Well, I don't really   know the answer to that. This is really not a  computer science question.

This is a philosophy   question. So, I'm not going to try to deal with  that one here because I'm not even sure how the  

13:27

answer would be. But another thing that gets us  back into this area though is understanding.

So,   a system can spit out a lot of things, but  it actually understand what it's saying. Um,   does it really know what the meaning of the  things are?

Seems like it's done a lot of that,  

13:44

but there's always the question of is this really  just simulating? Is it simulating thought?

Well,   I don't know. I'll tell you there's a lot  of people I've talked to and I think they   may be only simulating thought and simulating  intelligence.

So, again, it's a little hard to to   draw the line clearly, but uh this seems to be an  limitation where the AI maybe doesn't understand  

14:06

the biggest broadest context that we'd like it  to understand. Uh judgment.

So remember when I   was talking about data, information, knowledge  and wisdom. Well, this is that last one.

This is   the business of wisdom and judgment. And in this  case, is the system able to make good judgments?  

14:26

Maybe ethical judgments. Can it determine what  is right and what's wrong?

Again, can people do   that? Some people have a real hard problem with  those kind of judgments.

So, it's hard for us   to program a system that will if we can't figure  out what those rules would be. But we we certainly  

14:42

know that right now these are limitations that  the systems have. How about in terms of judging   something that's just very subjective like the  quality of something, maybe music?

You know,   what I think is really great music, you may  not think. So, you know, you say, "Well, Jeff,   you have no judgment at all." Uh, but I have a  different view of that.

But these systems are they  

15:02

able they're able to generate music and they're  able to throw away stuff that is just absolute   gibberish but can they tell what is going to be a  hit and what's not going to be for instance in the   music area. So there's a lot of of work here in  this space so that it's able to do some of those   qualitative judgments as well.

How about this one  common sense? And I'm going to really put that one  

15:24

in uh in quotes because air quotes because um I  mean is it really all that common? It seems like   again we have limitations with people.

So we can't  really expect a system to be able to perfectly do   what we consider to be common sense because we all  might have a different idea about that. Certainly  

15:42

there are some things that we know and the systems  ought to be able to understand that but today   there are some certainly some limitations to  that. How about in terms of goal setting?

Well,   some people would say that with today's agentic  AI that a system can in fact set its own goals and  

16:00

go off and accomplish those things. And what I'm  going to make a distinction here is that we have   micro goals.

These are sort of the small things  that we need to do if I give you a a larger task   and the macro goals. So the larger task, this  is what needs to be done.

This is how I go about  

16:18

doing it. And right now today's agents are able  to do these kind of micro goals, the goals within   the larger objective, but the big goal, why would  we do this in the first place?

That's maybe still   uh without uh beyond its reach at the moment. And  then sensation, how about this?

Does this does a  

16:36

AI system really sense things? Does it understand  what's happening?

What how things feel? How things   taste?

Um that sort of stuff. The things that  are of the senses.

Well, we're building robots   that are able to certainly see and hear. Can  they taste?

In some cases, maybe to an extent,  

16:52

but there's a lot of other things that go into uh  these kinds of sensations that we haven't put all   together in one system. And then here's the really  big one, I think, and that's deep emotions.

Is a   system really able to feel the same way that we  do? Is it able to experience joy?

Is it able to  

17:12

experience sadness, loss, uh, accomplishment? Does  it really get what all that's about?

And again,   I know some people who don't really do all  that particularly well. So, so this is one   of the things that is difficult to put into a  system and we can simulate it today, but is it  

17:31

really feeling these kinds of things? So, I would  suggest to you these are some of the things that   to one degree or another are limitations with  today's AI.

Now, what is the role for humans   and for AI? How do we work together?

How do we  make sure this is a tool that works for us? Well,  

17:47

people really should be over here doing this kind  of stuff. Answering the what question.

What is it   we want to do? That's the overall macrolevel goal,  the objective and answering the question why.   What's the purpose of this?

Is there meaning in  what we're doing? What's the ultimate thing that  

18:05

we're trying to accomplish? And without purpose,  all of this is just meaningless work.

So people   are still far better at that kind of thing. And we  should be the ones controlling this tool that way.   Over here on this side, once we've told the system  what needs to be done, AI can many cases with an  

18:23

agent figure out the how and go off and perform  it, actually do it. Agents are able to automate a   lot of things much faster than a person could.

and  they can do it in an optimized way, but they need   to know what to do in the first place. We need  to know why.

So, if you look at a history of AI,  

18:41

it felt like for the longest time we were making  very little progress and then all of a sudden it   just took off. And we're at this this inflection  point where the developments and where all of this   is going to go, no one really knows.

But what I  can say this for sure, we can look at a history  

19:00

of milestones that we've accomplished already.  And we can look at lots of future research, things   that still need to be done, which is actually  very exciting. If you're someone that enjoys   it and the possibilities of problem solving, then  we're going to be able to do a lot more work and  

19:16

ultimately we're going to get end up with systems  that do things we didn't even imagine yet.   So my advice to you if you start looking at  the limitations of AI today I would say don't   become preoccupied with those because the people  who have and have asserted that AI will never do  

19:33

this that or the other thing have generally been  wrong. My advice to you, don't bet against AI.