Data Driven Peter Voss on Artificial General Intelligence, Personalizing Personal Assistants, and Motorcycles
In this episode of Data Driven, Frank and Andy speak with Peter Voss about Artificial General Intelligence, Personalizing Personal Assistants, and Motorcycles
Sponsor
Sponsor: Audible.com - Get a free audio book and support DataDriven - visit thedatadrivenbook.com!
Guest Bio
Peter Voss is the world’s foremost authority in Artificial General Intelligence.
His company Aigo (https://www.aigo.ai/) has created the world’s first intelligent cognitive assistant.
Aigo was funded with a personal investment of $10 million dollars. They currently manage millions of personalized customer service inquiries for household name-brands
Notable Quotes
Aigo is Peter's company. BAILeY's Introduction (00:00)
The east coast has been blanketed with snow. (01:30)
The Expanse books (03:00)
Coding for curiosity? - Frank (11:50)
"Models don't dynamically learn." - Peter (13:00)
Three waves: Logic programming, Deep learning / neural networks, cognitive architecture / intelligence (14:00)
Intelligence v. sentience? - Frank (15:50)
What about bots being "led astray?" - Andy (18:30)
On programming morality... (21:30)
AI Safety is a better description - Peter (22:30)
Asimov's three laws of robotics - Frank (23:15)
On delimmas - Peter (24:15)
"Morality should be about human flourishing." - Peter (25:15)
Are we using digital means to do something analog? - Andy (27:55)
Peter is trained as an electronics engineer. (28:05)
"Context is always super-important." - Peter (28:30)
"You need a feedback system." - Peter (30:00)
AIGO is Peter's company. (31:00)
The three meanings of personal. (34:00)
"Exo-cortex" (33:50)
On context switches (38:30)
Did you find AI or did AI find you? (41:00)
"I took five years off to study..." - Peter (43:00)
What's your favorite part of your current gig? (44:10)
When I'm not working, I enjoy ___. (45:00)
I think the coolest thing in technology today is ___. (45:30)
I look forward to the day when I can use technology to ___. (46:25)
Something interesting or different about yourself (47:00)
How Not to Die (48:00)
Where can people learn more about Peter? (49:00)
Book reading / listening recommendations? (49:00)
The Mind's I (50:00)
Peter's articles on Medium (52:00)
Get a free audio book and support DataDriven - visit thedatadrivenbook.com! (00:00)
Transcript
The following transcript is AI generated.
00:00:01 BAILeY
Hello and welcome to data driven.
00:00:03 BAILeY
The podcast where we explore the emerging fields of data science, machine learning, and artificial intelligence.
00:00:11 BAILeY
In this episode, Frank and Andy speak with Peter Voss, peterboat.
00:00:15 BAILeY
Peter Voss is the world's foremost authority, an artificial general intelligence or AGI.
00:00:21 BAILeY
In fact, he is the one who coined the term in 2001 and published a book on the topic in 2002.
00:00:28 BAILeY
He is a serial.
00:00:29 BAILeY
AI entrepreneur technology innovator who has for the past 20 years, then dedicated to advancing artificial general intelligence.
00:00:38 BAILeY
Today he is focused on his company, IGO, which is developing and selling increasingly advanced AGI systems for large enterprise customers.
00:00:47 BAILeY
Peter also has a keen interest in the interrelationship between philosophy, psychology, ethics, futurism and computer science.
00:00:56 BAILeY
I think you will find this interview a fascinating look at the future of AI.
00:01:01 BAILeY
Now on with the show.
00:01:05 Frank
Hello and welcome to data driven, the podcast where we explore the emerging fields of data science, machine learning and artificial intelligence.
00:01:13 Frank
If you like to think of data as the new oil, then you can think of us like well.
00:01:18 Frank
Car Talk because we focus on where the rubber meets the road and with me on this epic virtual road trip down the information highway because we're still locked in quarantine.
00:01:29 Frank
As always, Andy Leonard.
00:01:30 Frank
How's it going and?
00:01:31 Andy
Good Frank, how are you?
00:01:33 Frank
I'm doing well.
00:01:34 Frank
We had a bit of snow.
00:01:36 Frank
We're recording this on Monday, February 1st and the East Coast has been blanketed in some snow.
00:01:37 Peter Voss
Yes.
00:01:45 Andy
Yeah, we got more than we've gotten, probably since 2018 or so. About four inches here in FarmVille and then almost an inch of ice on top of that, which always makes it fun, right?
00:01:58 Frank
Yeah, the ice is worse than the snow on.
00:02:00 Frank
Basically so I went out, walk the dog today and one of the dogs and it was crunch, crunch, crunch.
00:02:06 Frank
So there's a nice layer of ice over everything which is going to make driving later fun, but I do have.
00:02:13 Frank
I do have the an all wheel drive car which is fantastic.
00:02:17 Frank
I will never not own one of those again.
00:02:19 Andy
Nice.
00:02:21 Frank
Yeah, you've seen it's the CRV.
00:02:23 Andy
Yes, yeah, it's nice you did well.
00:02:26 Frank
I dubbed it the Rocinante.
00:02:31 Andy
In case our listeners are not familiar with that, with what Frank is referring to, it is not the old novel.
00:02:40 Andy
Frank is not tilting at windmills instead.
00:02:44 Andy
And if I got that reference wrong, correct me.
00:02:46 Andy
I'll just edit that out.
00:02:47 Frank
Oh, you are right, it's from this AM Oh my God, I forgot new book on Cody.
00:02:48 Andy
Not sure.
00:02:51 Andy
Donkey Quixoti wasn't.
00:02:53 Frank
Yeah yeah Cervantes I was gonna say from Cervantes book and I'm like oh what was the name of that?
00:02:53 Andy
Yeah so.
00:02:59 Frank
Which is the opposite of how most people think, but that's what I do.
00:02:59 Frank
OK, good.
00:03:02 Andy
There we go, but it is actually a reference to both the books and a series, The expanse of which Frank and I are great fans, so.
00:03:12 Frank
Awesome, but you know who's not covered in snow today.
00:03:13 Andy
I like it.
00:03:15 Andy
Who is not covered in snow their guest.
00:03:16 Andy
Our guest.
00:03:18 Frank
Who lives in?
00:03:18 Frank
Yeah.
00:03:20 Frank
I'm assuming sunny or Smokey I guess depending on the time of year California Peter Voss Peter welcome to the show.
00:03:29 Peter Voss
Thank you, yes, it's we've got snow on the mountains here, but it's very sunny.
00:03:36 Peter Voss
It's it's nice and we have a lot less smog these days.
00:03:41 Andy
Very good.
00:03:41 Frank
Nice so you are the.
00:03:46 Frank
One of the world's, or if not the world's foremost authority in AGI or artificially artificial general intelligence, and I believe you are the one that coined the term.
00:03:58 Peter Voss
Yes, correct and 2001 myself and two other people. We coined the term artificial general intelligence AGI to really distinguish the kind of work we were doing from, you know, specialized narrow AI which is.
00:04:18 Peter Voss
Pretty much what everybody else is doing.
00:04:20 Peter Voss
The original dream of artificial intelligence was of course, to have systems that can think and learn the way humans do, but that turned out to be a lot lot harder than people thought.
00:04:31 Peter Voss
So over the years, AI really turned into narrow AI using human ingenuity to figure out how to solve one particular problem, like playing chess or.
00:04:41 Peter Voss
Container optimization or medical diagnosis and then to write a program or to train data to do that to solve that particular problem.
00:04:51 Peter Voss
But it's really the external intelligence of the program or the data scientists that is then encoded.
00:04:58 Peter Voss
To solve that problem, whereas we wanted to get back to the original dream of having a thinking machine that it can figure out how to do these things and and learn more humans do so.
00:05:09 Peter Voss
That's why we felt we had to.
00:05:12 Peter Voss
You know, coin a separate term to distinguish it from narrow AI.
00:05:16 Frank
Interesting.
00:05:18 Frank
So for years, AGI has been.
00:05:21 Frank
Kind of thought the stuff of science fiction.
00:05:24 Frank
I think there was a lot of optimistic people like you said that thought we would have it by now.
00:05:29 Frank
I know this is kind of a loaded question, but one do you think we'll ever get there and two, what's the sort of time frame we're looking at?
00:05:38 Peter Voss
Yes, it's an interesting question, so absolutely, I believe it's it's.
00:05:42 Peter Voss
Possible, and in fact the reason we got together. We wrote a book called Artificial General Intelligence. As I said in 2001 was because we believe the time is ripe to get back to this original dream that the technology had advanced enough. Both hardware and software technology and cognitive psychology. Cognitive science.
00:06:02 Peter Voss
That we now understood enough and had fundamentally had the tools in place to tackle this problem and to say.
00:06:11 Peter Voss
So I I absolutely believe that it can be solved soon, and in fact we will leave.
00:06:18 Peter Voss
We are on on the way of solving this problem now in terms of time frame.
00:06:24 Peter Voss
Normally the way I answer this question is I don't measure it in time.
00:06:28 Peter Voss
I measured in dollars.
00:06:31 Frank
I like that time is money, so I guess.
00:06:34 Frank
That's a reasonable correlation.
00:06:35 Peter Voss
Yeah, and and the reason I do, I say that is because.
00:06:39 Peter Voss
Still, today almost nobody is working on AGI. You know, 99% of all the effort in artificial intelligence is still on narrow AI, so if this continues, it will take a long long time for us to reach human level AGI. But if that changes.
00:07:00 Peter Voss
And you know the kind of funding that's going into deep learning machine learning suddenly was applied to AGI.
00:07:06 Peter Voss
Then I think it could easily happen at less than 10.
00:07:09 BAILeY
Yes.
00:07:10 Frank
Oh wow.
00:07:11 Andy
Very cool, so I'm curious is there any like lead in does?
00:07:16 Andy
Does time and money invested in deep learning and narrow AI?
00:07:23 Andy
Does any of that help move the cost?
00:07:25 Andy
Say further the cause for AGI?
00:07:29 Peter Voss
Slightly, I believe, you know.
00:07:32 Peter Voss
Obviously, any advances in languages and data collection in hardware development and the general experience.
00:07:42 Peter Voss
In that sense, it does help it.
00:07:44 Peter Voss
But in another sense, it's actually the opposite.
00:07:46 Peter Voss
It's actually hindering it because a whole generation of software engineers and data scientists are now coming into the field, believing that deep learning machine learning is a way to do it.
00:08:00 Peter Voss
And all we need is more data, more horsepower and will solve this problem.
00:08:05 Peter Voss
And that's I think barking up the wrong tree, and it's a it's a dead end.
00:08:10 Peter Voss
So in that sense, what's happening today with deep learning?
00:08:12 Peter Voss
Machine learning is actually counter to achieving.
00:08:16 Andy
GI interesting very interesting.
00:08:20 Frank
Was it always that way or it's just the way the market kind of went frenzied over just narrowed AI?
00:08:26 Peter Voss
Why?
00:08:26 Peter Voss
Well, we've had several windows of AI.
00:08:30 Peter Voss
You know the the disappointments over the decades.
00:08:33 Peter Voss
You know, when we had expert systems, people believe that you know they would really, you know, show real intelligence and then it kind of fizzled out.
00:08:42 Peter Voss
And so we've had.
00:08:43 Peter Voss
We've had various windows, and but of course, deep learning machine learning has been so spectacularly successful in several areas.
00:08:52 Peter Voss
You know, image recognition, improving speech recognition, and you know various other fields that just, you know, it's the only game in town as it has been very, very successful.
00:09:04 Peter Voss
But people are also starting to realize what the limitations are of it.
00:09:11 Peter Voss
So yeah, it's it's kind of at the moment.
00:09:14 Peter Voss
The only game in town, and it has really been successful in many.
00:09:17 Andy
Areas, So what are those limitations?
00:09:20 Andy
And how does AGI addressing?
00:09:23 Peter Voss
Yeah, so fundamentally when you think about intelligence, you know if you think about just common sense.
00:09:32 Peter Voss
If we talk to a person and we judge them to be intelligent or to be totally non intelligent, the kind of things we expect is that they can learn.
00:09:43 Peter Voss
Immediately that when you say something a, they understand what you're saying and they integrate that knowledge with their existing knowledge so you know if you say my sister's moving through Seattle next week or something.
00:10:01 Peter Voss
That knowledge needs to fit in somewhere.
00:10:04 Peter Voss
You know you know the person who's talking.
00:10:06 Peter Voss
You may know who the sister is, or you may not know who the sister is.
00:10:10 Peter Voss
You probably know what Seattle is.
00:10:13 Peter Voss
You may have images of, you know, rain pouring down all the time or whatever, but so you integrate that knowledge.
00:10:21 Peter Voss
And if you're not cleared, my maybe the person has two sisters, so then you would ask her, do you mean your older sister you know your younger sister?
00:10:30 Peter Voss
And so we expect an intelligent human to basically do.
00:10:35 Peter Voss
You know what's technically called one shot?
00:10:37 Peter Voss
Learning?
00:10:38 Peter Voss
You hear something once you see an image.
00:10:40 Peter Voss
Once you learn that and you integrate it into your existing knowledge base.
00:10:46 Peter Voss
And if you're not sure how to interpret it.
00:10:49 Peter Voss
Then you ask clarifying.
00:10:50 Peter Voss
Questions until you know what it what it is.
00:10:54 Peter Voss
So you have deep understanding you have disambiguation.
00:10:59 Peter Voss
You have learning instant learning, one shot learning.
00:11:03 Peter Voss
You have long term memory.
00:11:05 Peter Voss
You remember that next week you you know if you paid attention, you will remember that and you have reasoning about.
00:11:12 Peter Voss
30 now deep learning machine learning as it's done today, really doesn't offer any of those.
00:11:20 Peter Voss
So if you if you had a human and you told them something and they didn't remember it, they didn't understand that they didn't ask for clarification.
00:11:27 Peter Voss
You wouldn't think of them as being very intelligent, would you?
00:11:33 Frank
No, I mean, my kids are smart, but when I tell them to bring the trash cans back from the street, they'll conveniently forget.
00:11:39 Frank
But I, I think I know where you're going with that, yes?
00:11:42 BAILeY
All right?
00:11:44 Frank
But the question I have, it sounds like you're trying to and I know this is going to be not really good analogy.
00:11:50 Frank
Or maybe it is you're trying to code for curiosity.
00:11:54 Peter Voss
That's very much part of it, but you know even deeper is understanding.
00:11:59 Peter Voss
Basically, when you have some input, do you?
00:12:02 Peter Voss
Do you understand you know what the implications are, how it fits in with the rest of the knowledge that you have?
00:12:08 Peter Voss
And you know, even that, that's sort of more even more fundamental than curiosity.
00:12:13 Peter Voss
But yeah, curiosity is then wanting to gather more information, so this is inherently an interactive process.
00:12:22 Peter Voss
You know, an intelligent person would ask follow up questions you know they would want to kind of.
00:12:29 Peter Voss
Fill in the pieces of the puzzle and you know that they can be more.
00:12:33 Peter Voss
In fact effective in their communication on their or their job.
00:12:37 Frank
Right so.
00:12:37 Peter Voss
So yes, that's definitely part of it.
00:12:40 Frank
So calling back to your example of someone's sister moving to Seattle you you would ask, you know, I didn't know you had a sister or how many sisters do you have or how many siblings do you have and.
00:12:51 Frank
Where is she moving to?
00:12:52 Frank
Why?
00:12:53 Frank
I guess that's kind of.
00:12:55 Frank
I guess it's all about building that knowledge map inside.
00:12:58 Frank
Your head or then your head being could be a program I guess.
00:12:58 BAILeY
Exactly.
00:12:59 BAILeY
OK.
00:13:02 Peter Voss
Yeah, and deep learning machine learning really doesn't allow for that at all.
00:13:07 Peter Voss
You know you accumulate masses of data and you train a model, but that model is then static.
00:13:14 Peter Voss
It's a read only model.
00:13:15 Peter Voss
You know, it doesn't dynamically learn, so it may have a sort of a knowledge graph, but even that knowledge graph is.
00:13:23 Peter Voss
Is very opaque, it's.
00:13:26 Peter Voss
Yeah, it's not scrutable you know and and this is this is such a big problem with deep learning machine learning that you don't know why it gives a certain response, which is a huge...