Latest / The DEX Show: A Show for IT Change Makers / Reality Bytes #47: Eliza to Neural Networks: The AI Workplace Journey

Transcript

Reality Bytes #47: Eliza to Neural Networks: The AI Workplace Journey · The DEX Show: A Show for IT Change Makers

0:01
Welcome Item change Makers to the Deck Show with Tim Flower and Tom McGrath.
0:07
Let's get into it. Hello everybody, Welcome back to Reality Bites. We
0:12
have a pretty full a pretty full team. In fact, I think we
0:16
have minus Tim, we have the entire Reality Bites panels. So we've got
0:20
Dina, Oriana, Sean and Megan and as listeners will have noticed, we
0:26
are really delving deep. We are preoccupied with AI and the implications of AI
0:31
for decks and for the future of decks. And we've been having across the
0:35
Deck Show a lot of fun with this topic, all really enjoying it.
0:40
And then our very own Meghan Break She recently stepped up this week and contributed
0:46
a quite brilliant blog to the next Link blog and the link will be in
0:50
the description as ever, and it's called the History of AI in the Workplace,
0:56
Megan, and it reminded me a little bit, Meghan of a There's
1:00
one of my one of my favorite writers about tech is a guy called Charles
1:06
Petzold. Have you ever read him? Not that I know of, but
1:07
maybe he's called Code The Hidden Language of Computer Hardware and Software, where he
1:11
kind of walks you through the development of computers from the very very first kind
1:19
of mechanical adding machines. And he does it so incrementally and slowly and clearly
1:23
that you you you could almost convince yourself that you understand the technologies you move
1:29
through A but wonderful writer. We're actually going to reach out to him see
1:32
if he'd like to come on the show. But this remembrance, Megan,
1:34
was prompted by your excellent history of AI in the workplace. So you go
1:41
back in it right to the very beginning of the earliest origins of AI,
1:46
and I thought this would be a great place to sort of tee up the
1:48
wider discussion Megan, tell us about tell us about those origins and don't don't
1:53
don't neglect Eliza for first interactive who certain I certainly will. Yeah. My
2:00
first draft of that article was, I would say, twice the length of
2:02
what is currently published, and went all the way back to like the invention
2:06
of statistics and probability models. Yeah, because that's you know, like the
2:10
AI that we see today with machine learning. It basically has its roots in
2:15
that mathematics. But yeah, so the development of AI, I mean it
2:20
was first it was in like sci fi novels, right, like the idea
2:22
of computers that could make decisions like humans was this like sort of intriguing idea.
2:27
And then you have famously Alan Turing. He makes the Turing machine,
2:30
which was I believe more theoretical than actually practical, and the Turing test,
2:36
which is that test can you tell the difference if it's a machine or a
2:39
human, intriguing in controversial tests to test if AI is really effective. And
2:46
then you had this scientist McCarthy who there was this Dartmouth conference in the fifties
2:53
I believe, where he brought together kind of all the best sort of computer
2:55
researchers working at the time and said, essentially, let's try to see if
3:00
we can make computers make decisions like humans. And they those sort of first
3:07
approaches used a programming language that he had come up with called Lisp programming,
3:10
and it was a rules based approach. So if this condition, then this
3:15
action, and trying to create computer programs and computer algorithms that could sort of
3:22
simulate human decision making with if theyn so could you feed it enough conditions that
3:27
the computer could make a decision and you would think it was a human being
3:30
making that decision. They were limited by basically the computational power and the memory
3:35
available in computers in the fifties and sixties because back then computers took up an
3:38
entire room, so you know, they didn't really they weren't really successful,
3:43
but they were trying to create computers that could make medical diagnoses because they figured,
3:47
you know, doctors sometimes get it wrong or it takes too long to
3:51
make a medical diagnosis, and people suffer. So they kind of were trying,
3:54
I think, to eliminate some human suffering by make computers they could make
4:00
medical diagnoses. So that was kind of the fifties, and then in the
4:03
sixties you have this release of Eliza, which is this really intriguing computer program
4:08
which I found out yesterday as I was reading more about It was actually made
4:12
by a researcher at MIT who was trying to show how basically simple human and
4:16
computer interaction was, like how superficial. But he made this program that was
4:23
modeled after a Rugerian psychotherapist, which is that kind of psychotherapy where you say,
4:27
like, I'm feeling sad and they're like, why do you think you're
4:30
feeling sad? And so it's just kind of they say what you say back
4:32
to you. And there's still a version of it. You can find it
4:35
online. It's hosted by I think New Jersey Institute of Technology. So I
4:40
opened it up this morning. It pops up and says, hello, I'm
4:43
Eliza. I'll be your therapist today. And I said hi. Yeah,
4:46
I said hi, I'm feeling tired today, and it said, did you
4:48
come to me because you're feeling tired today? And I said, no,
4:53
I feel happy, and it said, tell me more about such feelings.
4:57
So it's very simple, right, It just kind of spits back kind of
5:00
exactly what you've said, or like a rote response. But apparently in the
5:02
sixties they released it to some people and there were people who thought they were
5:05
speaking to an actual human being. So, you know, at least back
5:09
then, I think today, like with our experience with computers, we would
5:13
obviously see through it. But at the time, I mean, it fooled
5:16
people. Some people thought they were talking to a human and actually had like
5:18
an emotional response to the conversations with Eliza. So I thought that was a
5:23
really interesting, you know, sort of first entry point to the development of
5:27
AI, that that was how kind of computers were being used so it didn't
5:31
really enter the workplace in a serious way till the eighties. And I mean
5:36
I could keep talking. I'm basically reciting. You're ahead of my next question,
5:42
Magan, which is because I think you do hone in very nicely in
5:46
the article of the fact that those origins which you just described are explicitly benign,
5:51
shall we say, right, that both developed in order to help people,
5:57
And I think that kind of feeds into your overall kind of conception of
6:00
what AI is, or at least should be. And we'll come back to
6:04
that. But but now, yeah, the entry into the workplace is precisely
6:08
where I hoped you go next anyway. Yeah, so yeah, So basically
6:14
they continued research in that rules based system that I was talking about earlier,
6:17
right the list programming that if this then that there was another one in Europe
6:20
that was if certain conditions, then certain conclusions. But it was still sort
6:27
of if then logic. And you know, after the whole medical diagnosis thing
6:32
didn't really take off in the fifties and sixties, a lot of funding disappeared,
6:39
but there was the invention of what they called an expert system, where
6:43
you basically fed a bunch of knowledge into these computers. And in the eighties
6:48
you had the release of the first PCs I think right like the first like
6:53
DOS computers, and so that kind of revolutionized computing in a huge way,
6:58
right in a like in a huge jo joo joos way. All of a
7:00
sudden, computers didn't take up a whole room. Everybody could have one.
7:04
And what that also meant was you didn't just have one central IT team who
7:09
controlled the mainframe, who were the only ones who could build applications and build
7:15
ways to use computers. Now you have all these individual computers, and it
7:17
like we talk about today with decks, it kind of democratized building out computer
7:23
programs and applications. And so you had these expert systems which were these sort
7:29
of if then logic that could help I believe, build these applications and could
7:34
kind of like know things. And these started to get used in the workplace
7:39
I think by computer programmers to help them build different computer kind of applications.
7:45
But they had some limitations. I think full scale automation was the goal with
7:48
an expert system that you could just tell it to do something and it would
7:51
do it, and it couldn't really achieve that and some of the articles I
7:56
read suggested that it teams who were used to having centralized control of everything.
8:01
We're a little peeved, should we say that suddenly like everybody felt like it
8:05
they had a right to engineer the workplace computers. So, you know,
8:09
the expert systems did fall out of use sort of towards the mid nineties.
8:13
I think the last sort of surge of them is that famous IBM computer I
8:20
forget its name right now, deep Blue or something that beat a chess player
8:24
in a chess game. So that was that was sort of the last like
8:28
famous expert system, because it knew if someone does this play, I do
8:33
a different play, and it was able to be a chess master in a
8:35
timed game. But that's kind of the last example of an expert system,
8:39
and then they sort of fell out of use, and intriguingly, like that
8:43
rules based technology that basically half a century about AI research have been dedicated to
8:52
isn't really the technology that underpins the generative AI, the predictive analytics everything we're
8:56
using today. That the whole AI revolution that we're experiencing now is based on
9:01
a different technology that was also invented in the fifties, but just was like
9:07
disregarded by researchers for like half a century until they reconceptualized it. Yeah.
9:13
Yeah, the neural networks they're called. Pedro talked about it when he was
9:16
on the podcast. He talked about actually how he noted like, oh,
9:20
we could have invested more in neural networks earlier. Yeah, it was that
9:26
comment that he made that made me ask. I started by research using chat
9:30
GPT for this article, and I said to chat gpt, like, can
9:33
you tell me the history of neural networks because Pedro had mentioned them, And
9:39
yeah, they were first invented in like the four conceptualized basically in the forties,
9:43
and then there was something called the perceptron that was invented in the fifties.
9:48
And I'm not going to pretend that I have a complete comprehensive understanding of
9:52
how neural networks work, because I don't. I think you have to get
9:54
like a PhD to do that. But they basically have like weighted neurons.
10:01
It's meant to kind of model I think a neural network like what humans have.
10:05
Yeah, and so they were sort of invented back then, but I
10:09
think we're not very sophisticated and so we're not the emphasis of AI research.
10:13
But then it was reconceptualized in the late nineties is machine learning, and that
10:16
has exploded into what we see today with the computational power and everything, the
10:22
incredible data lakes we have today. Like that technology just took off. So
10:26
I did think that was interesting that you had sort of fifty years of research
10:28
into one topic that went away and then it was something else from the beginning
10:33
that ended up spurring what we see today. I thought that was interesting one,
10:37
and we'll come back to you, but before we do, let's hear
10:41
a bit about neural networks from our very own six PhDs. Dina no PhDs
10:48
yet, so let's just keep it as simple as possible. The neural networks.
10:52
Imagine your brain has tiny helpers called neurons, and they pass secret notes
10:58
to each other. And when you see or when they see a picture of
11:01
a dog, some of the neurons get super excited and say, oh,
11:05
hey, it's a dog. So the neurals talk to each other and then
11:09
they like special They have their own special paths that are called connections, and
11:16
they start to have built to start to build actually connection next to a connection
11:20
and they start to build a neural network, which is like a team of
11:24
these magical neurons working together. So they solve a puzzle. It becomes bigger
11:28
and bigger and bigger, recognizing shapes, recognizing examples or learning from examples,
11:33
and once they do recognize a shape, they go to the bigger one,
11:37
build more neural networks, build more connections, so over time they become super
11:43
good at things like recognizing dogs, cats, even faces and photos. So
11:48
how does the neural network work or it actually transform the AI by combining all
11:54
of the historical context and the great stuff that we've heard before into bigger mathematical
12:01
insights, into breakthroughs and shaping the field as we know it today. So
12:05
all of this is very nice theoretical conversation. What does that mean to a
12:11
day to day human being now? Is example, that's what you use in
12:13
the facial recognition, that's the logic and math behind it, that's the science
12:18
behind it, Social media, healthcare, natural language processing or NLPs, and
12:24
all of those things are the core of building AI. So whether we used
12:28
any form of AI in the background to actually ask us or ask we get
12:33
those answers to start where we go, this is how in very simple format,
12:37
which I hope was clear to actually show you how we can build on
12:41
top of it, and all of the great inventions and progression of AI is
12:46
based off of that in simple form. It's a wonderful definition, and I
12:52
think I followed it. I do have I have a question for you,
12:54
data of interest. When it comes to distinguishing between digital technology per SE and
13:01
AI, how clearly do you feel able to make that distinction. So back
13:09
to the example Megan was speaking about earlier, with what back back in the
13:15
days, right now, I think the most distinctive part is going to be
13:18
the emotional part. Right, we can go and play around that emotional part.
13:24
This is something that AIS or Yeah, So for AI's they cannot distinguish
13:28
much. However, it has to be in something that they have pre programmed
13:33
to do or patterns that they have seen. So once you start to teach
13:35
them how to identify human behavior or how to identify emotions and explain them,
13:41
which is going to be amazing if we can do that, because it's very
13:46
very simple in its core. However, human or feelings are very hard to
13:52
explain and pinpoint because they don't even have a regular pattern. It could be
13:58
all explorative, it could be all guessing game kind of thing. So as
14:01
long as you're keeping your questions or the asks that we try and get to
14:07
AI into a very mathematical or theoretical analytical format, that's super You're very easy
14:15
to identify that. But when you start to navigate into the feelings aspect of
14:20
things that what makes humans humans, it's not just the decision making, it's
14:24
going to be a little bit harder to identify. However, I know AIS
14:28
now are learning better from the pattern of things, and we still if we
14:33
play on that silver lining for now, I think we're still ahead of the
14:37
game. We're able to easily identify when it's confusing when it comes to feelings
14:41
or how do you feel about that? If you're tired and then sad and
14:45
then hungry, how would that make you feel, etcetera, etcetera, or
14:48
react into a certain situation. So putting feelings in the mix, I think
14:52
would impact the answer. So if we try and do that and we don't
14:56
get the right answer, then we can safely say the the machine was behind
15:01
that, not another human being. Yeah, I love it. It's funny
15:03
how our discussions of AI so quickly kind of find themselves centered upon what makes
15:09
us human, right, which I think we'll come back to Megan's articles central
15:13
thesis, but before we do that, Sean, you're somebody who's written quite
15:18
a fair bit across the Decks Hub on the kind of marriage of AI and
15:22
DEX. Listening to what you've heard today, what stands out to you most
15:28
about that potentiality in the light of his history. Sure, so, I
15:35
think the potentiality of its probably best express with what we already have is one
15:41
of the features in the Next Thing product for Next Thing Assist. So that's
15:46
taking our own proprietary AI and chat gipt kind of like chatchept but on steroids.
15:52
And it works really well if you're an IT manager, if you're on
15:56
that side of things, just because you're able to use one single interface,
16:00
ask questions, you know, for example, how's team's going, and it
16:04
will pull up you know, fifty percent of devices or experiences issue by the
16:11
way, you might want to check out X, Y and z uh and
16:14
then based off of that you can actually take actions immediately within Assist. So
16:18
that is something that's very promising, and you have to be excited about if
16:22
you work in an IT. I do think I was thinking about this earlier.
16:29
Do you guys remember like Microsoft Clippy. Do you remember like the paper
16:33
clip thing? Yeah, so it was with popping eyes. Yeah, it
16:38
was very annoying and it was only around for a few years and then they
16:41
pulled it. So that's kind of like on the other side of the divides
16:45
for US for end users, for employees, and that clearly you know,
16:48
it wasn't technically an AI feature, but it was trying to like prescribe or
16:52
figure out what you were doing. Obviously had very limited technology, so that
16:56
would be the extreme comparison of something like assist for US. I haven't seen
17:03
anything yet that's sort of helping you out on screen. Now. It probably
17:08
exists somewhere, But I also think it brings up a lot of questions as
17:11
if you even want that. So it's sort of like having a colleague in
17:15
the room with you helping you out for certain projects. But do we really
17:19
want somebody kind of over our shoulder at all times suggesting things? And like
17:25
Dina was saying, this also gets into like the emotional side of work.
17:30
Right, You're on a team meeting and some people, you know, maybe
17:33
inspire you, some people bore you. So it's like, do you want
17:36
to add more colleagues to the mix when you're trying to focus, So it
17:40
has to be it's a very My point is simply really depends on the future
17:44
for end user technology when I'm viewing it for my side, but for it
17:48
managers, it seems to be quite promising. I'm shocked sure on that any
17:52
of our colleagues bol you absolutely none theoretically in theory exactly could be. It
18:00
could be that you're on a team call excellent. So I think you know,
18:04
Deck reintroduces for human into this picture pretty pretty forthrightly, Megum. But
18:11
you make the argument, as I've mentioned, that the history of AI has
18:15
been specifically human from those early psychological and medical applications or attempted applications, and
18:22
the AI do I have this right, it's like an intrinsically human centric technology
18:26
of that it should be? Yeah, I think should be is probably a
18:30
better take. I was struck by those first applications, those first sort of
18:36
thought experiments were in pursuit of resolving some kind of human challenge. And I
18:45
do think that the elements of AI that are the most valued and the ones
18:52
that I you know, a lot of conversations going around about AI include anxiety
18:57
around what's this going to do to our careers, what's this going to do
19:02
to our society, you know, up all the way up to you know,
19:06
your two thousand and one. Sorry, I can't do that day of
19:07
like leaving the man out in space, right, So, you know,
19:11
you do have that sort of anxiety around AI. But I think where it
19:17
where it brings the most value is where it helps us solve our human challenges.
19:21
Right where it helps us either make a medical diagnosis faster or in the
19:26
case of you know, digital playe experience, make employees feel a little bit
19:32
more relaxed at work and less stressed out about their machines always causing problems.
19:36
So my sort of thesis for the for my article is, yeah, that
19:41
AI is is at its heart a human endeavor, and that it is the
19:47
most successful when applied that way, and that it makes it a nice pairing
19:52
with the DEX field because the DEX is also at its core a human approach,
19:57
a human centric approach to it, and so AI when applied to this
20:02
field, you know, I hope will expand rapidly because they're so well paired.
20:07
I think that's a really good point. And I thinking about it as
20:08
well. It's an unusually significant it's a hugely significant technology with unusual origins excuse
20:17
me, with unusual origins that don't come from war effectively, right, Like
20:22
most of the most significant technologies on that level come comes straight out of conflict,
20:30
and it does seem to have striking the benign origins. But Oriana,
20:36
listening to Megan's argument there, how would you qualify or contest this perspective or
20:41
how would you what guidance would you say should be paramount as we begin to
20:48
integrate AI into our lives to this extent in terms of ensuring that it does
20:53
remain human centric. Maybe that's a better way of putting it, rather than
20:56
pitting you against Megan ins to it. Yeah, I mean, I think
21:03
fundamentally Megan and I tend to agree on this idea, right that there's certainly
21:10
the ideal that we see and then there's possibly the real practice. And I
21:17
think it's really interesting that Megan brings up healthcare. You bring up or because
21:22
a lot of this research and development has been funded by the military, right
21:29
our military that's where so much of our scientific research funding comes from, even
21:34
for things like medical technology, and that can be a really great thing,
21:37
but it's also you know, there's there's sort of the marketing perspective of how
21:41
do we talk about this in a way that's palatable to people, And then
21:45
there's those applications absolutely exist in our being developed at the same time, right,
21:52
those are part and parcel of that research, and they always have been
21:57
if you look into you know, all of those labs for example. So
22:04
I think, yeah, the question of what makes something human centric is really
22:08
one that's super relevant to me because that's you know, what I studied and
22:11
what I tried to focus on, and starting with just the medical example,
22:18
trying to solve a human problem doesn't make the solution itself, in my mind
22:23
human centric. I think what makes technology human centric is really the idea that
22:29
it is, as Don Norman, one of the founders of my field,
22:33
says, life enhancing and positive. And obviously that's you know, very nebulous
22:38
in and of itself. So I think we have to ask ourselves, you
22:41
know, does it actually have that effect for people as a whole in the
22:48
long term, because I think it's really easy to take sort of an individual
22:51
perspective and say one person has this positive interaction, therefore the system is good,
22:57
rather than it is possible to have a positive interaction, the system is
23:02
not necessarily life enhancing. And I think there's a lot of mythologization of technology.
23:11
I need to find the source for this, but there's a sort of
23:15
some writing on this concept of like we create mythology around technology. Right.
23:19
It's very easy for us to build stories around things, to personify things.
23:23
This is such a human impulse, right, or we're storytellers, we personify
23:29
things. We have to pomorphize things. Right. If you see a dog
23:32
and you're like, oh, it's smiling, Well is it smiling or is
23:34
it bearing its deep? That means something really different for a dog or for
23:37
a human. It's so easy for us to impose our emotions onto things around
23:44
us, including technology. Right. So this is one of the issues that
23:48
comes up with things like an AI therapist, right, is that it can
23:52
be built in a way that is very deeply fallawed and potentially even harmful to
23:59
hear humans, and we will still potentially relate to it on this very profound
24:04
emotional level. So I think there's Yeah. I think there's a few specific
24:08
factors that are important for like an ethical AI system. Right. Some of
24:15
the ones that people experts in the field, not me, have brought up
24:18
are things like, because you can't reverse engineer the decisions, you can't see
24:23
what factors are going into making a decision. And this is really important in
24:29
the healthcare example, because it's not just that it's being used to make diagnoses.
24:33
It's actually much more frequently used in things like health insurance denials, and
24:37
there the system is engineered to reduce the amount of spend for the company,
24:42
right, it's increasing efficiency to reduce cost that is not human centric. One
24:48
more thing which I think is really really important to talk about when it comes
24:51
to human applications of II, which is that there are mountains of evidence at
24:56
every stage of research. We have seen that there's a ton of buyer in
25:00
the training data and that that creates really bad results. And in medicine there's
25:06
a ton of research on the bias in medicine and all of that becomes formalized
25:14
and then obscured when you put it into an AI, and it gives people
25:19
in some ways plausible deniability for making decisions that are potentially very very harmful and
25:27
that we have seen in the research to be very harmful. Fascinating or on
25:34
it, Megan, anything you want to add to that and have the last
25:37
word, only that Oreana always elevates my thinking whatever they need to share on
25:44
this podcasts, and I really appreciate that commentary. And all I will add
25:48
is that that bias element of it I think is so important. I mean,
25:55
everything you said obviously is incredibly important to factor in. I definitely was
25:57
looking at it with rose colored gloves and sort of an optimistic take. And
26:03
I do want to just flag I one interesting piece of my research that didn't
26:07
make it into the article is actually that the Eliza bot was sort of one
26:11
of the earlier times that AI went out into the world, like where human
26:18
beings not part of the research were really interacting with it, and it raised
26:21
a lot of ethical questions about bias and AI that I think Oreana mentioned we're
26:26
still grappling with today because Eliza was a gender chatbot, so it was perceived
26:32
to be a woman, and that created a lot of biases in the way
26:37
people interacted with it and the way people considered it and its expertise and that
26:41
kind of it was sort of the first time people started asking how does human
26:45
bias apply in the development of a machine? And I think that that's a
26:49
question that we really still haven't solved for and in fact sort of forged ahead
26:53
without considering. In the example at Oreana gave of the data that you trained
26:59
current model and the bias that gets sort of weoven into that. So I
27:03
do think there's a lot for us to overcome. I think the ideal of
27:07
AI that is human centric is something I hope we can strive for. But
27:11
yeah, I think when it becomes profit centered and the human rights in the
27:18
human life is forgotten is when we go astraya and we head more towards that
27:22
dystopia that sci fi writers have been warning us about for decades. Fantastic and
27:30
fascinating. And you can read Megan's article on the next Think blog, not
27:34
on the Hub, there'll be a link to that blow And if you want
27:40
to explore the implications of AI for our industry, for the deck industry,
27:45
for it in general, you would do very well to book your place now
27:48
at Experience twenty four. It's in Boston on the fifteenth to the sixteenth of
27:53
October. It's in London to the fifth and sixth of November, so hopefully
27:57
there's some way of getting to one of those two locations if you can.
28:00
It's absolutely free and you'll see some of some representatives of the world of world
28:07
leading digital workplaces and many many other things besides. And there will guarantee be
28:14
some very very in depth and intriguing explorations and demonstrations of the power of AI
28:19
today in dex so please do check that out. You can also find those
28:25
links in the show notes. Just remains for me to thank everybody, particularly
28:30
you Meghan for writing and bringing us that brilliant article, and I'll see you
28:33
all very soon. Bye. To make sure that you never miss an episode,
28:37
subscribe to the show in Apple Podcasts, Spotify or your favorite podcast player,
28:42
and if you're listening on Apple Podcasts, make sure to leave a rating
28:45
of the show. Just have the number of stars you think the podcast deserves.
28:49
If you'd like to learn more about how Next Thing can help me improve
28:52
your digital employee experience, head over to next Think dot com. Thank you
28:56
so much for listening. Until next time, but Dott