July 17, 2024 • 40 mins
Major breakthroughs in artificial intelligence research often reshape the design and utility of Al in both business and society. In this special rebroadcast episode of Smart Talks with IBM, Malcolm Gladwell and Jacob Goldstein explore the conceptual underpinnings of modern Al with Dr. David Cox, VP of Al models at IBM Research. They talk foundation models, self-supervised machine learning, and the practical applications of Al and data platforms like watsonx in business and technology.
When we first aired this episode last year, the concept of foundation models was just beginning to capture our attention. Since then, this technology has evolved and redefined the boundaries of what's possible. Businesses are becoming more savvy about selecting the right models and understanding how they can drive revenue and efficiency.
This is a paid advertisement from IBM. The conversations on this podcast don't necessarily represent IBM's positions, strategies or opinions.
Visit us at ibm.com/smarttalks
See omnystudio.com/listener for privacy information.
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Speaker 1 (00:00):
Hey everyone, it's Robert and Joe here. Today we've got
something a little bit different to share with you. It
is a new season of the Smart Talks with IBM
podcast series.
Speaker 2 (00:09):
Today we are witnessed to one of those rare moments
in history, the rise of an innovative technology with the
potential to radically transform business and society forever. The technology,
of course, is artificial intelligence, and it's the central focus
for this new season of Smart Talks with IBM.
Speaker 1 (00:25):
Join hosts from your favorite Pushkin podcasts as they talk
with industry experts and leaders to explore how businesses can
integrate AI into their workflows and help drive real change
in this new era of AI. And of course, host
Malcolm Gladwell will be there to guide you through the
season and throw in his two cents as well.
Speaker 2 (00:43):
Look out for new episodes of Smart Talks with IBM
every other week on the iHeartRadio app, Apple Podcasts, or
wherever you get your podcasts. And learn more at IBM
dot com slash smart Talks.
Speaker 3 (00:56):
Hey, it's Jacob Goldstein for Smart Talks with IBM. Last
year I had the pleasure of sitting down with doctor
David Cox, VP of AI Models at IBM Research. We
explored the fascinating world of AI foundation models and their
revolutionary potential for business automation and innovation. When we first
aired this episode, the concept of foundation models was just
(01:19):
beginning to capture our attention. Since then, this technology has
evolved and redefined the boundaries of what's possible. Businesses are
becoming more savvy about selecting the right models and understanding
how they can drive revenue and efficiency. As I listened
back to the conversation, it was interesting to reflect on
some new developments and ideas that have emerged, and many
(01:41):
of these we will continue to explore throughout the season,
like how to play an active role in choosing the
best model for your needs. Whether you're a longtime listener
or tuning in for the first time, I'm certain you'll
find doctor Cox's insights as thought provoking as ever. Thanks
as always for joining us. Now let's dive in.
Speaker 4 (02:01):
Hello, Hello, Welcome to Smart Talks with IBM, a podcast
from Pushkin Industries, iHeartRadio and IBM. I'm Malcolm Gladwell. Our
guest today is doctor David Cox, VP of AI Models
at IBM Research and IBM Director of the MIT IBM
Watson AI Lab, a first of its kind industry academic
(02:24):
collaboration between IBM and MIT focused on the fundamental research
of artificial intelligence. Over the course of decades, David Cox
watched as the AI revolution steadily grew from the simmering
ideas of a few academics and technologists into the industrial
boom we are experiencing today. Having dedicated his life to
(02:48):
pushing the field of AI towards new horizons, David has
both contributed to and presided over many of the major
breakthroughs in artificial intelligence. In today's episode, you'll hear David
explain some of the conceptual underpinnings of the current AI landscape,
things like foundation models, in surprisingly comprehensible terms, I might add,
(03:11):
we'll also get into some of the amazing practical applications
for AI in business, as well as what implications AI
will have for the future of work and design. David
spoke with Jacob Goldstein, host of the Pushkin podcast What's
Your Problem. A veteran business journalist, Jacob has reported for
The Wall Street Journal, the Miami Herald, and was a
(03:32):
longtime host of the NPR program Planet Money. Okay, let's
get to the interview.
Speaker 3 (03:41):
Tell me about your job at IBM.
Speaker 5 (03:43):
So I wear two hats at IBM. So one, I'm
the IBM Doctor of the MIT IBM Watson AI Lab.
So that's a joint lab between IBM and MIT where
we try and invent what's next in AI. It's been
running for about five years, and then more recently I
started as Vice president for AI Models, and I'm in
charge of building IBM's foundation models, you know, building these
(04:07):
these big models, generative models that allow us to have
all kinds of new exciting capabilities in AI.
Speaker 3 (04:11):
So, so I want to talk to you a lot
about foundation models, about genitive AI. But before we get
to that, let's just spend a minute on the on
the IBM MIT collaboration. Where did that partnership start, How
did it originate?
Speaker 5 (04:27):
Yeah, So, actually it turns out that MIT and IBM
have been collaborating for a very long time in the
area of AI. In fact, the term artificial intelligence was
coined in a nineteen fifty six workshop that was held
at Dartmouth. It was actually organized by an IBM or
Nathaniel Rochester, who led the development of the IBM seven
and one. So we've really been together in AIS since
(04:49):
the beginning and as AI kept accelerating more and more
and more, I think there was a really interesting decision
to say, let's make this a formal partnership. So IBM
in twenty seventeen and now, so it'll be committing close
to a quarter billion dollars over ten years to have
this joint lab with MIT, and we we located ourselves
right on the campus and we've been developing very very
(05:12):
deep relationships where we can you know, really get to
know each other, work shoulder to shoulder, conceiving what we
should work on next, and then executing the projects. And
it's really you know, very few entities like this exist
between academia industry. It's been really fun of the last
five years to be a part of it.
Speaker 3 (05:29):
And what do you think are some of the most
important outcomes of this collaboration between IBM and MIT.
Speaker 5 (05:36):
Yeah, so we're really kind of the tip of the
sphere for for IBM's the I strategy. So we're really looking,
you know, what's coming ahead, and you know, in areas
like foundation models, you know, as the field changes, MIT
people are interested in working on you know, faculty, students
and staff are interested in working on what's the latest thing,
what's the next thing. We at IBM Research are very
(05:58):
much interested in the same. We can kind of put
out feelers, you know, interesting things that we're seeing in
our research, interesting things we're hearing in the field. We
can go and chase those opportunities. So when something big comes,
like the big change that's been happening lately with foundation models,
we're ready to jump on it. That's really the purpose,
that's that's the lab functioning the way it should. We're
also really interested in how do we advance you know,
(06:21):
AI that can help with climate change or you know,
build better materials and all these kinds of things that
are you know, a broader aperture sometimes than what we
might consider just looking at the product portfolio of IBM,
and that that gives us again a breadth where we
can see connections that we might not have seen otherwise.
We can you know, think things that help out society
and also help out our customers.
Speaker 3 (06:43):
So the last whatever six months, say, there has been
this wild rise in the public's interest in AI right
clearly coming out of these generative AI models that are
really accessible you know, certainly chat GPT language models like that,
as well as models that generate images like mid journey.
(07:05):
I mean, can you just sort of briefly talk about
the breakthroughs in AI that have made this moment feel
so exciting, so revolutionary for artificial intelligence.
Speaker 5 (07:16):
Yeah. You know, I've been studying AI basically my entire
adult life. Before I came to IBM, I was a
professor at Harvard. I've been doing this a long time,
and I've gotten used to being surprised. It sounds like
a joke, but it's serious, Like I'm getting used to
being surprised at the acceleration of the pace again. It
tracks actually a long way back. You know, there's lots
(07:38):
of things where there was an idea that just simmered
for a really long time. Some of the key math
behind the stuff that we have today, which is amazing.
There's an algorithm called back propagation, which is sort of
key to training neural networks that's been around, you know,
since the eighties in wide use. And really what happened
(07:59):
was it simmered for a long time, and then enough
data and enough compute came so we had enough data
because you know, we all started carrying multiple cameras around
with us. Our mobile phones have all you know, all
these cameras and this we put everything on the Internet,
and there's all this data out there. We caught a
lucky break that there was something called the graphics processing unit,
(08:21):
which turns out to be really useful for doing these
kinds of algorithms, maybe even more useful than it is
for doing graphics. They're greater graphics too, And things just
kept kind of adding to the snowball. So we had
deep learning, which is sort of a rebrand of neural
networks that I mentioned from the eighties, and that was
enabled again by data because we digitalized the world and
(08:43):
compute because we kept building faster and faster and more
powerful computers, and then that allowed us to make this
this big breakthrough. And then, you know, more recently, using
the same building blocks, that inexorable rise of more and
more and more data, that technology called self supervised learning.
Where the key difference there in traditional deep learning, you know,
(09:07):
for classifying images, you know, like is this a cat
or is this a dog? And a picture those technologies
require supper visions, so you have to take what you
have and then you have to label it. So you
have to take a picture of a cat and then
you label it as a cat, and it turns out that,
you know, that's very powerful, but it takes a lot
of time to label gats and to label dogs, and
(09:28):
there's only so many labels that exist in the world.
So what really changed more recently is that we have
self supervised learning where you don't have to have the labels.
We can just take unannotated data. And what that does
is it allows you use even more data. And that's
really what drove this this latest sort of rage. And
then and then all of a sudden we start getting
these these really powerful models. And then really, this has
(09:52):
been simmering technologies, right, this has been happening for a
while and progressively getting more and more powerful. One of
the things that really happened with CHATGBT and technologies like
Stable Diffusion and mid Journey was that they made it
visible to the public. You know, you put it out
(10:12):
there the public can touch and feel and they're like, wow,
not only is there palpable change, and wow this you know,
I can talk to this thing. Wow, this thing can
generate an image. Not only that, but everyone can touch
and feel and try. My kids can use some of
these AI art generation technologies, and that's really just launched.
(10:32):
You know, it's like a propelled slingshot at us into
a different regime. In terms of the public awareness of
these technologies.
Speaker 3 (10:39):
You mentioned earlier in the conversation foundation models, and I
want to talk a little bit about that. I mean,
can you just tell me, you know, what are foundation
models for AI and why are they a big deal?
Speaker 5 (10:52):
Yeah, So this term foundation model was coined by a
group at Stanford, and I think it's actually a really
apt term because remember I said, you know, one of
the big things that unlocked this latest excitement was the
fact that we could use large amounts of unannotated data.
We could train a model. We don't have to go
through the painful effort of labeling each and every example.
(11:16):
You still need to have your model do something you
wanted to do. You still need to tell it what
you want to do. You can't just have a model
that doesn't, you know, have any purpose. But what a
foundation models that provides a foundation, like a literal foundation.
You can sort of stand on the shoulders of giants.
You can have them these massively trained models, and then
do a little bit on top. You know, you could
use just a few examples of what you're looking for
(11:39):
and you can get what you want from the model,
So just a little bit on top now gets to
the results that a huge amount of effort used to
have to put in, you know, to get from the
ground up to that level.
Speaker 3 (11:50):
I was trying to think of of an analogy for
sort of foundation models versus what came before, and I
don't know that I came up with a good one,
but the best I could do was this. I want
you to tell me if it's plausible. It's like before
foundation models, it was like you had these sort of
single use kitchen appliances. You could make a waffle iron
(12:11):
if you wanted waffles, or you could make a toaster
if you wanted to make toast. But a foundation model
is like like an oven with a range on top.
So it's like this machine and you could just cook
anything with this machine.
Speaker 5 (12:24):
Yeah, that's a great analogy. They're very versatile. The other
piece of it, too, is that they dramatically lower the
effort that it takes to do something that you want
to do. And sometimes I used to say about the
old world of AI, would say, you know, the problem
with automation is that it's too labor intensive. H sounds
like I'm making a joke.
Speaker 3 (12:44):
Indeed, famously, if automation does one thing, it substitutes machines
or computing power for labor. Right, So what does that
mean to say AI is or automation is too labor intensive.
Speaker 5 (12:57):
It sounds like I'm making a joke, but I'm actually serious.
What I mean is that the effort it took the
old regime to automate something was very, very high. So
if I need to go and curate all this data,
collect all this data, and then carefully label all these examples,
that labeling itself might be incredibly expensive and time. So
(13:17):
and we estimate anywhere between eighty to ninety percent of
the effort it takes to feel an AI solution actually
is just spent on data, so that that has some consequences,
which is the threshold for bothering. You know, if you're
going to only get a little bit of value back
from something, are you going to go through this huge
effort to curate all this data and then when it
(13:40):
comes time to train the model, you need highly skilled
people expensive or hard to find in the labor market.
You know, are you really going to do something that's
just a tiny, little incremental thing. Now, you're going to
do the only the highest value things that weren't at
level because.
Speaker 3 (13:55):
You have to essentially build the whole machine from scratch,
and there aren't many things where it's worth that much
work to build a machine that's only going to do
one narrow thing.
Speaker 5 (14:06):
That's right, and then you tackle the next problem and
you basically have to start over. And you know, there
are some nuances here, like for images, you can pre
train a model on some other tasks and change it around.
So there are some examples of this like non recurring
cost that we have in the old world too, But
by and large, it's just a lot of effort. It's hard.
It takes, you know, a large level of skill to implement.
(14:30):
One analogy that I like is, you know, think about
it as you know, you have a river of data,
you know, running through your company or your institution. Traditional
AI solutions are kind of like building a dam on
that river. You know, dams are very expensive things to build.
They require highly specialized skills and lots of planning. And
you know, you're only going to put a dam on
(14:51):
a river that's big enough that you're gonna get enough
energy out of it that it was worth trouble. You're
gonna get a lot of value out of that dam.
If you have a river like that, you know, a
river of data, but it's actually the vast majority of
the water you know in your kingdom actually isn't in
that river. It's in puddles and greeks and babid brooks,
And you know, there's a lot of value left on
(15:14):
the table because it's like, well, I can't there's nothing
you can do about it. It's just that that's too
low value. So it takes too much effort, so I'm
just not going to do it. The return around investment
just isn't there, so you just end up not automating
things because it's too much of a pain. Now what
foundation models do is they say, well, actually, no, we
can train a base model a foundation that you can
(15:35):
work on that we don't we don't care. We don't
specify what the task is ahead of time. We just
need to learn about the domain of data. So if
we want to build something that can understand English language,
there's a ton of English language text available out in
the world. We can now train models on huge quantities
of it, and then it learned the structure. It learned
(15:56):
how language, you know, good part of how language works
on all that unlabeled data. And then when you roll
up with your task, you know, I want to solve
this particular problem, you don't have to start from scratch.
You're starting from a very very very high place. So
that just gives you the ability to just you know, now,
all of a sudden, everything is accessible. All the puddles
(16:16):
and greeks and babbling brooks and kettlepons, you know, those
are all accessible now. And that's that's very exciting. But
it just changes the equation on what kinds of problems
you could use AI to solve.
Speaker 3 (16:27):
And so foundation models basically mean that automating some new
task is much less labor intensive, The sort of marginal
effort to do some new automation thing is much lower
because you're building on top of the foundation model rather
than starting from scratch. Absolutely, so that is that is
like the exciting good news. I do feel like there's
(16:52):
a little bit of a countervailing idea that's worth talking
about here, and that is the idea that even though
there are these foundation models that are really powerful, that
are relatively easy to build on top of, it's still
the case, right that there is not some one size
fits all foundation model. So you know, what does that
mean and why is that important to think about in
(17:13):
this context?
Speaker 5 (17:14):
Yeah, so we believe very strongly that there isn't just
one model to rule them all. There's a number of
reasons why that could be true. One which I think
is important and very relevant today is how much energy
these models can consume. So these models, you know, can
get very very large. So one thing that we're starting
(17:39):
to see or starting to believe, is that you probably
shouldn't use one giant sledgehammer model to solve every single problem,
you know, like we should pick the right size model
to solve the problem. We shouldn't necessarily assume that we
need the biggest, baddest model for every little use case.
And we're also seeing that, you know, small models that
(17:59):
are trained to like to specialize on particular domains can
actually outperform much bigger models. So bigger isn't always even better.
Speaker 3 (18:07):
So they're more efficient and they do the thing you
want them to do better as well.
Speaker 5 (18:12):
That's right. So Stanford, for instance, a group of Stanford
trained a model. It is a two point seven billion
parameter model, which isn't terribly big by today's standards. They
trained it just on the biomedical literature, you know, this
is the kind of thing that universities do, and what
they showed was that this model was better at answering
questions about the biomedical literature than some models that were
(18:33):
one hundred billion parameters, you know, many times larger. So
it's a little bit like you know, asking an expert
for help on something versus asking the smartest person. You know,
the smartest person you know may be very smart, but
they're not going to be expertise. And then as an
added bonus, you know, this is now a much smaller model,
it's much more efficient to run. We are you know,
(18:54):
you know, it's cheaper. So there's lots of different advantages there.
So I think we're going to see attention in the
industry between vendors that say, hey, this is the one,
you know, big model, and then others that say, well, actually,
you know, there's there's you know, lots of different tools
we can use that all have this nice quality that
we outlined at the beginning, and then we should really
(19:15):
pick the one that makes the most sense for the
task at hand.
Speaker 3 (19:19):
So there's sustainability basically efficiency, another kind of set of
issues that come up a lot with AI A are bias, hallucination.
Can you talk a little bit about bias and hallucination,
what they are and how you're working to mitigate those problems.
Speaker 5 (19:34):
Yeah, so there are lots of issues still as amazing
as these technologies are, and they are amazing, let's let's
be very clear, lots of great things we're going to
enable with these kinds of technologies. Bias isn't a new problem.
So you know, basically we've seen this since the beginning
of AI. If you train a model on data that
(19:55):
has a bias in it, the model is going to
recapitulate that bias when it provides it's answers. So every time,
you know, if all the text you have says, you know,
it's more likely to refer to female nurses and male scientists,
then you're going to you know, get models that you know.
For instance, there was an example where a machine learning
based translation system translated from Hungarian to English. Hungarian doesn't
(20:17):
have gendered pronouns. English does, and when you ask them
to translate, it would translate they are a nurse to
she is a nurse, translate they are a scientist. To
he is a scientist. And that's not because the people
who wrote the algorithm were building in bias and coding
in like, oh, it's got to be this way. It's
because the data was like that. You know, we have
biases in our society and they're reflected in our data
(20:40):
and our text and our images everywhere. And then the
models they're just mapping from what they what they've seen
in their training data to to the result that you're
trying to get them to do and to give, and
then these biases come out. So there's a very active
program of research and you know, we do quite a
bit at IBM research and my T but also all
(21:03):
over the community and industry and academia trying to figure
out how do we explicitly remove these biases, how do
we identify them, how do you know, how do we
build tools that allow people to audit their systems to
make sure they aren't biased. So this is a really
important thing. And you know, again this was here since
the beginning, you know, of machine learning and AI, but
(21:24):
foundation models and large language models and generative AI just
bring it into sharper even sharper focus because there's just
so much data and it's sort of building in baking
in all these different biases we have, so that that's
that's absolutely a problem that these models have. Another one
that you mentioned was hallucinations. So even the most impressive
(21:45):
of our models will often just make stuff up. You know,
the technical term that the field has chosen is hallucination.
To give you an example, I asked chat tbt to
create a biography of David Cox at IBM, and you know,
it started off really well. You know, the identified that
I was the director of the MNT IBM Watson and
(22:05):
said a few words about that, and then it proceeded
to create an authoritative but completely fake biography of me
where I was British, I was born in the UK,
I went to British university, you know universities in the UK.
I was professor the authority.
Speaker 3 (22:21):
Right, it's the certainty that that is weird about it, right,
It's it's dead certain that you're from the UK, et cetera.
Speaker 5 (22:28):
Absolutely, yeah, it has all kinds of flourishes like I
want awards in the UK. So yeah, it's it's problematic
because it kind of pokes a lot of weak spots
in our human psychology where if something sounds coherent, We're
likely to assume it's true. We're not used to interacting
with people who eloquently and authoritatively, you know, emit complete nonsense,
(22:52):
like yeah, you know, we can debate about that, but.
Speaker 3 (22:55):
Yeah, we can debate about that. But yes, it's the
sort of blive confidence throws you off when you realize
it's completely wrong.
Speaker 5 (23:03):
Right, that's right. And we do have a little bit
of like a great and powerful oz sort of vibe
going sometimes where we're like, well, you know, the AI
is all knowing and therefore whatever it says must be true.
But these things will make up stuff, you know, very aggressively,
and you know, you everyone can try asking it for
their their bio. You'll you'll get something that You'll always
(23:26):
get something that's of the right form, that has the
right tone. But you know, the facts just aren't necessarily there.
So that's obviously a problem. We need to figure out
how to close those gaps, fix those problems. There's lots
of ways we can use them much more easily.
Speaker 4 (23:40):
I'd just like to say, faced with the awesome potential
of what these technologies might do, it's a bit encouraging
to hear that even chat GPT has a weakness for
inventing flamboyant, if fictional versions of people's lives, and while
entertaining ourselves with chat GPT and mid journey is important,
the way lpeople use consumer facing chatbots and generative AI
(24:04):
is just fundamentally different from the way an enterprise business
uses AI. How can we harness the abilities of artificial
intelligence to help us solve the problems we face in
business and technology. Let's listen on as David and Jacob
continue their conversation.
Speaker 3 (24:21):
We've been talking in a somewhat abstract way about AI
in the ways it can be used. Let's talk in
a little bit more of a specific way. Can you
just talk about some examples of business challenges that can
be solved with automation, with this kind of automation we're
talking about.
Speaker 5 (24:39):
Yeah, so the really really, this guy's the limit. There's
a whole set of different applications that these models are
really good at. And basically it's a superset of everything
we used to use AI for in business. So you know,
the simple kinds of things are like, hey, if I
have text and I have product reviews, and I want
to be able to tell if these are positive or negative.
(25:00):
You know, like let's look at all the negative reviews,
so we can have a human look through them and
see what was up. Very common business use case. You
can do it with traditional deep learning based AI. So so
there's things like that that are you know, it's very
prosaic sort of we were already doing it. We've been
doing it for a long time. Then you get situations
that are that were harder for the old day. I like,
(25:21):
if I'm I want to impress something like I want
to I have like say I have a chat transcript,
Like a customer called in and they had a complaint.
They called back. Okay, now a new you know, a
person on the line needs to go read the old
transcript to catch up. Wouldn't it be better if we
could just summarize that, just condense it all down a
(25:41):
quick little paragraph. You know, customer called they we upset
about this, rather than having to read the blow by blow.
There's just lots of settings like that where summarization is
really helpful. Hey, you have a meeting and I'd like
to just automatically you know, have had that meeting or
that email or whatever. I'd like to just have a
condensed down so I can really quickly get to the
heart of the matter. These models are are really good
(26:02):
at doing that. They're also really good at question answering.
So if I want to find out, what's how many
vacation days do I have? I can now interact in
natural language with a system that can go and that
it has access to our HR policies, and I can
actually have a you know, a multi turn conversation where
I can, you know, like I would have with you know, somebody,
(26:22):
you know, actual HR professional or customer service representative. So
a big part, you know, of what this is doing
is it's it's putting an interface. You know, when we
think of computer interfaces, we're usually thinking about UI user
interface elements where I click on menus and there's buttons
and all this stuff. Increasingly, now we can just talk,
(26:44):
you know, you just in words. You can describe what
you want, you want to answer, ask a question, you
want to sort of command the system to do something,
rather than having to learn how to do that clicking buttons,
which might be inefficient, Now we can just sort of
spell it out.
Speaker 3 (26:57):
Interesting, right, the graphical user interface that we all sort
of default to, that's not like the state of nature, Right,
that's a thing that was invented and just came to
be the standard way that we interact with computers. And
so you could imagine, as you're saying, like chat essentially
chatting with the machine could could become a sort of
standard user interface, just like the graphical user interface, did
(27:20):
you know over the past several decades.
Speaker 5 (27:22):
Absolutely, And I think those kinds of conversational interfaces are
going to be hugely important for increasing our productivity. It's
just a lot easier if I if I have to
learn how to use a tool, or I don't have
to kind of have awkward, you know, interactions from the computer.
I can just tell it what I want and I
can understand it. Could you know, potentially even ask questions
back to clarify and have those kinds of conversations that
(27:45):
can be extremely powerful. And in fact, one area where
that's going to I think be absolutely game changing is
in code. When we write code. You know, programming languages
are a way for us to sort of match between
a very sloppy way of talking and the very exact
way that you need to command a computer to do
(28:05):
what you wanted to do. They're cumbersome to learn, they
can you know, create very complex systems that are very
hard to reason about. And we're already starting to see
the ability to just write down what you want and
AI will generate the code for you. And I think
we're just going to see a huge revolution of like
we just converse you and we can have a conversation
to say what we want, and then the computer can
actually not only do fixed actions and do things for us,
(28:29):
but it can actually even write code to do new things,
you know, and generate software itself. Given how much software
we have, how much craving we have for software, like
we'll never have enough software in our world. Uh, you know,
the ability to have AI systems as a helper in that,
I think we're going to see a lot of a
lot of value there.
Speaker 3 (28:48):
So if you if you think about the different ways
AI might be applied to business, I mean you've talked
about a number of the sort of classic use cases.
What are some of the more out there use cases.
What are some you know, unique ways you could imagine
AI being applied to business.
Speaker 5 (29:06):
You know, there's really disguised the limit. I mean, we
have one project that I'm kind of a fan of
where we actually were working with a mechanical engineering professor
at MIT working on a classic problem, how do you
build linkage systems which are like you imagine bars and
joints and overs.
Speaker 3 (29:24):
You know, the things that are building a thing, building
a physical machine of some kind.
Speaker 5 (29:29):
Like real like metal and you know nineteenth century just
old school industrial revolution. Yeah yeah, yeah, but you know
the little arm that's that's holding up my microphone in
front of me. Cranes, get build your buildings, you know,
parts of your engines. This is like classical stuff. It
turns out that you know, humans, if you want to
build an advanced system, you decide what like curve you
(29:50):
want to create, and then a human together with a
computer program, can build a five or six bar linkage
and then that's kind of where you top out it
because it gets too comp replicated to work more than that.
We built a generative AI system that can build twenty
bar linkages, like arbitrarily complex. So these are machines that
are beyond the capability of a human to design themselves.
(30:13):
Another example, we have an AI system that can generate
electronic circuits. You know, we had a project where we're
working where we're building better power converters which allow our
computers and our devices to be more efficient, save energy,
you know, less less carbone. But I think the world
around us has always been shaped by technology. If we
look around, you know, just think about how many steps
(30:35):
and how many people and how many designs went into
the table and the chair and the lamp. It's really
just astonishing. And that's already you know, the fruit of
automation and computers and those kinds of tools. But we're
going to see that increasingly be product also of AI.
It's just going to be everywhere around us. Everything we
touch is going to have been helped in some way
(30:56):
to get to you by.
Speaker 3 (30:58):
A you know, that is a pretty profound transformation that
you're talking about in business. How do you think about
the implications of that both for the sort of you know,
business itself and also for employees.
Speaker 5 (31:12):
Yeah, so I think for businesses this is gonna cut costs,
make new opportunities to like customers, you know, like there's
just you know, it's sort of all upside right like
for the for the workers, I think the story is
mostly good too. You know, like how many things do
you do in your day that you'd really rather not?
Speaker 2 (31:33):
Right?
Speaker 5 (31:34):
You know, and we're used to having things we don't
like automated away. You know, we we didn't. You know,
if you didn't like walking many miles to work, then
you know, like you can have a car and you
can drive there. Or we used to have a huge
fraction over ninety percent of the US population engaged in agriculture,
and then we mechanized it. How very few people work
in agriculture, a small number of people can do the
(31:54):
work of a large number of people. And then you know,
things like email, and yeah, they've led to huge productor
the enhancements because I don't need to be writing letters
and sending them in the mail. I can just instantly
communicate with people. We just become more effective, Like our
jobs have transformed, whether it's a physical job like agriculture
or whether it's a knowledge worker job where you're sending
(32:16):
emails and communicating with people and coordinating teams. We've just
gotten better. And you know, the technology has just made
us more productive. And this is just another example. Now,
you know, there are people who worry that you know,
will be so good at that that maybe jobs will
be displaced, and that's a legitimate concern. But just like
(32:36):
how in agriculture, you know, it's not like suddenly we
had ninety percent of the population unemployed. You know, people
transitioned to other jobs. And the other thing that we
found too, is that our appetite for doing more things
is as humans is sort of insatiable. So even if
we can dramatically increase how much one human can do,
(32:58):
that doesn't necessarily mean we're going to do fixed amount
of stuff. There's an appetite to have even more, so
we're going to you can continue to grow, grow the pie.
So I think at least certainly in the near term,
you know, we're going to see a lot of drudgery
go away from work. We're going to see people be
able to be more effective at their jobs. You know,
we will see some transformation in jobs and like we've
(33:19):
seen that before, and the technology least has the potential
to make our lives a lot easier.
Speaker 3 (33:27):
So IBM recently launched Watson X, which includes Watson X
dot AI. Tell me about that, Tell me about you
know what it is and the new possibilities that it
opens up.
Speaker 5 (33:38):
Yeah. So so Watson X is obviously a bit of
a new branding on the Watson brand.
Speaker 2 (33:46):
T J.
Speaker 5 (33:47):
Watson that was the founder of IBM and our EI
technologies have had the Watson brand. Watson X is a
recognition that that there's something new, there's something that actually
has changed the game.
Speaker 3 (33:59):
You know.
Speaker 5 (33:59):
We gone from this old world of automation is to
labor intensive to this new world of possibilities where it's
much easier to use AI. And what watsonex does it
brings together tools for businesses to harness that power. So
whatsonex dot AI foundation models that our customers can use.
(34:21):
It includes tools that make it easy to run, easy
to deploy, easy to experiment. There's a watsonex dot Data
component which allows you to sort of organize and access
to your data. So what we're really trying to do
is give our customers a cohesive set of tools to
harness the value of these technologies and at the same
(34:43):
time be able to manage the risks and other things
that you have to keep an eye on in an
enterprise context.
Speaker 3 (34:50):
So we talk about the guests on this show as
new creators, by which we mean people who are creatively
applying technology in businesiness to drive change. And I'm curious
how creativity plays a role in the research that you do.
Speaker 5 (35:08):
I Honestly, I think the creative aspects of this job,
this is what makes this work exciting. You know, I
should say, you know, the folks who work in my
organization are doing the creating, and I guess you're.
Speaker 3 (35:24):
Doing the managing so that they could do the creator.
Speaker 5 (35:27):
I'm helping them be their best, and I still get
to get involved in the weeds of the research as
much as I can. But you know, there's something really
exciting about inventing, you know, like one of the nice
things about doing invention and doing research on AI. In industries,
it's usually grounded and a real problem that somebody is having.
(35:48):
You know, a customer wants to solve this problem. It's
losing money, or there wuld be a new opportunity. You
identify that problem and then you build something that's never
been built before to do that. And I think that's
honestly the adrenaline rush that keeps all of us in
this field. How do you do something that nobody else
(36:08):
on earth has has done before or tried before, so
that that kind of creativity and there's also creativity as
well and identifying what those problems are, being able to
understand the places where the technology is close enough to
solving a problem and doing that matchmaking between problems that
(36:28):
are now solvable, you know, and an AI where the
field is moving so fast, this is constantly growing horizon
of things that we might be able to solve. So
that matchmaking, I think, is also a really interesting creative problem.
So I think I think that's that's that's why it's
so much fun. And it's a fun environment we have
here too. It's you know, people drawing on whiteboards and
(36:50):
writing on pages of math and like in a movie,
like in a movie, yes, straight from sexual casting.
Speaker 3 (36:58):
Drawing, the drawing on the window, writing on the window,
and sharp absolutely so, so let's close with the really
long view. How do you imagine AI and people working
together twenty years from now?
Speaker 5 (37:16):
Yeah, it's really hard to make predictions. The vision that
I I like, actually this came from an MIT economist
named David Ottur, which was imagine AI almost as a
natural resource. You know, we know how natural resources work, right,
(37:38):
Like there's an or we can dig up out of
the earth that comes from kind of springs from the earth,
or we usually think of that in terms of physical stuff.
With AI, you can almost think of it as like
there's a new kind of abundance potentially twenty years from now,
where not only can we have things we can build
or eat or use or burn or whatever. Now we have,
you know, this ability to do things and understand things
(37:59):
and do intellectual work, and I think we can get
to a world where automating things is just seamless. We're
surrounded by capability to augment ourselves to get things done.
And you could think of that in terms of like, oh,
that's going to displace our jobs, because eventually the AI
system is going to do everything we can do. But
(38:19):
you could also think of it in terms of like, wow,
that's just so much abundance that we now have, and
really how we use that abundance is sort of up
to us, you know, like when you can writing software
is super easy and fast and anybody can do it.
Just think about all the things you can do now, like,
think about all the new activities and go about all
the ways we could use that to enrich our lives.
(38:40):
That's where I'd like to see us in twenty years.
You know, we can we can do just so much
more than we were able to do before abundance.
Speaker 3 (38:50):
Great, thank you so much for your time.
Speaker 5 (38:53):
Yeah, It's been a pleasure. Thanks for inviting me.
Speaker 4 (38:57):
What a far ranging, deep conversation. I'm mesmerized by the
vision David just described. A world where natural conversation between
mankind and machine can generate creative solutions to our most
complex problems. A world where we view AI not as
our replacements, but as a powerful resource we can tap
(39:17):
into and exponentially boost our innovation and productivity. Thanks so
much to doctor David Cox for joining us on smart Talks.
We deeply appreciate him sharing his huge breadth of AI
knowledge with us and for explaining the transformative potential of
foundation models in a way that even I can understand.
(39:39):
We eagerly await his next great breakthrough. Smart Talks at
IBM is produced by Matt Romano, David jaw nishe Venkat
and Royston Preserve with Jacob Goldstein. We're edited by Lydia
Jean Kott. Our engineers are Jason Gambrel, Sarah Bouguer and
Ben Holliday. Theme song by Gramma's Scope. Special thanks to
(40:01):
Carli Megliori, Andy Kelly, Kathy Callahan and the Eight Bar
and IBM teams, as well as the Pushkin marketing team.
Smart Talks with IBM is a production of Pushkin Industries
and iHeartMedia. To find more Pushkin podcasts, listen on the
iHeartRadio app, Apple Podcasts, or wherever you listen to podcasts.
(40:23):
Him Malcolm Gladwell. This is a paid advertisement from IBM.
Hey everyone, it's Robert and Joe here. Today we've got
something a little bit different to share with you. It
is a new season of the Smart Talks with IBM
podcast series.
Speaker 2 (00:09):
Today we are witnessed to one of those rare moments
in history, the rise of an innovative technology with the
potential to radically transform business and society forever. The technology,
of course, is artificial intelligence, and it's the central focus
for this new season of Smart Talks with IBM.
Speaker 1 (00:25):
Join hosts from your favorite Pushkin podcasts as they talk
with industry experts and leaders to explore how businesses can
integrate AI into their workflows and help drive real change
in this new era of AI. And of course, host
Malcolm Gladwell will be there to guide you through the
season and throw in his two cents as well.
Speaker 2 (00:43):
Look out for new episodes of Smart Talks with IBM
every other week on the iHeartRadio app, Apple Podcasts, or
wherever you get your podcasts. And learn more at IBM
dot com slash smart Talks.
Speaker 3 (00:56):
Hey, it's Jacob Goldstein for Smart Talks with IBM. Last
year I had the pleasure of sitting down with doctor
David Cox, VP of AI Models at IBM Research. We
explored the fascinating world of AI foundation models and their
revolutionary potential for business automation and innovation. When we first
aired this episode, the concept of foundation models was just
(01:19):
beginning to capture our attention. Since then, this technology has
evolved and redefined the boundaries of what's possible. Businesses are
becoming more savvy about selecting the right models and understanding
how they can drive revenue and efficiency. As I listened
back to the conversation, it was interesting to reflect on
some new developments and ideas that have emerged, and many
(01:41):
of these we will continue to explore throughout the season,
like how to play an active role in choosing the
best model for your needs. Whether you're a longtime listener
or tuning in for the first time, I'm certain you'll
find doctor Cox's insights as thought provoking as ever. Thanks
as always for joining us. Now let's dive in.
Speaker 4 (02:01):
Hello, Hello, Welcome to Smart Talks with IBM, a podcast
from Pushkin Industries, iHeartRadio and IBM. I'm Malcolm Gladwell. Our
guest today is doctor David Cox, VP of AI Models
at IBM Research and IBM Director of the MIT IBM
Watson AI Lab, a first of its kind industry academic
(02:24):
collaboration between IBM and MIT focused on the fundamental research
of artificial intelligence. Over the course of decades, David Cox
watched as the AI revolution steadily grew from the simmering
ideas of a few academics and technologists into the industrial
boom we are experiencing today. Having dedicated his life to
(02:48):
pushing the field of AI towards new horizons, David has
both contributed to and presided over many of the major
breakthroughs in artificial intelligence. In today's episode, you'll hear David
explain some of the conceptual underpinnings of the current AI landscape,
things like foundation models, in surprisingly comprehensible terms, I might add,
(03:11):
we'll also get into some of the amazing practical applications
for AI in business, as well as what implications AI
will have for the future of work and design. David
spoke with Jacob Goldstein, host of the Pushkin podcast What's
Your Problem. A veteran business journalist, Jacob has reported for
The Wall Street Journal, the Miami Herald, and was a
(03:32):
longtime host of the NPR program Planet Money. Okay, let's
get to the interview.
Speaker 3 (03:41):
Tell me about your job at IBM.
Speaker 5 (03:43):
So I wear two hats at IBM. So one, I'm
the IBM Doctor of the MIT IBM Watson AI Lab.
So that's a joint lab between IBM and MIT where
we try and invent what's next in AI. It's been
running for about five years, and then more recently I
started as Vice president for AI Models, and I'm in
charge of building IBM's foundation models, you know, building these
(04:07):
these big models, generative models that allow us to have
all kinds of new exciting capabilities in AI.
Speaker 3 (04:11):
So, so I want to talk to you a lot
about foundation models, about genitive AI. But before we get
to that, let's just spend a minute on the on
the IBM MIT collaboration. Where did that partnership start, How
did it originate?
Speaker 5 (04:27):
Yeah, So, actually it turns out that MIT and IBM
have been collaborating for a very long time in the
area of AI. In fact, the term artificial intelligence was
coined in a nineteen fifty six workshop that was held
at Dartmouth. It was actually organized by an IBM or
Nathaniel Rochester, who led the development of the IBM seven
and one. So we've really been together in AIS since
(04:49):
the beginning and as AI kept accelerating more and more
and more, I think there was a really interesting decision
to say, let's make this a formal partnership. So IBM
in twenty seventeen and now, so it'll be committing close
to a quarter billion dollars over ten years to have
this joint lab with MIT, and we we located ourselves
right on the campus and we've been developing very very
(05:12):
deep relationships where we can you know, really get to
know each other, work shoulder to shoulder, conceiving what we
should work on next, and then executing the projects. And
it's really you know, very few entities like this exist
between academia industry. It's been really fun of the last
five years to be a part of it.
Speaker 3 (05:29):
And what do you think are some of the most
important outcomes of this collaboration between IBM and MIT.
Speaker 5 (05:36):
Yeah, so we're really kind of the tip of the
sphere for for IBM's the I strategy. So we're really looking,
you know, what's coming ahead, and you know, in areas
like foundation models, you know, as the field changes, MIT
people are interested in working on you know, faculty, students
and staff are interested in working on what's the latest thing,
what's the next thing. We at IBM Research are very
(05:58):
much interested in the same. We can kind of put
out feelers, you know, interesting things that we're seeing in
our research, interesting things we're hearing in the field. We
can go and chase those opportunities. So when something big comes,
like the big change that's been happening lately with foundation models,
we're ready to jump on it. That's really the purpose,
that's that's the lab functioning the way it should. We're
also really interested in how do we advance you know,
(06:21):
AI that can help with climate change or you know,
build better materials and all these kinds of things that
are you know, a broader aperture sometimes than what we
might consider just looking at the product portfolio of IBM,
and that that gives us again a breadth where we
can see connections that we might not have seen otherwise.
We can you know, think things that help out society
and also help out our customers.
Speaker 3 (06:43):
So the last whatever six months, say, there has been
this wild rise in the public's interest in AI right
clearly coming out of these generative AI models that are
really accessible you know, certainly chat GPT language models like that,
as well as models that generate images like mid journey.
(07:05):
I mean, can you just sort of briefly talk about
the breakthroughs in AI that have made this moment feel
so exciting, so revolutionary for artificial intelligence.
Speaker 5 (07:16):
Yeah. You know, I've been studying AI basically my entire
adult life. Before I came to IBM, I was a
professor at Harvard. I've been doing this a long time,
and I've gotten used to being surprised. It sounds like
a joke, but it's serious, Like I'm getting used to
being surprised at the acceleration of the pace again. It
tracks actually a long way back. You know, there's lots
(07:38):
of things where there was an idea that just simmered
for a really long time. Some of the key math
behind the stuff that we have today, which is amazing.
There's an algorithm called back propagation, which is sort of
key to training neural networks that's been around, you know,
since the eighties in wide use. And really what happened
(07:59):
was it simmered for a long time, and then enough
data and enough compute came so we had enough data
because you know, we all started carrying multiple cameras around
with us. Our mobile phones have all you know, all
these cameras and this we put everything on the Internet,
and there's all this data out there. We caught a
lucky break that there was something called the graphics processing unit,
(08:21):
which turns out to be really useful for doing these
kinds of algorithms, maybe even more useful than it is
for doing graphics. They're greater graphics too, And things just
kept kind of adding to the snowball. So we had
deep learning, which is sort of a rebrand of neural
networks that I mentioned from the eighties, and that was
enabled again by data because we digitalized the world and
(08:43):
compute because we kept building faster and faster and more
powerful computers, and then that allowed us to make this
this big breakthrough. And then, you know, more recently, using
the same building blocks, that inexorable rise of more and
more and more data, that technology called self supervised learning.
Where the key difference there in traditional deep learning, you know,
(09:07):
for classifying images, you know, like is this a cat
or is this a dog? And a picture those technologies
require supper visions, so you have to take what you
have and then you have to label it. So you
have to take a picture of a cat and then
you label it as a cat, and it turns out that,
you know, that's very powerful, but it takes a lot
of time to label gats and to label dogs, and
(09:28):
there's only so many labels that exist in the world.
So what really changed more recently is that we have
self supervised learning where you don't have to have the labels.
We can just take unannotated data. And what that does
is it allows you use even more data. And that's
really what drove this this latest sort of rage. And
then and then all of a sudden we start getting
these these really powerful models. And then really, this has
(09:52):
been simmering technologies, right, this has been happening for a
while and progressively getting more and more powerful. One of
the things that really happened with CHATGBT and technologies like
Stable Diffusion and mid Journey was that they made it
visible to the public. You know, you put it out
(10:12):
there the public can touch and feel and they're like, wow,
not only is there palpable change, and wow this you know,
I can talk to this thing. Wow, this thing can
generate an image. Not only that, but everyone can touch
and feel and try. My kids can use some of
these AI art generation technologies, and that's really just launched.
(10:32):
You know, it's like a propelled slingshot at us into
a different regime. In terms of the public awareness of
these technologies.
Speaker 3 (10:39):
You mentioned earlier in the conversation foundation models, and I
want to talk a little bit about that. I mean,
can you just tell me, you know, what are foundation
models for AI and why are they a big deal?
Speaker 5 (10:52):
Yeah, So this term foundation model was coined by a
group at Stanford, and I think it's actually a really
apt term because remember I said, you know, one of
the big things that unlocked this latest excitement was the
fact that we could use large amounts of unannotated data.
We could train a model. We don't have to go
through the painful effort of labeling each and every example.
(11:16):
You still need to have your model do something you
wanted to do. You still need to tell it what
you want to do. You can't just have a model
that doesn't, you know, have any purpose. But what a
foundation models that provides a foundation, like a literal foundation.
You can sort of stand on the shoulders of giants.
You can have them these massively trained models, and then
do a little bit on top. You know, you could
use just a few examples of what you're looking for
(11:39):
and you can get what you want from the model,
So just a little bit on top now gets to
the results that a huge amount of effort used to
have to put in, you know, to get from the
ground up to that level.
Speaker 3 (11:50):
I was trying to think of of an analogy for
sort of foundation models versus what came before, and I
don't know that I came up with a good one,
but the best I could do was this. I want
you to tell me if it's plausible. It's like before
foundation models, it was like you had these sort of
single use kitchen appliances. You could make a waffle iron
(12:11):
if you wanted waffles, or you could make a toaster
if you wanted to make toast. But a foundation model
is like like an oven with a range on top.
So it's like this machine and you could just cook
anything with this machine.
Speaker 5 (12:24):
Yeah, that's a great analogy. They're very versatile. The other
piece of it, too, is that they dramatically lower the
effort that it takes to do something that you want
to do. And sometimes I used to say about the
old world of AI, would say, you know, the problem
with automation is that it's too labor intensive. H sounds
like I'm making a joke.
Speaker 3 (12:44):
Indeed, famously, if automation does one thing, it substitutes machines
or computing power for labor. Right, So what does that
mean to say AI is or automation is too labor intensive.
Speaker 5 (12:57):
It sounds like I'm making a joke, but I'm actually serious.
What I mean is that the effort it took the
old regime to automate something was very, very high. So
if I need to go and curate all this data,
collect all this data, and then carefully label all these examples,
that labeling itself might be incredibly expensive and time. So
(13:17):
and we estimate anywhere between eighty to ninety percent of
the effort it takes to feel an AI solution actually
is just spent on data, so that that has some consequences,
which is the threshold for bothering. You know, if you're
going to only get a little bit of value back
from something, are you going to go through this huge
effort to curate all this data and then when it
(13:40):
comes time to train the model, you need highly skilled
people expensive or hard to find in the labor market.
You know, are you really going to do something that's
just a tiny, little incremental thing. Now, you're going to
do the only the highest value things that weren't at
level because.
Speaker 3 (13:55):
You have to essentially build the whole machine from scratch,
and there aren't many things where it's worth that much
work to build a machine that's only going to do
one narrow thing.
Speaker 5 (14:06):
That's right, and then you tackle the next problem and
you basically have to start over. And you know, there
are some nuances here, like for images, you can pre
train a model on some other tasks and change it around.
So there are some examples of this like non recurring
cost that we have in the old world too, But
by and large, it's just a lot of effort. It's hard.
It takes, you know, a large level of skill to implement.
(14:30):
One analogy that I like is, you know, think about
it as you know, you have a river of data,
you know, running through your company or your institution. Traditional
AI solutions are kind of like building a dam on
that river. You know, dams are very expensive things to build.
They require highly specialized skills and lots of planning. And
you know, you're only going to put a dam on
(14:51):
a river that's big enough that you're gonna get enough
energy out of it that it was worth trouble. You're
gonna get a lot of value out of that dam.
If you have a river like that, you know, a
river of data, but it's actually the vast majority of
the water you know in your kingdom actually isn't in
that river. It's in puddles and greeks and babid brooks,
And you know, there's a lot of value left on
(15:14):
the table because it's like, well, I can't there's nothing
you can do about it. It's just that that's too
low value. So it takes too much effort, so I'm
just not going to do it. The return around investment
just isn't there, so you just end up not automating
things because it's too much of a pain. Now what
foundation models do is they say, well, actually, no, we
can train a base model a foundation that you can
(15:35):
work on that we don't we don't care. We don't
specify what the task is ahead of time. We just
need to learn about the domain of data. So if
we want to build something that can understand English language,
there's a ton of English language text available out in
the world. We can now train models on huge quantities
of it, and then it learned the structure. It learned
(15:56):
how language, you know, good part of how language works
on all that unlabeled data. And then when you roll
up with your task, you know, I want to solve
this particular problem, you don't have to start from scratch.
You're starting from a very very very high place. So
that just gives you the ability to just you know, now,
all of a sudden, everything is accessible. All the puddles
(16:16):
and greeks and babbling brooks and kettlepons, you know, those
are all accessible now. And that's that's very exciting. But
it just changes the equation on what kinds of problems
you could use AI to solve.
Speaker 3 (16:27):
And so foundation models basically mean that automating some new
task is much less labor intensive, The sort of marginal
effort to do some new automation thing is much lower
because you're building on top of the foundation model rather
than starting from scratch. Absolutely, so that is that is
like the exciting good news. I do feel like there's
(16:52):
a little bit of a countervailing idea that's worth talking
about here, and that is the idea that even though
there are these foundation models that are really powerful, that
are relatively easy to build on top of, it's still
the case, right that there is not some one size
fits all foundation model. So you know, what does that
mean and why is that important to think about in
(17:13):
this context?
Speaker 5 (17:14):
Yeah, so we believe very strongly that there isn't just
one model to rule them all. There's a number of
reasons why that could be true. One which I think
is important and very relevant today is how much energy
these models can consume. So these models, you know, can
get very very large. So one thing that we're starting
(17:39):
to see or starting to believe, is that you probably
shouldn't use one giant sledgehammer model to solve every single problem,
you know, like we should pick the right size model
to solve the problem. We shouldn't necessarily assume that we
need the biggest, baddest model for every little use case.
And we're also seeing that, you know, small models that
(17:59):
are trained to like to specialize on particular domains can
actually outperform much bigger models. So bigger isn't always even better.
Speaker 3 (18:07):
So they're more efficient and they do the thing you
want them to do better as well.
Speaker 5 (18:12):
That's right. So Stanford, for instance, a group of Stanford
trained a model. It is a two point seven billion
parameter model, which isn't terribly big by today's standards. They
trained it just on the biomedical literature, you know, this
is the kind of thing that universities do, and what
they showed was that this model was better at answering
questions about the biomedical literature than some models that were
(18:33):
one hundred billion parameters, you know, many times larger. So
it's a little bit like you know, asking an expert
for help on something versus asking the smartest person. You know,
the smartest person you know may be very smart, but
they're not going to be expertise. And then as an
added bonus, you know, this is now a much smaller model,
it's much more efficient to run. We are you know,
(18:54):
you know, it's cheaper. So there's lots of different advantages there.
So I think we're going to see attention in the
industry between vendors that say, hey, this is the one,
you know, big model, and then others that say, well, actually,
you know, there's there's you know, lots of different tools
we can use that all have this nice quality that
we outlined at the beginning, and then we should really
(19:15):
pick the one that makes the most sense for the
task at hand.
Speaker 3 (19:19):
So there's sustainability basically efficiency, another kind of set of
issues that come up a lot with AI A are bias, hallucination.
Can you talk a little bit about bias and hallucination,
what they are and how you're working to mitigate those problems.
Speaker 5 (19:34):
Yeah, so there are lots of issues still as amazing
as these technologies are, and they are amazing, let's let's
be very clear, lots of great things we're going to
enable with these kinds of technologies. Bias isn't a new problem.
So you know, basically we've seen this since the beginning
of AI. If you train a model on data that
(19:55):
has a bias in it, the model is going to
recapitulate that bias when it provides it's answers. So every time,
you know, if all the text you have says, you know,
it's more likely to refer to female nurses and male scientists,
then you're going to you know, get models that you know.
For instance, there was an example where a machine learning
based translation system translated from Hungarian to English. Hungarian doesn't
(20:17):
have gendered pronouns. English does, and when you ask them
to translate, it would translate they are a nurse to
she is a nurse, translate they are a scientist. To
he is a scientist. And that's not because the people
who wrote the algorithm were building in bias and coding
in like, oh, it's got to be this way. It's
because the data was like that. You know, we have
biases in our society and they're reflected in our data
(20:40):
and our text and our images everywhere. And then the
models they're just mapping from what they what they've seen
in their training data to to the result that you're
trying to get them to do and to give, and
then these biases come out. So there's a very active
program of research and you know, we do quite a
bit at IBM research and my T but also all
(21:03):
over the community and industry and academia trying to figure
out how do we explicitly remove these biases, how do
we identify them, how do you know, how do we
build tools that allow people to audit their systems to
make sure they aren't biased. So this is a really
important thing. And you know, again this was here since
the beginning, you know, of machine learning and AI, but
(21:24):
foundation models and large language models and generative AI just
bring it into sharper even sharper focus because there's just
so much data and it's sort of building in baking
in all these different biases we have, so that that's
that's absolutely a problem that these models have. Another one
that you mentioned was hallucinations. So even the most impressive
(21:45):
of our models will often just make stuff up. You know,
the technical term that the field has chosen is hallucination.
To give you an example, I asked chat tbt to
create a biography of David Cox at IBM, and you know,
it started off really well. You know, the identified that
I was the director of the MNT IBM Watson and
(22:05):
said a few words about that, and then it proceeded
to create an authoritative but completely fake biography of me
where I was British, I was born in the UK,
I went to British university, you know universities in the UK.
I was professor the authority.
Speaker 3 (22:21):
Right, it's the certainty that that is weird about it, right,
It's it's dead certain that you're from the UK, et cetera.
Speaker 5 (22:28):
Absolutely, yeah, it has all kinds of flourishes like I
want awards in the UK. So yeah, it's it's problematic
because it kind of pokes a lot of weak spots
in our human psychology where if something sounds coherent, We're
likely to assume it's true. We're not used to interacting
with people who eloquently and authoritatively, you know, emit complete nonsense,
(22:52):
like yeah, you know, we can debate about that, but.
Speaker 3 (22:55):
Yeah, we can debate about that. But yes, it's the
sort of blive confidence throws you off when you realize
it's completely wrong.
Speaker 5 (23:03):
Right, that's right. And we do have a little bit
of like a great and powerful oz sort of vibe
going sometimes where we're like, well, you know, the AI
is all knowing and therefore whatever it says must be true.
But these things will make up stuff, you know, very aggressively,
and you know, you everyone can try asking it for
their their bio. You'll you'll get something that You'll always
(23:26):
get something that's of the right form, that has the
right tone. But you know, the facts just aren't necessarily there.
So that's obviously a problem. We need to figure out
how to close those gaps, fix those problems. There's lots
of ways we can use them much more easily.
Speaker 4 (23:40):
I'd just like to say, faced with the awesome potential
of what these technologies might do, it's a bit encouraging
to hear that even chat GPT has a weakness for
inventing flamboyant, if fictional versions of people's lives, and while
entertaining ourselves with chat GPT and mid journey is important,
the way lpeople use consumer facing chatbots and generative AI
(24:04):
is just fundamentally different from the way an enterprise business
uses AI. How can we harness the abilities of artificial
intelligence to help us solve the problems we face in
business and technology. Let's listen on as David and Jacob
continue their conversation.
Speaker 3 (24:21):
We've been talking in a somewhat abstract way about AI
in the ways it can be used. Let's talk in
a little bit more of a specific way. Can you
just talk about some examples of business challenges that can
be solved with automation, with this kind of automation we're
talking about.
Speaker 5 (24:39):
Yeah, so the really really, this guy's the limit. There's
a whole set of different applications that these models are
really good at. And basically it's a superset of everything
we used to use AI for in business. So you know,
the simple kinds of things are like, hey, if I
have text and I have product reviews, and I want
to be able to tell if these are positive or negative.
(25:00):
You know, like let's look at all the negative reviews,
so we can have a human look through them and
see what was up. Very common business use case. You
can do it with traditional deep learning based AI. So so
there's things like that that are you know, it's very
prosaic sort of we were already doing it. We've been
doing it for a long time. Then you get situations
that are that were harder for the old day. I like,
(25:21):
if I'm I want to impress something like I want
to I have like say I have a chat transcript,
Like a customer called in and they had a complaint.
They called back. Okay, now a new you know, a
person on the line needs to go read the old
transcript to catch up. Wouldn't it be better if we
could just summarize that, just condense it all down a
(25:41):
quick little paragraph. You know, customer called they we upset
about this, rather than having to read the blow by blow.
There's just lots of settings like that where summarization is
really helpful. Hey, you have a meeting and I'd like
to just automatically you know, have had that meeting or
that email or whatever. I'd like to just have a
condensed down so I can really quickly get to the
heart of the matter. These models are are really good
(26:02):
at doing that. They're also really good at question answering.
So if I want to find out, what's how many
vacation days do I have? I can now interact in
natural language with a system that can go and that
it has access to our HR policies, and I can
actually have a you know, a multi turn conversation where
I can, you know, like I would have with you know, somebody,
(26:22):
you know, actual HR professional or customer service representative. So
a big part, you know, of what this is doing
is it's it's putting an interface. You know, when we
think of computer interfaces, we're usually thinking about UI user
interface elements where I click on menus and there's buttons
and all this stuff. Increasingly, now we can just talk,
(26:44):
you know, you just in words. You can describe what
you want, you want to answer, ask a question, you
want to sort of command the system to do something,
rather than having to learn how to do that clicking buttons,
which might be inefficient, Now we can just sort of
spell it out.
Speaker 3 (26:57):
Interesting, right, the graphical user interface that we all sort
of default to, that's not like the state of nature, Right,
that's a thing that was invented and just came to
be the standard way that we interact with computers. And
so you could imagine, as you're saying, like chat essentially
chatting with the machine could could become a sort of
standard user interface, just like the graphical user interface, did
(27:20):
you know over the past several decades.
Speaker 5 (27:22):
Absolutely, And I think those kinds of conversational interfaces are
going to be hugely important for increasing our productivity. It's
just a lot easier if I if I have to
learn how to use a tool, or I don't have
to kind of have awkward, you know, interactions from the computer.
I can just tell it what I want and I
can understand it. Could you know, potentially even ask questions
back to clarify and have those kinds of conversations that
(27:45):
can be extremely powerful. And in fact, one area where
that's going to I think be absolutely game changing is
in code. When we write code. You know, programming languages
are a way for us to sort of match between
a very sloppy way of talking and the very exact
way that you need to command a computer to do
(28:05):
what you wanted to do. They're cumbersome to learn, they
can you know, create very complex systems that are very
hard to reason about. And we're already starting to see
the ability to just write down what you want and
AI will generate the code for you. And I think
we're just going to see a huge revolution of like
we just converse you and we can have a conversation
to say what we want, and then the computer can
actually not only do fixed actions and do things for us,
(28:29):
but it can actually even write code to do new things,
you know, and generate software itself. Given how much software
we have, how much craving we have for software, like
we'll never have enough software in our world. Uh, you know,
the ability to have AI systems as a helper in that,
I think we're going to see a lot of a
lot of value there.
Speaker 3 (28:48):
So if you if you think about the different ways
AI might be applied to business, I mean you've talked
about a number of the sort of classic use cases.
What are some of the more out there use cases.
What are some you know, unique ways you could imagine
AI being applied to business.
Speaker 5 (29:06):
You know, there's really disguised the limit. I mean, we
have one project that I'm kind of a fan of
where we actually were working with a mechanical engineering professor
at MIT working on a classic problem, how do you
build linkage systems which are like you imagine bars and
joints and overs.
Speaker 3 (29:24):
You know, the things that are building a thing, building
a physical machine of some kind.
Speaker 5 (29:29):
Like real like metal and you know nineteenth century just
old school industrial revolution. Yeah yeah, yeah, but you know
the little arm that's that's holding up my microphone in
front of me. Cranes, get build your buildings, you know,
parts of your engines. This is like classical stuff. It
turns out that you know, humans, if you want to
build an advanced system, you decide what like curve you
(29:50):
want to create, and then a human together with a
computer program, can build a five or six bar linkage
and then that's kind of where you top out it
because it gets too comp replicated to work more than that.
We built a generative AI system that can build twenty
bar linkages, like arbitrarily complex. So these are machines that
are beyond the capability of a human to design themselves.
(30:13):
Another example, we have an AI system that can generate
electronic circuits. You know, we had a project where we're
working where we're building better power converters which allow our
computers and our devices to be more efficient, save energy,
you know, less less carbone. But I think the world
around us has always been shaped by technology. If we
look around, you know, just think about how many steps
(30:35):
and how many people and how many designs went into
the table and the chair and the lamp. It's really
just astonishing. And that's already you know, the fruit of
automation and computers and those kinds of tools. But we're
going to see that increasingly be product also of AI.
It's just going to be everywhere around us. Everything we
touch is going to have been helped in some way
(30:56):
to get to you by.
Speaker 3 (30:58):
A you know, that is a pretty profound transformation that
you're talking about in business. How do you think about
the implications of that both for the sort of you know,
business itself and also for employees.
Speaker 5 (31:12):
Yeah, so I think for businesses this is gonna cut costs,
make new opportunities to like customers, you know, like there's
just you know, it's sort of all upside right like
for the for the workers, I think the story is
mostly good too. You know, like how many things do
you do in your day that you'd really rather not?
Speaker 2 (31:33):
Right?
Speaker 5 (31:34):
You know, and we're used to having things we don't
like automated away. You know, we we didn't. You know,
if you didn't like walking many miles to work, then
you know, like you can have a car and you
can drive there. Or we used to have a huge
fraction over ninety percent of the US population engaged in agriculture,
and then we mechanized it. How very few people work
in agriculture, a small number of people can do the
(31:54):
work of a large number of people. And then you know,
things like email, and yeah, they've led to huge productor
the enhancements because I don't need to be writing letters
and sending them in the mail. I can just instantly
communicate with people. We just become more effective, Like our
jobs have transformed, whether it's a physical job like agriculture
or whether it's a knowledge worker job where you're sending
(32:16):
emails and communicating with people and coordinating teams. We've just
gotten better. And you know, the technology has just made
us more productive. And this is just another example. Now,
you know, there are people who worry that you know,
will be so good at that that maybe jobs will
be displaced, and that's a legitimate concern. But just like
(32:36):
how in agriculture, you know, it's not like suddenly we
had ninety percent of the population unemployed. You know, people
transitioned to other jobs. And the other thing that we
found too, is that our appetite for doing more things
is as humans is sort of insatiable. So even if
we can dramatically increase how much one human can do,
(32:58):
that doesn't necessarily mean we're going to do fixed amount
of stuff. There's an appetite to have even more, so
we're going to you can continue to grow, grow the pie.
So I think at least certainly in the near term,
you know, we're going to see a lot of drudgery
go away from work. We're going to see people be
able to be more effective at their jobs. You know,
we will see some transformation in jobs and like we've
(33:19):
seen that before, and the technology least has the potential
to make our lives a lot easier.
Speaker 3 (33:27):
So IBM recently launched Watson X, which includes Watson X
dot AI. Tell me about that, Tell me about you
know what it is and the new possibilities that it
opens up.
Speaker 5 (33:38):
Yeah. So so Watson X is obviously a bit of
a new branding on the Watson brand.
Speaker 2 (33:46):
T J.
Speaker 5 (33:47):
Watson that was the founder of IBM and our EI
technologies have had the Watson brand. Watson X is a
recognition that that there's something new, there's something that actually
has changed the game.
Speaker 3 (33:59):
You know.
Speaker 5 (33:59):
We gone from this old world of automation is to
labor intensive to this new world of possibilities where it's
much easier to use AI. And what watsonex does it
brings together tools for businesses to harness that power. So
whatsonex dot AI foundation models that our customers can use.
(34:21):
It includes tools that make it easy to run, easy
to deploy, easy to experiment. There's a watsonex dot Data
component which allows you to sort of organize and access
to your data. So what we're really trying to do
is give our customers a cohesive set of tools to
harness the value of these technologies and at the same
(34:43):
time be able to manage the risks and other things
that you have to keep an eye on in an
enterprise context.
Speaker 3 (34:50):
So we talk about the guests on this show as
new creators, by which we mean people who are creatively
applying technology in businesiness to drive change. And I'm curious
how creativity plays a role in the research that you do.
Speaker 5 (35:08):
I Honestly, I think the creative aspects of this job,
this is what makes this work exciting. You know, I
should say, you know, the folks who work in my
organization are doing the creating, and I guess you're.
Speaker 3 (35:24):
Doing the managing so that they could do the creator.
Speaker 5 (35:27):
I'm helping them be their best, and I still get
to get involved in the weeds of the research as
much as I can. But you know, there's something really
exciting about inventing, you know, like one of the nice
things about doing invention and doing research on AI. In industries,
it's usually grounded and a real problem that somebody is having.
(35:48):
You know, a customer wants to solve this problem. It's
losing money, or there wuld be a new opportunity. You
identify that problem and then you build something that's never
been built before to do that. And I think that's
honestly the adrenaline rush that keeps all of us in
this field. How do you do something that nobody else
(36:08):
on earth has has done before or tried before, so
that that kind of creativity and there's also creativity as
well and identifying what those problems are, being able to
understand the places where the technology is close enough to
solving a problem and doing that matchmaking between problems that
(36:28):
are now solvable, you know, and an AI where the
field is moving so fast, this is constantly growing horizon
of things that we might be able to solve. So
that matchmaking, I think, is also a really interesting creative problem.
So I think I think that's that's that's why it's
so much fun. And it's a fun environment we have
here too. It's you know, people drawing on whiteboards and
(36:50):
writing on pages of math and like in a movie,
like in a movie, yes, straight from sexual casting.
Speaker 3 (36:58):
Drawing, the drawing on the window, writing on the window,
and sharp absolutely so, so let's close with the really
long view. How do you imagine AI and people working
together twenty years from now?
Speaker 5 (37:16):
Yeah, it's really hard to make predictions. The vision that
I I like, actually this came from an MIT economist
named David Ottur, which was imagine AI almost as a
natural resource. You know, we know how natural resources work, right,
(37:38):
Like there's an or we can dig up out of
the earth that comes from kind of springs from the earth,
or we usually think of that in terms of physical stuff.
With AI, you can almost think of it as like
there's a new kind of abundance potentially twenty years from now,
where not only can we have things we can build
or eat or use or burn or whatever. Now we have,
you know, this ability to do things and understand things
(37:59):
and do intellectual work, and I think we can get
to a world where automating things is just seamless. We're
surrounded by capability to augment ourselves to get things done.
And you could think of that in terms of like, oh,
that's going to displace our jobs, because eventually the AI
system is going to do everything we can do. But
(38:19):
you could also think of it in terms of like, wow,
that's just so much abundance that we now have, and
really how we use that abundance is sort of up
to us, you know, like when you can writing software
is super easy and fast and anybody can do it.
Just think about all the things you can do now, like,
think about all the new activities and go about all
the ways we could use that to enrich our lives.
(38:40):
That's where I'd like to see us in twenty years.
You know, we can we can do just so much
more than we were able to do before abundance.
Speaker 3 (38:50):
Great, thank you so much for your time.
Speaker 5 (38:53):
Yeah, It's been a pleasure. Thanks for inviting me.
Speaker 4 (38:57):
What a far ranging, deep conversation. I'm mesmerized by the
vision David just described. A world where natural conversation between
mankind and machine can generate creative solutions to our most
complex problems. A world where we view AI not as
our replacements, but as a powerful resource we can tap
(39:17):
into and exponentially boost our innovation and productivity. Thanks so
much to doctor David Cox for joining us on smart Talks.
We deeply appreciate him sharing his huge breadth of AI
knowledge with us and for explaining the transformative potential of
foundation models in a way that even I can understand.
(39:39):
We eagerly await his next great breakthrough. Smart Talks at
IBM is produced by Matt Romano, David jaw nishe Venkat
and Royston Preserve with Jacob Goldstein. We're edited by Lydia
Jean Kott. Our engineers are Jason Gambrel, Sarah Bouguer and
Ben Holliday. Theme song by Gramma's Scope. Special thanks to
(40:01):
Carli Megliori, Andy Kelly, Kathy Callahan and the Eight Bar
and IBM teams, as well as the Pushkin marketing team.
Smart Talks with IBM is a production of Pushkin Industries
and iHeartMedia. To find more Pushkin podcasts, listen on the
iHeartRadio app, Apple Podcasts, or wherever you listen to podcasts.
(40:23):
Him Malcolm Gladwell. This is a paid advertisement from IBM.
Stuff To Blow Your Mind News
Hosts And Creators
Robert Lamb
Joe McCormick
Popular Podcasts
Dateline NBC
Current and classic episodes, featuring compelling true-crime mysteries, powerful documentaries and in-depth investigations. Follow now to get the latest episodes of Dateline NBC completely free, or subscribe to Dateline Premium for ad-free listening and exclusive bonus content: DatelinePremium.com
24/7 News: The Latest
The latest news in 4 minutes updated every hour, every day.
Therapy Gecko
An unlicensed lizard psychologist travels the universe talking to strangers about absolutely nothing. TO CALL THE GECKO: follow me on https://www.twitch.tv/lyleforever to get a notification for when I am taking calls. I am usually live Mondays, Wednesdays, and Fridays but lately a lot of other times too. I am a gecko.