The Truth about Artificial Intelligence (AI) with Dr.Alok Aggarwal (Author, Scientist, Entrepreneur)
October 3, 2024 • 52 mins
In this episode, we hosted a globally renowned and prolific guest Dr. Alok Aggarwal, founder, CEO, and Chief Data Scientist of Scry AI, the Author of the book "The Fourth Industrial Revolution & 100 Years of AI (1950-2050)" and an Inventor with 8 patents.
Dr. Aggarwal pioneered the concept of “Knowledge Process Outsourcing (KPO)”, “co-founded” Evalueserve (4000+, employees), “founded” IBM’s India Research Laboratory, founded Scry AI that builds proprietary AI products for enterprises globally.
He has published 125 research articles, taught 2 courses at the Massachusetts Institute of Technology (MIT), has a Ph.D from Johns Hopkins University and a B. Tech. from the Indian Institute of Technology (IIT) Delhi.
In this conversation with Pankaj, with insights drawn from his book, "Fourth Industrial Revolution in 100 Years of AI from 1950 to 2050," Dr. Alok presents a compelling argument for why AI is not just another technological trend but a catalyst for a new industrial revolution. He delves into the history of industrial revolutions to understand what makes AI stand out.
From steam engines to CPUs, each era has been marked by inventions that transformed societies. This episode offers a thorough analysis of how AI compares to these past innovations, while also cautioning against the hype that surrounds it.
He explains how AI's unique capabilities in classification, pattern recognition, and data processing are reshaping industries from banking and technology to healthcare and heavy engineering.
For entrepreneurs, the episode highlights the risks of getting caught up in AI hype without developing robust intellectual property and suggests strategies for creating high-value AI products.
In this podcast episode we spoke about the below topics, dive in:
03:55 - Historical Analysis of Industrial Revolutions
19:11 - The Impact of AI on Industries
34:05 - Navigating AI and Intellectual Property
45:27 - AI Transforming Services in India
Enjoyed the podcast? Please consider leaving a review on Apple Podcasts and subscribe wherever you are listening to this.
Follow Prime Venture Partners:
LinkedIn: https://www.linkedin.com/company/primevp/
Twitter: https://twitter.com/Primevp_in
This podcast is for you. Do let us know what you like about the podcast, what you don't like, the guests you'd like to have on the podcast and the topics you'd like us to cover in future episodes.
Please share your feedback here: https://primevp.in/podcastfeedback
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Dr.Alok Aggarwal (00:00):
Unfortunately, ai is again being hyped very,
very strong.
Businessweek, one of themagazines in the US, actually
had the front cover in 1984, Ibelieve it says AI.
It's finally here.
People get all excited andcreate hype cycles.
Ai machines are essentiallychai machines.
You train them just like youtrain your children to provide
(00:21):
learning.
Then you test them.
7-8% of the US economy easilyruns on COBOL and most of the US
Army Department of Defense runson COBOL.
So some of the nuclearsubmarines run on COBOL.
Pankaj Agrawal (00:34):
The GPT-40 is like a Ferrari for grocery
shopping.
Dr.Alok Aggarwal (00:37):
Mckinsey says it's about $15 trillion will
come out of it by 2030.
Everyone talks about China andUS.
People are forgetting thatIndia has the largest number of
software professionals right nowin the world.
Metaverse probably will occurmore like 15 to 20 years from
now.
Ai is not machine learning andmachine learning is not AI.
Pankaj Agrawal (01:02):
Hello everyone, my pleasure to welcome all of
you for the new episode of PrimePodcast.
The guest I have today issomeone that I've been really
looking forward to chat with.
He's an engineer and computerscientist.
He spent 15 plus years with IBMand did deep research at the
(01:22):
intersection of computer scienceoperations and data scientist.
He spent 15 plus years with IBMand did deep research at the
intersection of computer scienceoperations and data analytics.
He holds nine patents andseveral research publication.
He was also the foundingdirector of IBM research in
India office, which he helpedset up at IT Delhi, his alma
mater.
He later went on to foundE-ValueServe, a professional
(01:42):
services company that providesresearch in analytics, services
in intellectual property, legalprocesses, market and business
research and data analytics.
It went on to become a globalcompany with offices and
research centers in 15-pluscountries.
In his most recent role, he'scontributing meaningfully in AI
by founding Scry AI, whichprovides AI-based product
(02:03):
solutions and services in BFSItechnology and heavy engineering
and life sciences andhealthcare industries.
He's currently running thecompany as the CEO.
To top all of that, he recentlyauthored a book Demystifying AI
called Fourth IndustrialRevolution in 100 Years of AI
from 1950 to 2050.
(02:24):
Dr Alok, it's my pleasure towelcome you and I'm glad that we
are doing this conversation.
Thank you for having me Great.
So to kick this conversationoff right off the bat, dr Alok,
I'm reading your book and what Ifound very interesting is that
(02:44):
you mentioned that AI hasbrought about the fourth
industrial revolution and withthe past three, kind of
triggered by the discovery ofsteam engines, electrical motor
and, most importantly, cpus.
What are the characteristics ofthe industrial revolution and
why do you think AI has thepotential or is actually in the
midst of drawing out a fourthone?
What is so fundamentallydifferent about AI than the
(03:05):
previous revolutions?
Dr.Alok Aggarwal (03:08):
So every revolution, industrial
revolution, has its distinctness.
The interesting part, when Istarted writing the book is I
did not realize that nobody hadreally defined what an
industrial revolution was.
So actually, the first chaptersets up a framework for
industrial revolutions and whatare their characteristics.
(03:31):
How do we know that this is thefourth revolution, not the
fifth one or the third one, andso on.
So from that perspective, thefirst chapter actually sets up
eight characteristics of allindustrial revolutions that I
have seen, and I'm not saying byany means that these are all
the characteristics.
There may be more that I havenot discussed, or I do not even
(03:54):
know about.
The first one, typically in eachindustrial revolution has been
that it wasn't formed by oneinvention.
It was formed by severalinventions and they actually
often used each other to improve.
So an inventor or inventorswould use one invention to
(04:15):
improve another and so on.
So the first, for example, asyou said, the main invention was
the steam engine, which wasbeing used almost throughout,
and it became pervasive and verywell known.
So water and steaminfrastructure was created, and
what steam was obviously neededfor steam engines?
(04:36):
Steam also started as aninfrastructure.
Water and steam also startedfueling flower mills, for
example converting wheat toflour, textile mills and so on.
So the first characteristic ofall these inventions is that
there are several inventions.
I'll come to other revolutionsin a minute.
(04:56):
The second is that it at leastcreates one infrastructure for
the society.
In the first revolution it waswater and steam infrastructure.
In the second one it was withbecause of electricity or
electric motors.
Electric generation anddissemination or distribution
(05:18):
was the infrastructure andinterestingly, electricity not
only propelled or fueledelectric motors but also
electric bulbs and many otherthings which have nothing to do
with motors.
The third infrastructure againhad several inventions, but the
one which became theinfrastructure going forward was
(05:41):
electronic communication,wireless Wi-Fi and so on.
Before that, 1950, I mean 1995,most of us were using dialing,
using basically phone lines todial from home to the office,
which was extremely slow, andthat basically converted,
fundamentally changed that.
(06:01):
So those are the first three.
And the third characteristic isthat there is one invention
which basically becomespervasive during the entire
period and it becomes ubiquitous.
For example, steam enginesbegan to be used pretty much
everywhere in the 1800s, all theway from steam cranes or
(06:25):
railroads to steamboats,steamships and so on.
Second one electric motors.
There are about 3,500 differenttypes of electric motors.
They're being used everywhere.
In my room sitting here, I'msure there are at least six or
eight electric motors running inthe air fan, in the air
(06:45):
conditioning unit and so on.
And then there are five othercharacteristics which we can go
through in more detail.
But, going back to your question, this particular industrial
revolution has many inventions,so not only AI and inventions
(07:06):
related to data, but alsoinventions related to climate
change, related to Internet ofThings, robotics, gene editing,
personalized healthcare,blockchain, metaverse, ar, vr.
Not that all of these willhappen immediately.
In fact, most of theserevolutions were there for 40 to
(07:27):
80 years.
So we started here in 2011 forthe fourth industrial revolution
.
Probably it will continue till2050, if not 2060.
And some of these inventions,like Metaverse, probably will
occur more like 15 to 20 yearsfrom now, because we don't even
have the high-order broadbandcommunication or 60 wireless
(07:50):
communication.
But we can discuss the otherfive as we go along.
Why AI is different is that eachof these inventions, in each
revolution which becamepervasive, was markedly
different than the previous one.
For example, steam generatorsor steam engines were working
(08:11):
very differently than motors,and it moved from effectively
what I would call aresource-based infrastructure,
which is the resource was wateror steam to an electricity-based
infrastructure which wascreated by humans.
Electricity was being generatedby humans.
The third infrastructure wasquite different from the second
(08:35):
because the infrastructure wasthat in communication and there
was nothing being generated, soto speak.
It was just being communicatedfrom one end to the other.
And the fourth infrastructure,with AI, it's now about data,
which we can say it's very muchlike electricity or like steam
(08:57):
engines I mean steam or water orlike broadband but data has
many, many facets and it's verymuch different.
So, although people say data isthe new oil or like broadband,
but data has many, many facetsand it's very much different.
So, although people say data isthe new oil or data is the new
electricity, they're actuallyobfuscating a lot of the issues
(09:17):
or a lot of the important pointsthat data has.
And AI is, of course, usingdata to be the next pervasive
invention.
We already have people do notrealize often already have more
than 2,500 use cases of AI whichare operational in nature.
Everyone is today's hyped upabout LLMs and GPTs, but
(09:41):
actually there many, manyinstances where AI is already
being used and will be used by2050, my belief is at least in
100,000 use cases.
Pankaj Agrawal (09:54):
AI by design, what you can say a lot more
intelligent than the previousthree.
It has a lot more capabilitiesaround reasoning, about
contextual understanding, aboutvision understanding and basis,
which it can enable, as you said, right, multiple of these use
cases.
Right, so do you think, um,because of that and you know,
(10:14):
for example, many of the bigtech leaders whether it's on
that pitch I came out with thestatement, right, you know that
invention of ai is is as, uh, asinstrumental, if not more, than
electricity, than the inventionof electricity.
Right Now, that's a bigstatement.
Arguably, that there are vestedinterests and so on and so
forth, right, but do you kind ofbelieve that the acceleration
(10:39):
or the unlock of use casesthrough AI could be a lot more
pervasive than what the previousones have been?
Or do you think, you know, it'stoo early to kind of really
figure that out, because AI isfundamentally, you know, I would
(11:00):
say, meaningful intelligenceover the past three, right, so
do you think it would affect howit gets kind of seeped into the
mainstream?
Dr.Alok Aggarwal (11:11):
Yes, two aspects.
One is that, unfortunately, aiis again being hyped very, very
strongly, partly by varioustechnology companies, partly by
the media.
This is not the first time thatthis hype has happened, in fact
, when artificial intelligencewas initially created by Alan
(11:33):
Turing by talking about theimitation game.
After that there was aconference in 1956 in Dartmouth
College, and that's when thename was given.
And during 1956 to 1973, ai washyped equally strongly, maybe
more, maybe a bit less, butabout the same in many ways.
(11:54):
In fact, a movie was created,2001, a Space Odyssey, in which
the computer HAL 9000, isconsidered to be an artificial
general intelligence computerwith emotions, abilities to
scheme, arrogance, all the humancharacteristics and much more,
and it can beat people in chessand so on.
(12:16):
So, from that perspective, thatwas the first hype.
It went sour, it went bust in1973.
The second hype started in 1980with expert systems, and that's
a very important thing.
What is AI?
Ai, in effect, is a combinationof machine learning and expert
(12:37):
systems, expert systems beingknowledge-based, subject matter
expertise.
And even for humans, withoutsubject matter expertise, we are
not.
Without subject matterexpertise, we are not able to do
most of the work.
We become eventually knowledgeworkers in some field or the
other or are doing a basic worklike typing and so on.
So, from that perspective,1980s again there was a.
(13:00):
It was short-lived, it wentbust in 1987, 88.
And then this is now again thehype.
Interestingly, all hype cycleshave the same thing.
Businessweek, one of themagazines in US, actually had
the front cover in 1984, Ibelieve.
It says AI it's finally here.
And isn't that almost whateveryone is saying, including Mr
(13:24):
Pichai and others AI is finallyhere.
Having said that, all inventionsthat is the fourth
characteristic I bring out ofindustrial revolutions All
industrial revolutions havethese very important inventions,
but they take time to seep intothe society.
Things did not happen overnight.
(13:46):
For example, railroads.
The first railroad was createdin 1830 between Liverpool and
Manchester and people got soexcited.
Obviously steam engines weregetting pervasive.
People got so excited.
Inventors, investors and thenthe public, media, public they
all put in a lot of money.
By 1890, it was oversaturated,the hype had died down.
(14:10):
Most of the railroads went bust.
Same thing we see during thesecond industrial revolution
with telegraphs.
In the third industrialrevolution with telecoms and
dot-coms I mean the telecommoney.
About $500 billion was spent inbillion, not million.
$500 billion was spent in tolay cables in sea, under water,
(14:34):
that is, under soil, wirelessWi-Fi and many of these
companies went bust.
So it takes time for inventionsto seep into the society and of
course, the time is reducing butit is not going away.
For example, in 1882, the firstelectric generation plant was
(14:55):
created by Thomas Elber Edisonin upstate New York.
That was in 1882, but it wasonly in 1925, that is, about 43
years later, that half ofAmericans got electricity.
So 43 years later that half ofAmericans got electricity.
So 43 years later.
Similarly, even more recently,the first handheld phone was
(15:17):
created in Motorola by Cooper in1971.
But it was only in the 1990sthat people actually started
using handheld phones.
I'm just talking about not evensmartphones like BlackBerry.
Yeah, started using handheldphones.
I'm just talking about not evensmartphones like Blackberry, or
that is something which is veryimportant to keep in mind that
things take time to seep intosociety.
(15:39):
I give eight reasons why that'sthe case.
Everyone falls into that,including me.
When I started EvaluServe, wegave the notion of knowledge
process outsourcing and we saidin 11 years or 10 years, this is
how much it will be and Indiawill gain so much out of it.
The rest of the world will gainso much out of it.
Not that fundamentally we werewrong.
(16:02):
We were wrong about the time.
It took at least two times whatwe had projected, what I had
projected.
And that is the fifthcharacteristic of industrial
revolutions.
Fourth being things take timeto seep into the society.
Fifth being that people get allexcited and create hype cycles
which go boom in the bus.
Of course, many of these hypecycles are good.
(16:24):
I mean, if there was no hypecycle, we would probably have no
railroads.
Most of the railroads that wehave to do are because of the
hype cycle that got created inEngland and in the US, and
that's a good thing.
So I actually am a very strongbeliever.
It may hurt the investors, whomay lose their shirt, but it
(16:45):
helps the society in the longrun.
So definitely, ai will play abig role, probably a bigger role
than electricity, althoughthat's hard to say at this point
, but I think that will happen.
Gradually the pace will beginto pick up, but it will happen.
As I said, we clearly haveabout 2,500 use cases that I
(17:08):
know of and I'm putting up about1,000 of them by the end of
August on our website.
There are many more and I thinkit will go to 100,000.
So, from that perspective,there are only 3,500 motors
electric motors of differentkinds but there may be about
100,000 different use cases ordifferent type of AI systems.
(17:29):
So from that perspective, mrPichai may be right.
Pankaj Agrawal (17:34):
Excellent.
I think that's a good segue todiscuss.
What is the present state of AIright?
You mentioned about 2,500 usecases out there and, with your
experience at SCRI and theresearch for the book, what
sectors and industries are youmost excited about with respect
to AI adoption?
Maybe some examples of the usecases that are getting
(17:55):
accelerated or maybe altogetherreplaced by AI first approach,
if you want to pick yourfavorite out of those thousand.
Dr.Alok Aggarwal (18:02):
So, first of all, I think pretty much all of
the industries and departmentsof organizations, or you can say
divisions of organizations,will be affected by by here.
There is absolutely no doubt inmy mind that, if we excuse me
if we look at the longtrajectory I mean if we assume
(18:22):
that the revolution started infourth industrial revolution
started in 2011, with IBMwinning the Jeopardy contest my
feeling is we have at leastanother perhaps 10 to 15 more
years to go before AI reallyseeps deeply into it.
So, as part of those 1,000 usecases, pretty much I would say,
(18:47):
are broken up into 19 industries, one government with 20, and
then 10 different departments ordifferent divisions of a
typical organization finance andaccounting, procurement,
marketing and sales, and so onand so forth.
So, from a perspective ofcompanies, you can look at AI.
(19:10):
Ultimately, if we take on thehype, what can AI do today or in
the near future?
One thing it is very good at isclassifying, that is, it
differentiates between the facesof dogs and cats.
That's a very simple example invision.
But a more interesting exampleis to differentiate whether a
(19:31):
person has skin cancer or aperson should be given out a
loan.
So these are classification.
Secondly, it is very good atfiguring out patterns.
So otherwise, AI, as AlanTuring said, AI machines are
essentially child machines.
You train them, just like youtrain your children to provide
(19:52):
learning.
Then you test them and if thetest fails it doesn't have
enough accuracy to train more,or you rewrite the AI program.
So from that perspective,already there is a lot of
innovation, a lot of use casesin the Internet of Things, For
example, figuring out whethersomebody has a water leakage or
(20:15):
a gas leakage in his or herhouse.
There is about 4% water aroundthe world gets leaked in water
pipes as the water companiessend the purified water, the
potable water, from their plantsto homes.
So a lot of this is getting now, because the sensors are
(20:38):
relatively inexpensive, that alot of this is happening already
and will continue to happen inthe area of supply chain
management and in generalinternet, happen in the area of
supply chain management and ingeneral internet.
Clearly that would be one ofthe big fields, if I would call
a vertical which will getaffected.
Another one will be banking andfinance insurance.
(20:58):
Now, insurance has a veryinteresting problem, especially
life insurance, because it goeson for 40, 50 years and
therefore they have so much datawhich is in paper-based data
and that paper is not the factsof today very, very poor quality
(21:19):
.
So converting no matter.
Even humans cannot understandit, Forget about AI
understanding it.
So that's a very importantproblem and I think it will be
resolved in the next 10 to 12years.
Another problem which is alongthe same lines is reverse
engineering of COBOL programs.
So COBOL was a language whichwas created in 1960-61 and even
(21:44):
though it's a very simplelanguage, because it's a simple
language, people never wrotedocumentation on it.
And even though it's a verysimple language, because it's a
simple language, people neverwrote documentation on it.
And even today about 7 to 8percent of US economy easily
runs on COBOL.
So all of insurance companiestoday I mean India fortunately
got saved because COBOL by 1980shad become already an older
(22:09):
language and newer languages hadcome in Java and so on C++, C,
Java.
But COBOL for the Western world,for the more developed world,
is a very, very big problembecause people like me I mean I
learned COBOL in IT in 1975, onecourse when, when we did not
(22:30):
have even tapes, we had a deckof cards which we used to feed
into the computers and I meanpeople like me pretty much.
Most of them have retired, Someeven unfortunately passed away.
There is no documentation andwhat happens all the time is
that people modify their COBOLprograms and those have bugs in
(22:51):
them.
Nobody can now figure out whereare the bugs.
So just to give you an idea,most of the US Army Department
of Defense runs on COBOL.
(23:11):
So some of the nuclear basicallyreverse engineering COBOL
programs so that you don't needeffectively, you need very few
if any, COBOL programmers.
The other advantage of reverseengineering of COBOL program
using AI is because it doesn'tknow the syntax of COBOL
programmers.
The other advantage of reverseengineering of COBOL program
using AI is because it doesn'tknow the syntax of COBOL.
(23:32):
It would actually find all theerrors, or most of the errors
that are in the COBOL program,because it's not transliterating
COBOL program into a Python ora Java program but into a flow
chart.
So then we've got many caseswhich are very interesting.
I mean, healthcare is anotherone which will be fundamentally
(23:54):
changed in the next 10 to 14years.
But again the hype far exceedswhat reality is, which is the
sad part.
And right now I mean, on theother hand, as I write in the
book, hype is good, because ifwe put in 50 billion dollars in
(24:15):
in llms and gpts, it is good forthe humanity, the human society
, because something good wouldcome out.
Problems are hard problems,they're not easy.
Yeah, so you need a lot ofpeople in a lot of months.
Pankaj Agrawal (24:31):
Interesting, and I think that's a good segue to
the next set of topics that Ihad in mind.
I mean, given that you'reserving large industries of all
shapes and sizes first with EVL,you serve now with Scry where
are these companies in the AIjourney?
Because what we constantly hearis that a lot of pilot testing
is going on but barely anythinghas moved into production.
Sequoia came out with a recentreport that a lot of CapEx is
(24:55):
being spent in enablinggenerative AI and powering LLMs.
Kind of make up for that, right.
Google, meta, microsoft, allthese guys I think they recently
had their quarterly earningsand all of them are spending
anywhere from 15 to $20 billionevery year, every quarter, not
(25:18):
every year every quarter asCapEx.
A large chunk of that isobviously going into procuring
the GPUs, you know, obviously.
And then NVIDIA is obviouslyNVIDIA stock has been, you know,
touching the skies, but none ofthat has resulted into the kind
of revenue, right?
Or maybe the production usecases, right?
So what do you hear from yourcustomers, right?
(25:39):
I mean, where are they in thejourney, right Among?
the use cases which are kind ofalready proven.
Yeah.
Dr.Alok Aggarwal (25:43):
Yeah.
So again I go back to thecharacteristics of the
Industrial Revolution.
I said the fourthcharacteristic was that it takes
time for even the bestinventions to seep into the
society, the fifth being thatpeople get hyped up about it and
therefore you have a boom-bustcycle.
And we are in a very strongboom-bust cycle on Google Buzz
(26:08):
site, because the problem is,everyone has gotten excited
about one very, very specificaspect of AI, which is GPTs and
LLMs, large language models andgenerative pre-trained transform
.
The first paper in this regardcame from Google Research itself
.
It was a research paper calledAttention is All you Need, and
after that Google created BERT,which was the first transformer.
(26:32):
It wasn't that good because itdidn't have enough parameters,
so to speak, without going intodetails, but since then the race
, or the war, began.
Llms are very interesting forhumans because it can write me a
recommendation letter, it cansummarize very large pieces of
(26:54):
text for me, it can improve myEnglish.
They've been trained.
These particular deep learningmodels, or deep learning
networks, are trained onanywhere from 400 million pages
to a billion pages of text andtables and so on.
So they're very, very good inwriting English.
They're almost perfect in that.
(27:15):
Imagine a kid going throughthat many novels If a novel is
about 400 pages long that's alarge novel, maybe 200 pages
long, talking about 2 millionnovels Then either the kid will
go crazy, but also the kid willprobably become excellent in
writing English, very cogent,very proficient, and that's what
you see in LLMs also.
(27:38):
They're extremely cogent,they're extremely proficient,
they're so good that they caneven fool lawyers and they fool
lawyers into believing thatnon-existing cases exist and
therefore lawyers into believingthat non-existing cases exist.
So, just like I was talkingabout the kid, that the kid will
go crazy and will becomeproficient, probably that's
what's happening.
(27:59):
We don't know what's happeningwith machine learning models,
but probably that's what ishappening with machine learning
models.
They're both crazy and that'swhy they they hallucinate and
they are extremely good atwriting, which is actually in a
very interesting way or a sadway for humans, because they're
writing so well.
(28:19):
We trust them and that's whatwe call machine endearment, but
then they hallucinate.
So we haven't really found outactually good ways of using
these particular deep learningnetworks for large language
models and DPTs in real sense ofthe word.
Yes, it can write a very goodpoem for me.
(28:40):
It can elaborate something forme.
It can write a story, but can Iuse it in any meaningful manner
?
And I think this is whatSequoia is talking about.
I think this is what alsoGoldman Sachs is talking about.
This is what I write in chapter11 of the book.
Not that they will not be usedAgain.
It goes back to the industrialrevolution and its
(29:02):
characteristics.
Ten years down the road, theywill be commonplace.
We would have figured out, butright now we don't even have the
first, second, third-mileproblem, several-mile problem
solved for these GPTs.
For example, they're trained onInternet data, but a company
has its own data, which ispaper-based, pdf-based, excel
(29:23):
spreadsheets and so on.
The first mile is you have toconvert all this data so that a
large language model can adjust.
So those are the ones why Ithink Sequoia.
I think there was also a smallcomment or a small article from
Barclays.
Sequoia and Goldman Sachs hadbeen on the forefront.
(29:45):
In December 2023, when I wrotethe book in chapter 11, I said
look, more heist than reality,because McKinsey says it's about
15 trillion dollars will comeout of it by 2030 or it will
affect 15 trillion dollars.
Now the world's GDP will beonly 150 trillion, so you're
talking about it affecting 10%of the economy of the world.
(30:07):
I don't think that's even goingto be close to where we are.
Having said that, there aremany areas which we don't even
today.
We don't even see where AI is,for example.
So many not so many, but manyairports, including in the US,
have started using facerecognition to speed up the
(30:30):
immigration process.
So in Dubai, if you go, it'syour face that it checks and if
it's 98% correct, it'll just letyou go.
Literally, it reduces your timefrom 10 minutes to 10 seconds.
Singapore has done the same.
Us government, slightly morecautious, does it all for
specific people that it hasalready screened in the past.
(30:54):
Or the global services people Imean travel services people.
So we don't even talk aboutthese.
Very soon, a lot of these thebiometric recognition, whether
it's pupils, face recognitionhands in many countries not in
the US or Europe, but in manyother countries will be used as
(31:15):
a way to check out of a grocerystore.
Not in the US, because the USis still worried about privacy,
and so is particularly Europe.
Similarly, if you look inagriculture, there is a lot
going on in agriculture intrying to understand various
aspects how much nitrogen isthere?
How do I implant seeds in thedirt without literally opening
(31:40):
up the dirt, because the momentyou open up the dirt which has
been for the last 3,000 years ormaybe 5,000 years that you
basically use an animal whichhas a machine, a small machine
in the sense it has a small toolin it which will open the dirt,
and then you throw seeds andthat always gets carbon dioxide
(32:04):
into the air, that always losesmoisture into the air and, above
all, it loses topsoil to theseas.
So a lot is happening inagriculture, and agriculture, I
believe, in 25 years, willfundamentally change.
So there are areas where thingsare moving and moving very
rapidly.
One area which I'm particularlyinterested in, and particularly
(32:26):
because of Scribe, isintelligent document processing.
There is so much paper and PDFdocuments, but these are either
unstructured or semi-structuredin nature.
You need to convert it into anelectronic form to be able to
automatically use AI, to playwith it, to give decisions,
(32:47):
support and so on.
But unless you can get thatcompletely resolved with very
little human intervention, youwill not be able to solve this
problem.
And that I mean out of the1,000 use cases, there are about
120 use cases on intelligentdocument processing alone what
all can do or should be able todo.
(33:09):
Some of them are already in someform or the other, are already
being implemented, the simplestones.
Pankaj Agrawal (33:21):
Got it.
So our audience has a largechunk of our audience are
founders who are alreadybuilding or kind of thinking of
building a business right.
So what areas within AI right,Do you think a new entrant can
contribute to and, in theprocess, build a large business?
Right, I mean, what should theykeep in mind while serving
(33:44):
incumbents?
Dr.Alok Aggarwal (33:45):
right, I would suggest they look into three
areas very particularly andparticularly because generally
the entrepreneurs aretechnologists, they miss out on
these areas quite a bit.
One is to make sure that youstart small with a small problem
.
But the area effectively.
(34:05):
You know that there are manyadjacencies, so I'll start with
the use case.
But there are many use caseswhich are very close by and I
can solve them in the long run.
Now that doesn't mean I willsolve them on day one.
It may take a seven yearjourney, five year journey.
And, by the way, I mean, unlessyou're an entrepreneur, unless
you're planning on selling thecompany for 100 million or 50
(34:27):
million, most companies take 20to 25 years really to build.
So you should keep in mind asto which particular direction
you want to go, whether you'rein it for about half of your
working life or you're in it forone or one third of your
working life or only for fiveyears of it.
Both are perfectly fine.
I'm not advocating one versusanother, but that's something
(34:50):
which is very important.
Having said that, that theyshould keep adjacencies in mind,
that I'm going to look at thisuse cases, but use case to make
a minimum viable product, butthen I'll go and expand it,
because those are the use casesthat can expand to.
Second thing that they shouldbe very careful about is AI is
(35:12):
not machine learning and machinelearning is not AI.
There is no equivalence betweenAI is a superset of machine
learning.
As I said, the second AI boomand bust happened because of
expert systems.
No matter how much we put in,even in LLMs and GPTs, how much
we train them 400 million pages,500 billion words they still
(35:38):
require context.
They still require and humansdo.
By the way also, I mean I'm notinto accounting Somebody comes
to me and says can you figureout what is the total operating
income of a bank?
Here is the report and supposethat number is not given.
I may not know what are theformulas I will use because I
(35:59):
don't have the context andobviously AI suffers a lot from
that.
I mean GPTs and LLMs.
A child will suffer a lot, evena human will suffer a lot.
So I think it is very importantfor them to understand the
distinction between machinelearning and say look, I need to
add actually knowledge matterinto it, subject matter,
(36:24):
expertise.
So that is a very, veryfundamental thing.
My own view is 90 percent ofcompanies startups will fail
because they do not include thesubject matter expertise into
the entire story.
(36:44):
The third is not to get caughtup with the hype.
A lot of people are just takingeither open-source LLM models
or just open-source models,putting them together and
creating a minimum viableproduct out of it.
Great, you can do something,you can create a chatbot out of
it, and so on and so forth.
If you're a services company,it may all work out, also
(37:07):
because you'll move on to thenext project.
But if you're in a productcompany, then you don't have any
intellectual propertyassociated.
So how do you sell?
I'll give you an example ofintelligent document processing
which basically takes the data,converts I mean which is scanned
data or PDF data converts itinto electronic data.
(37:28):
But unlike humans, which havetwo eyes, the machine converts
it into literally one dimension.
It loses the context of tables,loses the context of graphs,
charts, etc.
They have the second problem,which is to recreate tables with
pretty much 100 percentaccuracy, graphs and charts.
Then suppose I come and tellyou that look, it's 90 percent
(37:52):
correct.
You're my client.
Your immediate reaction is okay, that's pretty good, it's 90%
correct.
You're my client, yourimmediate reaction is okay,
that's pretty good, it's 90%correct.
But this is a 300-page documentyou just converted.
Will I have to review it fromleft to right for all the 300
pages, because I don't knowwhere the 10% is wrong?
So the question is if I canreconcile, for example, various
(38:16):
suppose there is a table in itwhich is income state, I can
reconcile and I can show youthat look or not me.
But the software says look inthis statement, a plus B plus C
equals D.
Then you don't have to reviewit.
So I can very, very clearlypoint out which are the issues,
(38:41):
where are the issues, and so onand so forth.
This is just one example thatyou add, because everyone I mean
there are 57 or 58 companiesthat we know of.
There may be more companies inintelligent document processing,
but none of them.
Everyone will say oh, you willneed a human in the loop.
Not clear that you will need ahuman in the loop.
In humans, yes, we have a makerand a checker, but it's not
(39:05):
that we use checker all the timebecause checker is expensive.
It's our two eyes which solvethe problem.
With my right eye closed, Ilook at you and I say it's
punctured.
With my left eye closed, I lookat you and say it's punctured.
The error of my being wrong issquared, because if my right eye
was 90% correct, error was 10%.
(39:26):
My left eye was 90% correct,error was 10%, and if both eyes
were working independent of eachother, the error becomes
squared.
That is only 1% wrong 10%.
Wrong there 10%.
So what I'm trying to say isthere are interesting areas that
people can go into.
I'm not saying everyone will,but I'm saying that in product
(39:50):
business you have to haveintellectual property that you
can defend and unfortunately Ido not see that right now.
This is one of the things wherepeople are rushing.
It's more than a gold rush I'llcall it a platinum or a diamond
rush that people are notrushing to figure out what the
issues are and how to solve them.
(40:11):
Rajesh.
Pankaj Agrawal (40:12):
KASTURIRANGANANI , could data be that
intellectual property?
So one common theme, or ratherthe playbook and I wrote about
it on LinkedIn recently that aplaybook is, as you said
identify a painful and a usecase, build a superior product
to solve that.
That becomes your sort of footin the door product generate
demand, bring along customers.
(40:34):
Doing that In the process,hopefully you are generating
some high-quality andproprietary data which you can
use to kind of keep on makingthe product better.
Do you think the data itselfcould be that IP?
Dr.Alok Aggarwal (40:47):
I don't think data will be the IP.
I think use of the right datais part of the design knowledge
and that's where subject matterexpertise comes in.
So what data will I use?
Because there is a lot of dataaround, but a lot of data is
also noisy and you can't use it.
Pankaj Agrawal (41:09):
Quality of data.
Dr.Alok Aggarwal (41:14):
So that goes back to subject matter expertise
.
What data do I use?
How do I get to creating aproduct which is superior?
And when I say superior, it hasintellectual property which is
hard for people to overcome.
The sad part about AI is or, ingeneral, software is yeah, you
(41:37):
can write patents, but they'reworth not even the paper they're
written on.
You give me a patent insoftware, in AI, because it's
math.
Eventually I can go around itfairly easy.
It's not like building acompany, and I know you folks
are in the venture capitalbusiness business and I know
many VCs who get all excitedabout patents.
(41:58):
Even though we could probablyin our company write about 30 to
40 patents, we have not writtena single one because we know
the moment we write a patent, wehave essentially given our
intellectual property away.
Somebody will just go around it.
Yeah, it's like reconciliationof intelligent document
processing.
(42:18):
We have not written a singlepatent on how do we reconcile?
How does the system learnformulas which are there in a
PDF document automatically, orhow does the formula get learned
in a football program?
Pankaj Agrawal (42:38):
I think rightly so.
Another theme which I agreewith that you pointed out that
subject matter expertise will berelevant.
So if you see, now Lama came upwith 405 billion parameters and
it is kind of giving a chase toGPT-4.0.
And a lot of experts are kindof saying that GPT 4.0 is like a
Ferrari for grocery shopping.
(42:58):
Right, it's just too over thetop for most of the things.
Right Versus open source modelthat you can access, you can
utilize, you can train on yourspecific use case, make it like
you know, work really reallywell and accurately for that
domain.
So that will be the future andthat's where subject matter
expertise becomes more and morekind of relevant.
(43:19):
I personally feel that modelsthemselves are probably the
fastest depreciating asset inhistory, probably.
Dr.Alok Aggarwal (43:29):
In some sense they're not even depreciating
because they become open sourcevery quickly.
Like Lama is open source, Imean Mixtral is open source.
I can choose a bunch of opensource models and then I can put
them together.
Now, what is intellectualproperty is?
Lama has, let's say, in aparticular metric, 90% accuracy,
mixtral has 88%, and all ofthese are open source.
(43:52):
Can I now put them together ina meaningful way so that my
accuracy becomes 92 percent inthe whole process?
That becomes my intellectualproperty.
Now, if I put subject matterexpertise on top of it, then I
can get to 95 96 percent andthen if I can reconcile
everything and show you thatthere is no hallucination,
(44:13):
because I'm showing you wherethe where the particular thing
got its answer from, then youare a happy person as a client,
right?
I mean then that you say look,you have really saved me time,
money, cost etc.
And above all, human labor.
So I think these things willtake time.
They always.
And that's why chapter one ofthe book is so important,
(44:35):
because it sets up a frameworkthat, look, let's not get
carried away.
I mean it's perfectly fine ifwe get carried away.
Also, because I do say thathype, boom bus cycles are
actually good for society.
I mean, we saw that withdriverless cars recently.
About $100 billion was spent inthere.
Yeah, some amount of it reallywent into research and
(44:58):
eventually we will solve thedriverless car problem.
It may be eventually, maybe 10to 15 years from now.
So I'm not even saying thatcycles are bad cycles For the
human society.
They are actually probably goodthings to have.
But investors and inventors ofcourse, they lose their shirt,
(45:21):
so they feel bad.
Yeah, that's it.
Pankaj Agrawal (45:26):
This has been super interesting.
I have one last question thatI've been kind of trying to
unpack for the last couple ofmonths as I go deeper into AI.
I think AI, with its power to,you know, either as a co-pilot
or you know, kind of make theoverall system more efficient,
has the ability to kind of addmeaningful value in services
(45:50):
businesses right, and there isan amazing concept that, rather
than software as a service,there could be certain areas
where service as a software hasthe right characteristics of a
venture scale business producingventure scale business in terms
of capital, efficient growth,velocity of growth, and so on
and so forth.
Given that you have experiencein both now with E-valueServe
(46:12):
with respect to KPO, and nowwith Scribe, you are
buildingalueServe, you know withrespect to KPO, and now with
Scry you are kind of buildingproducts for these customers,
large customers what is yourview?
I mean, if you were to, let'ssay, think of starting
E-ValueServe in today's world,would you do anything
differently?
You mentioned that it took youtwice the time you projected to
get to the scale right, or wheredo you think?
How should an investor thinkabout the services space
(46:34):
generally and how can it bepowered by that?
Dr.Alok Aggarwal (46:37):
Actually, after writing the book, I
decided this year to convert twothings To convert the book into
36 lectures of 45 minutes each,because not everyone reads
books these days, very fewpeople read books.
And by no means the book issmall.
It's actually 270 pages offairly dense material.
(46:57):
It's not intense because itdoesn't have math or computer
science, but it is dense.
You have to think about it,read it once more before it
begins to seep into your intothe brain, into your thought
process.
So, uh, so that's one.
And the second is to write atleast eight articles on India
(47:17):
and AI.
Because everyone talks aboutChina and US.
People are forgetting thatIndia has the largest number of
software professions right nowin the world 30 million
developers out of 30 million.
And that is 5.5 million softwaredevelopers actually working on
a daily basis in all thesecompanies.
(47:38):
And we're not talking aboutHTML and CSS developers, we're
talking about softwareengineering developers.
I mean, the world has about 25million and India has about 5.4.
So people forget about that andthat is going to be a very,
very important aspect aboutIndia's growth in AI which
(47:59):
people don't realize.
Having said that, so my firstchapter was first article was
the unsung heroes of AI how dataannotation will grow enormously
, meteorically in India.
The second one was, of course,loneliness in India will rise
because of AI.
(48:19):
The third one is that theservices industry in AI will be
transformed radically by AI inthe next 10 years.
Because you talked aboutEvaluServe.
When I think about EvaluServeall the time, because I'm still
invested in it fairly heavily,even though I'm on the board,
and if I were to redo it, Iwould just completely redo it
(48:42):
with software first and humanslater.
And about 50% of the work thatEvaluServe does and, to me,
about 70% of the work that TCSand others do, especially in the
BPO domain, like GenPan, likeeXcel service, 70% of the work
(49:03):
is within the next three to fiveyears can be reduced by a
factor of two which is a lotright, Because suddenly clients
begin to realize that you'vereduced it by a factor of two,
they'll say why are you chargingme X when you should be
charging me X over two?
And then you say, okay, there isa software cost.
(49:25):
They'll still pay you maybe 0.6, 0.7 times X, so everyone gains
in the whole.
Pankaj Agrawal (49:32):
And this it can also expand the overall demand.
Right, the Jevons paradox willkick in.
You reduce the cost oftechnology per unit consumption,
the demand will grow up.
Dr.Alok Aggarwal (49:43):
Absolutely, because second one is it will go
pay by the drink, not by theFTE, so to speak.
So you will go into, as yourightly said, small,
medium-sized businesses whocannot afford any Excel or any
value serve today will be ableto afford the next generation of
companies.
So I think, this is going to bea very, very fundamental change.
(50:05):
The only good news and thispaper should come out in
September the only good news Isee is that it is not again not
going to happen in the nextthree years.
It may take seven to 10 yearsfor this change to occur, but
this change is imminent.
This change is.
I mean I hope TCS, infosys etcare waking up to this change.
(50:26):
It's not that these people willgo away, but I mean you will
still have 50 percent.
You'll have people who will goupstream and so on.
India may actually gain a lot.
It is an opportunity for Indiato really play very strongly in
the industrial revolution,because the foundation is set,
because India has so manysoftware engineers, because
(50:48):
India can produce enormousnumber of data annotators who
can check for ground truth, whocan basically do data annotation
, who can tell you that look,this part is this portion,
because in supervised learningwe need data annotation.
Pankaj Agrawal (51:08):
Fantastic.
Thanks a lot, Really reallyappreciated doing this
conversation.
Of course, we'll recommend theaudience to check out your book.
We'll link it in the show notesand thanks for taking the time
to chat with us.
Dr.Alok Aggarwal (51:22):
Thank you, Pankaj.
Thank you so much.
Thanks, Jerome, for making itwork all of it.
Thanks, I'll write to youseparately.
Pankaj Agrawal (51:30):
Sir, I'll take your details from Jerome.
I would love for you to talk toEmpirical and at some point in
time, of course we would wanthim to spend some time in the US
, so of course we would love foryou guys to.
I mean, he's from your almamater, so that's alma mater,
building something in AI.
So I think two strong reasonsfor you to take the time out and
(51:51):
have a chat, and I think itwill be helpful for him as well.
Yeah, sure, absolutely All right, have a great day, sir.
Thank you.
Thank you Bye, bye.
Prime Venture Partners (52:08):
Dear listeners, thank you for
listening to this episode of thepodcast.
Subscribe now on your favoritepodcast app for free and you'll
be the first one to know whennew episodes are available.
(52:29):
Thank you, we would be reallygrateful if you leave us a
review on Apple Podcast.
To read the full transcript,find the link in the show notes.
Unfortunately, ai is again being hyped very,
very strong.
Businessweek, one of themagazines in the US, actually
had the front cover in 1984, Ibelieve it says AI.
It's finally here.
People get all excited andcreate hype cycles.
Ai machines are essentiallychai machines.
You train them just like youtrain your children to provide
(00:21):
learning.
Then you test them.
7-8% of the US economy easilyruns on COBOL and most of the US
Army Department of Defense runson COBOL.
So some of the nuclearsubmarines run on COBOL.
Pankaj Agrawal (00:34):
The GPT-40 is like a Ferrari for grocery
shopping.
Dr.Alok Aggarwal (00:37):
Mckinsey says it's about $15 trillion will
come out of it by 2030.
Everyone talks about China andUS.
People are forgetting thatIndia has the largest number of
software professionals right nowin the world.
Metaverse probably will occurmore like 15 to 20 years from
now.
Ai is not machine learning andmachine learning is not AI.
Pankaj Agrawal (01:02):
Hello everyone, my pleasure to welcome all of
you for the new episode of PrimePodcast.
The guest I have today issomeone that I've been really
looking forward to chat with.
He's an engineer and computerscientist.
He spent 15 plus years with IBMand did deep research at the
(01:22):
intersection of computer scienceoperations and data scientist.
He spent 15 plus years with IBMand did deep research at the
intersection of computer scienceoperations and data analytics.
He holds nine patents andseveral research publication.
He was also the foundingdirector of IBM research in
India office, which he helpedset up at IT Delhi, his alma
mater.
He later went on to foundE-ValueServe, a professional
(01:42):
services company that providesresearch in analytics, services
in intellectual property, legalprocesses, market and business
research and data analytics.
It went on to become a globalcompany with offices and
research centers in 15-pluscountries.
In his most recent role, he'scontributing meaningfully in AI
by founding Scry AI, whichprovides AI-based product
(02:03):
solutions and services in BFSItechnology and heavy engineering
and life sciences andhealthcare industries.
He's currently running thecompany as the CEO.
To top all of that, he recentlyauthored a book Demystifying AI
called Fourth IndustrialRevolution in 100 Years of AI
from 1950 to 2050.
(02:24):
Dr Alok, it's my pleasure towelcome you and I'm glad that we
are doing this conversation.
Thank you for having me Great.
So to kick this conversationoff right off the bat, dr Alok,
I'm reading your book and what Ifound very interesting is that
(02:44):
you mentioned that AI hasbrought about the fourth
industrial revolution and withthe past three, kind of
triggered by the discovery ofsteam engines, electrical motor
and, most importantly, cpus.
What are the characteristics ofthe industrial revolution and
why do you think AI has thepotential or is actually in the
midst of drawing out a fourthone?
What is so fundamentallydifferent about AI than the
(03:05):
previous revolutions?
Dr.Alok Aggarwal (03:08):
So every revolution, industrial
revolution, has its distinctness.
The interesting part, when Istarted writing the book is I
did not realize that nobody hadreally defined what an
industrial revolution was.
So actually, the first chaptersets up a framework for
industrial revolutions and whatare their characteristics.
(03:31):
How do we know that this is thefourth revolution, not the
fifth one or the third one, andso on.
So from that perspective, thefirst chapter actually sets up
eight characteristics of allindustrial revolutions that I
have seen, and I'm not saying byany means that these are all
the characteristics.
There may be more that I havenot discussed, or I do not even
(03:54):
know about.
The first one, typically in eachindustrial revolution has been
that it wasn't formed by oneinvention.
It was formed by severalinventions and they actually
often used each other to improve.
So an inventor or inventorswould use one invention to
(04:15):
improve another and so on.
So the first, for example, asyou said, the main invention was
the steam engine, which wasbeing used almost throughout,
and it became pervasive and verywell known.
So water and steaminfrastructure was created, and
what steam was obviously neededfor steam engines?
(04:36):
Steam also started as aninfrastructure.
Water and steam also startedfueling flower mills, for
example converting wheat toflour, textile mills and so on.
So the first characteristic ofall these inventions is that
there are several inventions.
I'll come to other revolutionsin a minute.
(04:56):
The second is that it at leastcreates one infrastructure for
the society.
In the first revolution it waswater and steam infrastructure.
In the second one it was withbecause of electricity or
electric motors.
Electric generation anddissemination or distribution
(05:18):
was the infrastructure andinterestingly, electricity not
only propelled or fueledelectric motors but also
electric bulbs and many otherthings which have nothing to do
with motors.
The third infrastructure againhad several inventions, but the
one which became theinfrastructure going forward was
(05:41):
electronic communication,wireless Wi-Fi and so on.
Before that, 1950, I mean 1995,most of us were using dialing,
using basically phone lines todial from home to the office,
which was extremely slow, andthat basically converted,
fundamentally changed that.
(06:01):
So those are the first three.
And the third characteristic isthat there is one invention
which basically becomespervasive during the entire
period and it becomes ubiquitous.
For example, steam enginesbegan to be used pretty much
everywhere in the 1800s, all theway from steam cranes or
(06:25):
railroads to steamboats,steamships and so on.
Second one electric motors.
There are about 3,500 differenttypes of electric motors.
They're being used everywhere.
In my room sitting here, I'msure there are at least six or
eight electric motors running inthe air fan, in the air
(06:45):
conditioning unit and so on.
And then there are five othercharacteristics which we can go
through in more detail.
But, going back to your question, this particular industrial
revolution has many inventions,so not only AI and inventions
(07:06):
related to data, but alsoinventions related to climate
change, related to Internet ofThings, robotics, gene editing,
personalized healthcare,blockchain, metaverse, ar, vr.
Not that all of these willhappen immediately.
In fact, most of theserevolutions were there for 40 to
(07:27):
80 years.
So we started here in 2011 forthe fourth industrial revolution
.
Probably it will continue till2050, if not 2060.
And some of these inventions,like Metaverse, probably will
occur more like 15 to 20 yearsfrom now, because we don't even
have the high-order broadbandcommunication or 60 wireless
(07:50):
communication.
But we can discuss the otherfive as we go along.
Why AI is different is that eachof these inventions, in each
revolution which becamepervasive, was markedly
different than the previous one.
For example, steam generatorsor steam engines were working
(08:11):
very differently than motors,and it moved from effectively
what I would call aresource-based infrastructure,
which is the resource was wateror steam to an electricity-based
infrastructure which wascreated by humans.
Electricity was being generatedby humans.
The third infrastructure wasquite different from the second
(08:35):
because the infrastructure wasthat in communication and there
was nothing being generated, soto speak.
It was just being communicatedfrom one end to the other.
And the fourth infrastructure,with AI, it's now about data,
which we can say it's very muchlike electricity or like steam
(08:57):
engines I mean steam or water orlike broadband but data has
many, many facets and it's verymuch different.
So, although people say data isthe new oil or like broadband,
but data has many, many facetsand it's very much different.
So, although people say data isthe new oil or data is the new
electricity, they're actuallyobfuscating a lot of the issues
(09:17):
or a lot of the important pointsthat data has.
And AI is, of course, usingdata to be the next pervasive
invention.
We already have people do notrealize often already have more
than 2,500 use cases of AI whichare operational in nature.
Everyone is today's hyped upabout LLMs and GPTs, but
(09:41):
actually there many, manyinstances where AI is already
being used and will be used by2050, my belief is at least in
100,000 use cases.
Pankaj Agrawal (09:54):
AI by design, what you can say a lot more
intelligent than the previousthree.
It has a lot more capabilitiesaround reasoning, about
contextual understanding, aboutvision understanding and basis,
which it can enable, as you said, right, multiple of these use
cases.
Right, so do you think, um,because of that and you know,
(10:14):
for example, many of the bigtech leaders whether it's on
that pitch I came out with thestatement, right, you know that
invention of ai is is as, uh, asinstrumental, if not more, than
electricity, than the inventionof electricity.
Right Now, that's a bigstatement.
Arguably, that there are vestedinterests and so on and so
forth, right, but do you kind ofbelieve that the acceleration
(10:39):
or the unlock of use casesthrough AI could be a lot more
pervasive than what the previousones have been?
Or do you think, you know, it'stoo early to kind of really
figure that out, because AI isfundamentally, you know, I would
(11:00):
say, meaningful intelligenceover the past three, right, so
do you think it would affect howit gets kind of seeped into the
mainstream?
Dr.Alok Aggarwal (11:11):
Yes, two aspects.
One is that, unfortunately, aiis again being hyped very, very
strongly, partly by varioustechnology companies, partly by
the media.
This is not the first time thatthis hype has happened, in fact
, when artificial intelligencewas initially created by Alan
(11:33):
Turing by talking about theimitation game.
After that there was aconference in 1956 in Dartmouth
College, and that's when thename was given.
And during 1956 to 1973, ai washyped equally strongly, maybe
more, maybe a bit less, butabout the same in many ways.
(11:54):
In fact, a movie was created,2001, a Space Odyssey, in which
the computer HAL 9000, isconsidered to be an artificial
general intelligence computerwith emotions, abilities to
scheme, arrogance, all the humancharacteristics and much more,
and it can beat people in chessand so on.
(12:16):
So, from that perspective, thatwas the first hype.
It went sour, it went bust in1973.
The second hype started in 1980with expert systems, and that's
a very important thing.
What is AI?
Ai, in effect, is a combinationof machine learning and expert
(12:37):
systems, expert systems beingknowledge-based, subject matter
expertise.
And even for humans, withoutsubject matter expertise, we are
not.
Without subject matterexpertise, we are not able to do
most of the work.
We become eventually knowledgeworkers in some field or the
other or are doing a basic worklike typing and so on.
So, from that perspective,1980s again there was a.
(13:00):
It was short-lived, it wentbust in 1987, 88.
And then this is now again thehype.
Interestingly, all hype cycleshave the same thing.
Businessweek, one of themagazines in US, actually had
the front cover in 1984, Ibelieve.
It says AI it's finally here.
And isn't that almost whateveryone is saying, including Mr
(13:24):
Pichai and others AI is finallyhere.
Having said that, all inventionsthat is the fourth
characteristic I bring out ofindustrial revolutions All
industrial revolutions havethese very important inventions,
but they take time to seep intothe society.
Things did not happen overnight.
(13:46):
For example, railroads.
The first railroad was createdin 1830 between Liverpool and
Manchester and people got soexcited.
Obviously steam engines weregetting pervasive.
People got so excited.
Inventors, investors and thenthe public, media, public they
all put in a lot of money.
By 1890, it was oversaturated,the hype had died down.
(14:10):
Most of the railroads went bust.
Same thing we see during thesecond industrial revolution
with telegraphs.
In the third industrialrevolution with telecoms and
dot-coms I mean the telecommoney.
About $500 billion was spent inbillion, not million.
$500 billion was spent in tolay cables in sea, under water,
(14:34):
that is, under soil, wirelessWi-Fi and many of these
companies went bust.
So it takes time for inventionsto seep into the society and of
course, the time is reducing butit is not going away.
For example, in 1882, the firstelectric generation plant was
(14:55):
created by Thomas Elber Edisonin upstate New York.
That was in 1882, but it wasonly in 1925, that is, about 43
years later, that half ofAmericans got electricity.
So 43 years later that half ofAmericans got electricity.
So 43 years later.
Similarly, even more recently,the first handheld phone was
(15:17):
created in Motorola by Cooper in1971.
But it was only in the 1990sthat people actually started
using handheld phones.
I'm just talking about not evensmartphones like BlackBerry.
Yeah, started using handheldphones.
I'm just talking about not evensmartphones like Blackberry, or
that is something which is veryimportant to keep in mind that
things take time to seep intosociety.
(15:39):
I give eight reasons why that'sthe case.
Everyone falls into that,including me.
When I started EvaluServe, wegave the notion of knowledge
process outsourcing and we saidin 11 years or 10 years, this is
how much it will be and Indiawill gain so much out of it.
The rest of the world will gainso much out of it.
Not that fundamentally we werewrong.
(16:02):
We were wrong about the time.
It took at least two times whatwe had projected, what I had
projected.
And that is the fifthcharacteristic of industrial
revolutions.
Fourth being things take timeto seep into the society.
Fifth being that people get allexcited and create hype cycles
which go boom in the bus.
Of course, many of these hypecycles are good.
(16:24):
I mean, if there was no hypecycle, we would probably have no
railroads.
Most of the railroads that wehave to do are because of the
hype cycle that got created inEngland and in the US, and
that's a good thing.
So I actually am a very strongbeliever.
It may hurt the investors, whomay lose their shirt, but it
(16:45):
helps the society in the longrun.
So definitely, ai will play abig role, probably a bigger role
than electricity, althoughthat's hard to say at this point
, but I think that will happen.
Gradually the pace will beginto pick up, but it will happen.
As I said, we clearly haveabout 2,500 use cases that I
(17:08):
know of and I'm putting up about1,000 of them by the end of
August on our website.
There are many more and I thinkit will go to 100,000.
So, from that perspective,there are only 3,500 motors
electric motors of differentkinds but there may be about
100,000 different use cases ordifferent type of AI systems.
(17:29):
So from that perspective, mrPichai may be right.
Pankaj Agrawal (17:34):
Excellent.
I think that's a good segue todiscuss.
What is the present state of AIright?
You mentioned about 2,500 usecases out there and, with your
experience at SCRI and theresearch for the book, what
sectors and industries are youmost excited about with respect
to AI adoption?
Maybe some examples of the usecases that are getting
(17:55):
accelerated or maybe altogetherreplaced by AI first approach,
if you want to pick yourfavorite out of those thousand.
Dr.Alok Aggarwal (18:02):
So, first of all, I think pretty much all of
the industries and departmentsof organizations, or you can say
divisions of organizations,will be affected by by here.
There is absolutely no doubt inmy mind that, if we excuse me
if we look at the longtrajectory I mean if we assume
(18:22):
that the revolution started infourth industrial revolution
started in 2011, with IBMwinning the Jeopardy contest my
feeling is we have at leastanother perhaps 10 to 15 more
years to go before AI reallyseeps deeply into it.
So, as part of those 1,000 usecases, pretty much I would say,
(18:47):
are broken up into 19 industries, one government with 20, and
then 10 different departments ordifferent divisions of a
typical organization finance andaccounting, procurement,
marketing and sales, and so onand so forth.
So, from a perspective ofcompanies, you can look at AI.
(19:10):
Ultimately, if we take on thehype, what can AI do today or in
the near future?
One thing it is very good at isclassifying, that is, it
differentiates between the facesof dogs and cats.
That's a very simple example invision.
But a more interesting exampleis to differentiate whether a
(19:31):
person has skin cancer or aperson should be given out a
loan.
So these are classification.
Secondly, it is very good atfiguring out patterns.
So otherwise, AI, as AlanTuring said, AI machines are
essentially child machines.
You train them, just like youtrain your children to provide
(19:52):
learning.
Then you test them and if thetest fails it doesn't have
enough accuracy to train more,or you rewrite the AI program.
So from that perspective,already there is a lot of
innovation, a lot of use casesin the Internet of Things, For
example, figuring out whethersomebody has a water leakage or
(20:15):
a gas leakage in his or herhouse.
There is about 4% water aroundthe world gets leaked in water
pipes as the water companiessend the purified water, the
potable water, from their plantsto homes.
So a lot of this is getting now, because the sensors are
(20:38):
relatively inexpensive, that alot of this is happening already
and will continue to happen inthe area of supply chain
management and in generalinternet, happen in the area of
supply chain management and ingeneral internet.
Clearly that would be one ofthe big fields, if I would call
a vertical which will getaffected.
Another one will be banking andfinance insurance.
(20:58):
Now, insurance has a veryinteresting problem, especially
life insurance, because it goeson for 40, 50 years and
therefore they have so much datawhich is in paper-based data
and that paper is not the factsof today very, very poor quality
(21:19):
.
So converting no matter.
Even humans cannot understandit, Forget about AI
understanding it.
So that's a very importantproblem and I think it will be
resolved in the next 10 to 12years.
Another problem which is alongthe same lines is reverse
engineering of COBOL programs.
So COBOL was a language whichwas created in 1960-61 and even
(21:44):
though it's a very simplelanguage, because it's a simple
language, people never wrotedocumentation on it.
And even though it's a verysimple language, because it's a
simple language, people neverwrote documentation on it.
And even today about 7 to 8percent of US economy easily
runs on COBOL.
So all of insurance companiestoday I mean India fortunately
got saved because COBOL by 1980shad become already an older
(22:09):
language and newer languages hadcome in Java and so on C++, C,
Java.
But COBOL for the Western world,for the more developed world,
is a very, very big problembecause people like me I mean I
learned COBOL in IT in 1975, onecourse when, when we did not
(22:30):
have even tapes, we had a deckof cards which we used to feed
into the computers and I meanpeople like me pretty much.
Most of them have retired, Someeven unfortunately passed away.
There is no documentation andwhat happens all the time is
that people modify their COBOLprograms and those have bugs in
(22:51):
them.
Nobody can now figure out whereare the bugs.
So just to give you an idea,most of the US Army Department
of Defense runs on COBOL.
(23:11):
So some of the nuclear basicallyreverse engineering COBOL
programs so that you don't needeffectively, you need very few
if any, COBOL programmers.
The other advantage of reverseengineering of COBOL program
using AI is because it doesn'tknow the syntax of COBOL
programmers.
The other advantage of reverseengineering of COBOL program
using AI is because it doesn'tknow the syntax of COBOL.
(23:32):
It would actually find all theerrors, or most of the errors
that are in the COBOL program,because it's not transliterating
COBOL program into a Python ora Java program but into a flow
chart.
So then we've got many caseswhich are very interesting.
I mean, healthcare is anotherone which will be fundamentally
(23:54):
changed in the next 10 to 14years.
But again the hype far exceedswhat reality is, which is the
sad part.
And right now I mean, on theother hand, as I write in the
book, hype is good, because ifwe put in 50 billion dollars in
(24:15):
in llms and gpts, it is good forthe humanity, the human society
, because something good wouldcome out.
Problems are hard problems,they're not easy.
Yeah, so you need a lot ofpeople in a lot of months.
Pankaj Agrawal (24:31):
Interesting, and I think that's a good segue to
the next set of topics that Ihad in mind.
I mean, given that you'reserving large industries of all
shapes and sizes first with EVL,you serve now with Scry where
are these companies in the AIjourney?
Because what we constantly hearis that a lot of pilot testing
is going on but barely anythinghas moved into production.
Sequoia came out with a recentreport that a lot of CapEx is
(24:55):
being spent in enablinggenerative AI and powering LLMs.
Kind of make up for that, right.
Google, meta, microsoft, allthese guys I think they recently
had their quarterly earningsand all of them are spending
anywhere from 15 to $20 billionevery year, every quarter, not
(25:18):
every year every quarter asCapEx.
A large chunk of that isobviously going into procuring
the GPUs, you know, obviously.
And then NVIDIA is obviouslyNVIDIA stock has been, you know,
touching the skies, but none ofthat has resulted into the kind
of revenue, right?
Or maybe the production usecases, right?
So what do you hear from yourcustomers, right?
(25:39):
I mean, where are they in thejourney, right Among?
the use cases which are kind ofalready proven.
Yeah.
Dr.Alok Aggarwal (25:43):
Yeah.
So again I go back to thecharacteristics of the
Industrial Revolution.
I said the fourthcharacteristic was that it takes
time for even the bestinventions to seep into the
society, the fifth being thatpeople get hyped up about it and
therefore you have a boom-bustcycle.
And we are in a very strongboom-bust cycle on Google Buzz
(26:08):
site, because the problem is,everyone has gotten excited
about one very, very specificaspect of AI, which is GPTs and
LLMs, large language models andgenerative pre-trained transform
.
The first paper in this regardcame from Google Research itself
.
It was a research paper calledAttention is All you Need, and
after that Google created BERT,which was the first transformer.
(26:32):
It wasn't that good because itdidn't have enough parameters,
so to speak, without going intodetails, but since then the race
, or the war, began.
Llms are very interesting forhumans because it can write me a
recommendation letter, it cansummarize very large pieces of
(26:54):
text for me, it can improve myEnglish.
They've been trained.
These particular deep learningmodels, or deep learning
networks, are trained onanywhere from 400 million pages
to a billion pages of text andtables and so on.
So they're very, very good inwriting English.
They're almost perfect in that.
(27:15):
Imagine a kid going throughthat many novels If a novel is
about 400 pages long that's alarge novel, maybe 200 pages
long, talking about 2 millionnovels Then either the kid will
go crazy, but also the kid willprobably become excellent in
writing English, very cogent,very proficient, and that's what
you see in LLMs also.
(27:38):
They're extremely cogent,they're extremely proficient,
they're so good that they caneven fool lawyers and they fool
lawyers into believing thatnon-existing cases exist and
therefore lawyers into believingthat non-existing cases exist.
So, just like I was talkingabout the kid, that the kid will
go crazy and will becomeproficient, probably that's
what's happening.
(27:59):
We don't know what's happeningwith machine learning models,
but probably that's what ishappening with machine learning
models.
They're both crazy and that'swhy they they hallucinate and
they are extremely good atwriting, which is actually in a
very interesting way or a sadway for humans, because they're
writing so well.
(28:19):
We trust them and that's whatwe call machine endearment, but
then they hallucinate.
So we haven't really found outactually good ways of using
these particular deep learningnetworks for large language
models and DPTs in real sense ofthe word.
Yes, it can write a very goodpoem for me.
(28:40):
It can elaborate something forme.
It can write a story, but can Iuse it in any meaningful manner
?
And I think this is whatSequoia is talking about.
I think this is what alsoGoldman Sachs is talking about.
This is what I write in chapter11 of the book.
Not that they will not be usedAgain.
It goes back to the industrialrevolution and its
(29:02):
characteristics.
Ten years down the road, theywill be commonplace.
We would have figured out, butright now we don't even have the
first, second, third-mileproblem, several-mile problem
solved for these GPTs.
For example, they're trained onInternet data, but a company
has its own data, which ispaper-based, pdf-based, excel
(29:23):
spreadsheets and so on.
The first mile is you have toconvert all this data so that a
large language model can adjust.
So those are the ones why Ithink Sequoia.
I think there was also a smallcomment or a small article from
Barclays.
Sequoia and Goldman Sachs hadbeen on the forefront.
(29:45):
In December 2023, when I wrotethe book in chapter 11, I said
look, more heist than reality,because McKinsey says it's about
15 trillion dollars will comeout of it by 2030 or it will
affect 15 trillion dollars.
Now the world's GDP will beonly 150 trillion, so you're
talking about it affecting 10%of the economy of the world.
(30:07):
I don't think that's even goingto be close to where we are.
Having said that, there aremany areas which we don't even
today.
We don't even see where AI is,for example.
So many not so many, but manyairports, including in the US,
have started using facerecognition to speed up the
(30:30):
immigration process.
So in Dubai, if you go, it'syour face that it checks and if
it's 98% correct, it'll just letyou go.
Literally, it reduces your timefrom 10 minutes to 10 seconds.
Singapore has done the same.
Us government, slightly morecautious, does it all for
specific people that it hasalready screened in the past.
(30:54):
Or the global services people Imean travel services people.
So we don't even talk aboutthese.
Very soon, a lot of these thebiometric recognition, whether
it's pupils, face recognitionhands in many countries not in
the US or Europe, but in manyother countries will be used as
(31:15):
a way to check out of a grocerystore.
Not in the US, because the USis still worried about privacy,
and so is particularly Europe.
Similarly, if you look inagriculture, there is a lot
going on in agriculture intrying to understand various
aspects how much nitrogen isthere?
How do I implant seeds in thedirt without literally opening
(31:40):
up the dirt, because the momentyou open up the dirt which has
been for the last 3,000 years ormaybe 5,000 years that you
basically use an animal whichhas a machine, a small machine
in the sense it has a small toolin it which will open the dirt,
and then you throw seeds andthat always gets carbon dioxide
(32:04):
into the air, that always losesmoisture into the air and, above
all, it loses topsoil to theseas.
So a lot is happening inagriculture, and agriculture, I
believe, in 25 years, willfundamentally change.
So there are areas where thingsare moving and moving very
rapidly.
One area which I'm particularlyinterested in, and particularly
(32:26):
because of Scribe, isintelligent document processing.
There is so much paper and PDFdocuments, but these are either
unstructured or semi-structuredin nature.
You need to convert it into anelectronic form to be able to
automatically use AI, to playwith it, to give decisions,
(32:47):
support and so on.
But unless you can get thatcompletely resolved with very
little human intervention, youwill not be able to solve this
problem.
And that I mean out of the1,000 use cases, there are about
120 use cases on intelligentdocument processing alone what
all can do or should be able todo.
(33:09):
Some of them are already in someform or the other, are already
being implemented, the simplestones.
Pankaj Agrawal (33:21):
Got it.
So our audience has a largechunk of our audience are
founders who are alreadybuilding or kind of thinking of
building a business right.
So what areas within AI right,Do you think a new entrant can
contribute to and, in theprocess, build a large business?
Right, I mean, what should theykeep in mind while serving
(33:44):
incumbents?
Dr.Alok Aggarwal (33:45):
right, I would suggest they look into three
areas very particularly andparticularly because generally
the entrepreneurs aretechnologists, they miss out on
these areas quite a bit.
One is to make sure that youstart small with a small problem
.
But the area effectively.
(34:05):
You know that there are manyadjacencies, so I'll start with
the use case.
But there are many use caseswhich are very close by and I
can solve them in the long run.
Now that doesn't mean I willsolve them on day one.
It may take a seven yearjourney, five year journey.
And, by the way, I mean, unlessyou're an entrepreneur, unless
you're planning on selling thecompany for 100 million or 50
(34:27):
million, most companies take 20to 25 years really to build.
So you should keep in mind asto which particular direction
you want to go, whether you'rein it for about half of your
working life or you're in it forone or one third of your
working life or only for fiveyears of it.
Both are perfectly fine.
I'm not advocating one versusanother, but that's something
(34:50):
which is very important.
Having said that, that theyshould keep adjacencies in mind,
that I'm going to look at thisuse cases, but use case to make
a minimum viable product, butthen I'll go and expand it,
because those are the use casesthat can expand to.
Second thing that they shouldbe very careful about is AI is
(35:12):
not machine learning and machinelearning is not AI.
There is no equivalence betweenAI is a superset of machine
learning.
As I said, the second AI boomand bust happened because of
expert systems.
No matter how much we put in,even in LLMs and GPTs, how much
we train them 400 million pages,500 billion words they still
(35:38):
require context.
They still require and humansdo.
By the way also, I mean I'm notinto accounting Somebody comes
to me and says can you figureout what is the total operating
income of a bank?
Here is the report and supposethat number is not given.
I may not know what are theformulas I will use because I
(35:59):
don't have the context andobviously AI suffers a lot from
that.
I mean GPTs and LLMs.
A child will suffer a lot, evena human will suffer a lot.
So I think it is very importantfor them to understand the
distinction between machinelearning and say look, I need to
add actually knowledge matterinto it, subject matter,
(36:24):
expertise.
So that is a very, veryfundamental thing.
My own view is 90 percent ofcompanies startups will fail
because they do not include thesubject matter expertise into
the entire story.
(36:44):
The third is not to get caughtup with the hype.
A lot of people are just takingeither open-source LLM models
or just open-source models,putting them together and
creating a minimum viableproduct out of it.
Great, you can do something,you can create a chatbot out of
it, and so on and so forth.
If you're a services company,it may all work out, also
(37:07):
because you'll move on to thenext project.
But if you're in a productcompany, then you don't have any
intellectual propertyassociated.
So how do you sell?
I'll give you an example ofintelligent document processing
which basically takes the data,converts I mean which is scanned
data or PDF data converts itinto electronic data.
(37:28):
But unlike humans, which havetwo eyes, the machine converts
it into literally one dimension.
It loses the context of tables,loses the context of graphs,
charts, etc.
They have the second problem,which is to recreate tables with
pretty much 100 percentaccuracy, graphs and charts.
Then suppose I come and tellyou that look, it's 90 percent
(37:52):
correct.
You're my client.
Your immediate reaction is okay, that's pretty good, it's 90%
correct.
You're my client, yourimmediate reaction is okay,
that's pretty good, it's 90%correct.
But this is a 300-page documentyou just converted.
Will I have to review it fromleft to right for all the 300
pages, because I don't knowwhere the 10% is wrong?
So the question is if I canreconcile, for example, various
(38:16):
suppose there is a table in itwhich is income state, I can
reconcile and I can show youthat look or not me.
But the software says look inthis statement, a plus B plus C
equals D.
Then you don't have to reviewit.
So I can very, very clearlypoint out which are the issues,
(38:41):
where are the issues, and so onand so forth.
This is just one example thatyou add, because everyone I mean
there are 57 or 58 companiesthat we know of.
There may be more companies inintelligent document processing,
but none of them.
Everyone will say oh, you willneed a human in the loop.
Not clear that you will need ahuman in the loop.
In humans, yes, we have a makerand a checker, but it's not
(39:05):
that we use checker all the timebecause checker is expensive.
It's our two eyes which solvethe problem.
With my right eye closed, Ilook at you and I say it's
punctured.
With my left eye closed, I lookat you and say it's punctured.
The error of my being wrong issquared, because if my right eye
was 90% correct, error was 10%.
(39:26):
My left eye was 90% correct,error was 10%, and if both eyes
were working independent of eachother, the error becomes
squared.
That is only 1% wrong 10%.
Wrong there 10%.
So what I'm trying to say isthere are interesting areas that
people can go into.
I'm not saying everyone will,but I'm saying that in product
(39:50):
business you have to haveintellectual property that you
can defend and unfortunately Ido not see that right now.
This is one of the things wherepeople are rushing.
It's more than a gold rush I'llcall it a platinum or a diamond
rush that people are notrushing to figure out what the
issues are and how to solve them.
(40:11):
Rajesh.
Pankaj Agrawal (40:12):
KASTURIRANGANANI , could data be that
intellectual property?
So one common theme, or ratherthe playbook and I wrote about
it on LinkedIn recently that aplaybook is, as you said
identify a painful and a usecase, build a superior product
to solve that.
That becomes your sort of footin the door product generate
demand, bring along customers.
(40:34):
Doing that In the process,hopefully you are generating
some high-quality andproprietary data which you can
use to kind of keep on makingthe product better.
Do you think the data itselfcould be that IP?
Dr.Alok Aggarwal (40:47):
I don't think data will be the IP.
I think use of the right datais part of the design knowledge
and that's where subject matterexpertise comes in.
So what data will I use?
Because there is a lot of dataaround, but a lot of data is
also noisy and you can't use it.
Pankaj Agrawal (41:09):
Quality of data.
Dr.Alok Aggarwal (41:14):
So that goes back to subject matter expertise
.
What data do I use?
How do I get to creating aproduct which is superior?
And when I say superior, it hasintellectual property which is
hard for people to overcome.
The sad part about AI is or, ingeneral, software is yeah, you
(41:37):
can write patents, but they'reworth not even the paper they're
written on.
You give me a patent insoftware, in AI, because it's
math.
Eventually I can go around itfairly easy.
It's not like building acompany, and I know you folks
are in the venture capitalbusiness business and I know
many VCs who get all excitedabout patents.
(41:58):
Even though we could probablyin our company write about 30 to
40 patents, we have not writtena single one because we know
the moment we write a patent, wehave essentially given our
intellectual property away.
Somebody will just go around it.
Yeah, it's like reconciliationof intelligent document
processing.
(42:18):
We have not written a singlepatent on how do we reconcile?
How does the system learnformulas which are there in a
PDF document automatically, orhow does the formula get learned
in a football program?
Pankaj Agrawal (42:38):
I think rightly so.
Another theme which I agreewith that you pointed out that
subject matter expertise will berelevant.
So if you see, now Lama came upwith 405 billion parameters and
it is kind of giving a chase toGPT-4.0.
And a lot of experts are kindof saying that GPT 4.0 is like a
Ferrari for grocery shopping.
(42:58):
Right, it's just too over thetop for most of the things.
Right Versus open source modelthat you can access, you can
utilize, you can train on yourspecific use case, make it like
you know, work really reallywell and accurately for that
domain.
So that will be the future andthat's where subject matter
expertise becomes more and morekind of relevant.
(43:19):
I personally feel that modelsthemselves are probably the
fastest depreciating asset inhistory, probably.
Dr.Alok Aggarwal (43:29):
In some sense they're not even depreciating
because they become open sourcevery quickly.
Like Lama is open source, Imean Mixtral is open source.
I can choose a bunch of opensource models and then I can put
them together.
Now, what is intellectualproperty is?
Lama has, let's say, in aparticular metric, 90% accuracy,
mixtral has 88%, and all ofthese are open source.
(43:52):
Can I now put them together ina meaningful way so that my
accuracy becomes 92 percent inthe whole process?
That becomes my intellectualproperty.
Now, if I put subject matterexpertise on top of it, then I
can get to 95 96 percent andthen if I can reconcile
everything and show you thatthere is no hallucination,
(44:13):
because I'm showing you wherethe where the particular thing
got its answer from, then youare a happy person as a client,
right?
I mean then that you say look,you have really saved me time,
money, cost etc.
And above all, human labor.
So I think these things willtake time.
They always.
And that's why chapter one ofthe book is so important,
(44:35):
because it sets up a frameworkthat, look, let's not get
carried away.
I mean it's perfectly fine ifwe get carried away.
Also, because I do say thathype, boom bus cycles are
actually good for society.
I mean, we saw that withdriverless cars recently.
About $100 billion was spent inthere.
Yeah, some amount of it reallywent into research and
(44:58):
eventually we will solve thedriverless car problem.
It may be eventually, maybe 10to 15 years from now.
So I'm not even saying thatcycles are bad cycles For the
human society.
They are actually probably goodthings to have.
But investors and inventors ofcourse, they lose their shirt,
(45:21):
so they feel bad.
Yeah, that's it.
Pankaj Agrawal (45:26):
This has been super interesting.
I have one last question thatI've been kind of trying to
unpack for the last couple ofmonths as I go deeper into AI.
I think AI, with its power to,you know, either as a co-pilot
or you know, kind of make theoverall system more efficient,
has the ability to kind of addmeaningful value in services
(45:50):
businesses right, and there isan amazing concept that, rather
than software as a service,there could be certain areas
where service as a software hasthe right characteristics of a
venture scale business producingventure scale business in terms
of capital, efficient growth,velocity of growth, and so on
and so forth.
Given that you have experiencein both now with E-valueServe
(46:12):
with respect to KPO, and nowwith Scribe, you are
buildingalueServe, you know withrespect to KPO, and now with
Scry you are kind of buildingproducts for these customers,
large customers what is yourview?
I mean, if you were to, let'ssay, think of starting
E-ValueServe in today's world,would you do anything
differently?
You mentioned that it took youtwice the time you projected to
get to the scale right, or wheredo you think?
How should an investor thinkabout the services space
(46:34):
generally and how can it bepowered by that?
Dr.Alok Aggarwal (46:37):
Actually, after writing the book, I
decided this year to convert twothings To convert the book into
36 lectures of 45 minutes each,because not everyone reads
books these days, very fewpeople read books.
And by no means the book issmall.
It's actually 270 pages offairly dense material.
(46:57):
It's not intense because itdoesn't have math or computer
science, but it is dense.
You have to think about it,read it once more before it
begins to seep into your intothe brain, into your thought
process.
So, uh, so that's one.
And the second is to write atleast eight articles on India
(47:17):
and AI.
Because everyone talks aboutChina and US.
People are forgetting thatIndia has the largest number of
software professions right nowin the world 30 million
developers out of 30 million.
And that is 5.5 million softwaredevelopers actually working on
a daily basis in all thesecompanies.
(47:38):
And we're not talking aboutHTML and CSS developers, we're
talking about softwareengineering developers.
I mean, the world has about 25million and India has about 5.4.
So people forget about that andthat is going to be a very,
very important aspect aboutIndia's growth in AI which
(47:59):
people don't realize.
Having said that, so my firstchapter was first article was
the unsung heroes of AI how dataannotation will grow enormously
, meteorically in India.
The second one was, of course,loneliness in India will rise
because of AI.
(48:19):
The third one is that theservices industry in AI will be
transformed radically by AI inthe next 10 years.
Because you talked aboutEvaluServe.
When I think about EvaluServeall the time, because I'm still
invested in it fairly heavily,even though I'm on the board,
and if I were to redo it, Iwould just completely redo it
(48:42):
with software first and humanslater.
And about 50% of the work thatEvaluServe does and, to me,
about 70% of the work that TCSand others do, especially in the
BPO domain, like GenPan, likeeXcel service, 70% of the work
(49:03):
is within the next three to fiveyears can be reduced by a
factor of two which is a lotright, Because suddenly clients
begin to realize that you'vereduced it by a factor of two,
they'll say why are you chargingme X when you should be
charging me X over two?
And then you say, okay, there isa software cost.
(49:25):
They'll still pay you maybe 0.6, 0.7 times X, so everyone gains
in the whole.
Pankaj Agrawal (49:32):
And this it can also expand the overall demand.
Right, the Jevons paradox willkick in.
You reduce the cost oftechnology per unit consumption,
the demand will grow up.
Dr.Alok Aggarwal (49:43):
Absolutely, because second one is it will go
pay by the drink, not by theFTE, so to speak.
So you will go into, as yourightly said, small,
medium-sized businesses whocannot afford any Excel or any
value serve today will be ableto afford the next generation of
companies.
So I think, this is going to bea very, very fundamental change.
(50:05):
The only good news and thispaper should come out in
September the only good news Isee is that it is not again not
going to happen in the nextthree years.
It may take seven to 10 yearsfor this change to occur, but
this change is imminent.
This change is.
I mean I hope TCS, infosys etcare waking up to this change.
(50:26):
It's not that these people willgo away, but I mean you will
still have 50 percent.
You'll have people who will goupstream and so on.
India may actually gain a lot.
It is an opportunity for Indiato really play very strongly in
the industrial revolution,because the foundation is set,
because India has so manysoftware engineers, because
(50:48):
India can produce enormousnumber of data annotators who
can check for ground truth, whocan basically do data annotation
, who can tell you that look,this part is this portion,
because in supervised learningwe need data annotation.
Pankaj Agrawal (51:08):
Fantastic.
Thanks a lot, Really reallyappreciated doing this
conversation.
Of course, we'll recommend theaudience to check out your book.
We'll link it in the show notesand thanks for taking the time
to chat with us.
Dr.Alok Aggarwal (51:22):
Thank you, Pankaj.
Thank you so much.
Thanks, Jerome, for making itwork all of it.
Thanks, I'll write to youseparately.
Pankaj Agrawal (51:30):
Sir, I'll take your details from Jerome.
I would love for you to talk toEmpirical and at some point in
time, of course we would wanthim to spend some time in the US
, so of course we would love foryou guys to.
I mean, he's from your almamater, so that's alma mater,
building something in AI.
So I think two strong reasonsfor you to take the time out and
(51:51):
have a chat, and I think itwill be helpful for him as well.
Yeah, sure, absolutely All right, have a great day, sir.
Thank you.
Thank you Bye, bye.
Prime Venture Partners (52:08):
Dear listeners, thank you for
listening to this episode of thepodcast.
Subscribe now on your favoritepodcast app for free and you'll
be the first one to know whennew episodes are available.
(52:29):
Thank you, we would be reallygrateful if you leave us a
review on Apple Podcast.
To read the full transcript,find the link in the show notes.
Popular Podcasts
Crime Junkie
Does hearing about a true crime case always leave you scouring the internet for the truth behind the story? Dive into your next mystery with Crime Junkie. Every Monday, join your host Ashley Flowers as she unravels all the details of infamous and underreported true crime cases with her best friend Brit Prawat. From cold cases to missing persons and heroes in our community who seek justice, Crime Junkie is your destination for theories and stories you won’t hear anywhere else. Whether you're a seasoned true crime enthusiast or new to the genre, you'll find yourself on the edge of your seat awaiting a new episode every Monday. If you can never get enough true crime... Congratulations, you’ve found your people. Follow to join a community of Crime Junkies! Crime Junkie is presented by audiochuck Media Company.
24/7 News: The Latest
The latest news in 4 minutes updated every hour, every day.
Stuff You Should Know
If you've ever wanted to know about champagne, satanism, the Stonewall Uprising, chaos theory, LSD, El Nino, true crime and Rosa Parks, then look no further. Josh and Chuck have you covered.