YT YouTube2Text

YouTube2Text → Transcripts → GAEA Talks - Human In The Lead, Not Human In The Loop with Cabinet Office's Dr Ravinder Singh

GAEA Talks - Human In The Lead, Not Human In The Loop with Cabinet Office's Dr Ravinder Singh — Transcript

by GAEA Talks · 10,431 words · 1,941 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00For doing this work, I need to use this
0:03model, this algorithm, this data.
0:06And that becomes my work.
0:09So, unless I do that,
0:11I will not be able to get efficient
0:13results. I may get results,
0:16but they will not be as efficient as I
0:19want.
0:20That's the real challenge which people
0:22are forgetting that machine learning
0:24will require time, training,
0:26explainability, and I'm using the data
0:29in the right way, in the right form.
0:32Second is is it sustainable? I'm using
0:35the technology which is sustainable
0:36[music]
0:36technology.
0:38Highest encryption which is available to
0:40you is 2048 [music]
0:42bits.
0:44Well, that's the highest one.
0:46And current computers been It will take
0:49years to solve it, but with quantum
0:51computers [music] may take
0:54few hours only.
0:56>> [music]
1:01[music]
1:11>> This episode of Guy Talks, it's a
1:13pleasure to have you, Ravinder Singh.
1:17Tell me about yourself.
1:19>> Okay, Graham. It's um I'm Ravinder
1:22Singh. Um currently working as head of
1:24digital and systems team within Cabinet
1:27Office, Government Commercial Function.
1:29So, my role is working
1:31uh
1:32across different teams, providing them
1:34technology tools, emerging technologies,
1:38AI, machine learning, blockchain, IoT,
1:40quantum computing,
1:42as well as we have a government
1:43commercial college which we run.
1:46And we have 106,000 users of that
1:48platform,
1:49which is
1:51largest uh
1:52after the Open University. So, we have
1:54one one largest after the Open
1:56University,
1:57which is good to see and and it comes
1:59with its own challenges.
2:01Uh but before this role, I was with the
2:04Government Digital Service, again in
2:05Cabinet Office.
2:07And we I was consulting technical
2:09architect working across various
2:10departments advising them on
2:12various technologies, tools,
2:14architecture, design, optimizations.
2:17Um so I was part of their technical
2:19design authorities.
2:21And before coming to
2:23the government services civil services
2:26uh 8 years back I was with private
2:29sector working with large-scale
2:31financial sector like JP Morgan, HSBC,
2:34Credit Suisse
2:36uh Accenture, 3i Infotech.
2:39Uh so that I have done a lot on that
2:41sector. Shell Oil is another company
2:43which I worked with implementing
2:45large-scale complex transformation and
2:48digital programs.
2:50Uh I did my PhD from King's College
2:53London.
2:54And after that my PhD I joined the civil
2:57services.
2:58And before um
3:00like I came to UK in 2004
3:02on highly skilled migrant program
3:05uh which was on first skilled
3:06professionals for visa on point based
3:08system. I came on that.
3:10And I was in India. Again, I was working
3:12as a civil servant there also
3:15working for technology development of
3:17Indian languages um
3:19in which Punjabi is one of the official
3:21languages. India's got 17 official
3:23languages. We created complete word
3:25processor which is like the MS Word but
3:28for Indian languages.
3:30And I had two copyrights of software
3:31there and one national award in in that
3:34for the algorithm which we wrote for the
3:37Indian languages spell checking. Well,
3:40that's slightly trickier than the
3:43English language.
3:44>> [snorts]
3:44>> So use the parameters of this but
3:46created that and
3:48so that's me in brief I You say.
3:50>> Okay.
3:52So you must have an interesting view of
3:55technology given
3:58this new wave of artificial intelligence
4:00is
4:01predominantly dominated in the US.
4:05Um
4:06to have the the UK view in it, but also
4:09the international view. And you you you
4:12talk about the the thing that you you
4:14did in India
4:16being that language and the nuances of
4:19it in a huge population. And then we
4:21have a massive audience in in India. And
4:24a lot of the talent, which is in in the
4:27United States, in the UK, and worldwide
4:30comes from India. So,
4:32as as a a hub of expertise, but now
4:36being in here and being obviously at the
4:37the cabinet office,
4:39you must have a an interesting
4:41perspective
4:43which is fundamentally different from
4:46the the Silicon Valley view.
4:49How would you see the state of AI today?
4:53>> Uh if I have to say the state of AI
4:56today,
4:58um of course AI has been in
5:01for years, in '70s, '80s, in different
5:04form and shape. We had Prolog, Lisp,
5:06then neural
5:07neural networks, fuzzy logic. All All
5:10have been used, but
5:12recent shift has happened because of the
5:16GPU power, which has become easily
5:18available at less cost. And uh various
5:22companies
5:23developing chips and servers, which can
5:26now cater to their needs. And easy
5:28access to those technologies. And that's
5:31the real shift which has happened. And
5:34which is good.
5:36Um of course the Silicon Valley and UK
5:40or India, if you think of Yeah,
5:42theoretically, I would say if I may say,
5:46uh Indian people very good, but
5:48practically, if these things are to be
5:50implemented, then US, UK, or Europe,
5:53they take the lead in those, and US
5:57probably are more taking risks
6:00and funding new opportunities and
6:02failing, failing fast, and learning from
6:05it, and then move forward.
6:06>> [snorts]
6:07>> And if I say, um, think of
6:10infrastructure on software as a service
6:12or infrastructure as a service, if I if
6:15I can say,
6:16then there has been forecast to reach
6:19around to 80 billion in 2026, which is
6:23around 35.6%
6:26jump from the previous years, which is
6:29huge.
6:30And most of that has been in US, again.
6:35Uh, so
6:36the
6:38financials, if you think of the
6:40manpower, if you think of
6:43they have that capacity, and going with
6:46AI machine learning,
6:49and to generative AI or LLM models,
6:52whether Anthropic or Claude or
6:55Mythos, whichever you want to call of
6:56Gemini,
6:59they are giving people now an
7:01opportunity to explore things which are
7:06they mean they fit it would have taken
7:08them a number of
7:10um,
7:10months or number of man hours to do
7:13that, and that activity can be done in
7:15lesser hours, and making it more
7:18efficient, whether just the back office
7:21operations, that's one part, but the
7:23front end of developing things also has
7:26changed, and I can quickly develop any
7:29application,
7:30any hardware also I can think of
7:33developing, and there are tools which
7:35can assist me in developing. So, that's
7:37the shift which has come,
7:40and um,
7:41people with
7:43little or less technical knowledge can
7:47also build things now with the tools
7:49which they they they can get get a hands
7:52on. And that's a real positive thing,
7:55but that's also that's a real worry
7:56thing also.
7:58>> AI is reshaping how organizations
8:00operate, innovate, and compete. The
8:02challenge now is making it deliver real
8:05business outcomes.
8:06Forrester's Technology and Innovation
8:09Forums help technology leaders do
8:11exactly that. Aligning AI with strategy,
8:15strengthening data foundations, and
8:17modernizing operations.
8:20Join CIOs and technology executives this
8:22fall in Austin, London, or New York to
8:26explore what's next and how to lead
8:28through it. Visit forrester.com/events/
8:36Our audience can use promo code gaiatech
8:39to get 10% off. That's g a i a t e c h.
8:45Gaiatech.
8:46>> Mhm. How about the risk then? Because I
8:48mean there's there's the democratized
8:50side, which is fantastic, and
8:53now more than ever anybody
8:55could attempt any anything and and do
8:59things which otherwise wouldn't have
9:00been possible.
9:01But with that comes risk
9:05from
9:08what considerations to put in.
9:10>> Mhm.
9:11>> And often when you do something, it's
9:13the cause and effect. You do something
9:15and think, "Oh, that's that's good."
9:17without really considering the negative
9:18consequences. And generative AI, it's
9:22it's amazing you can create things.
9:24>> Yeah.
9:25>> But should you create certain things? Uh
9:29Could it be exploited? And like like
9:32with everything, it's just having a
9:33responsible view. Yeah. know, anyone can
9:36drive a car. Yeah. Obviously, with a
9:38license.
9:38>> Yes.
9:39>> But
9:40drive it safely, don't drive it
9:41recklessly. Yes. So, it's more so a lot
9:44of this is a human problem.
9:46>> Yeah.
9:46>> And the humans which are making the
9:49decisions from the top, and I'm talking
9:51commercially as well, where there
9:53there's a particular motive through
9:56to a a smaller level where
9:59maybe you're distributing
10:02thinking
10:04to an MLM
10:06>> Yeah.
10:06>> and losing the value in between. Maybe
10:09you can't trust it.
10:11I don't really see the world in
10:13such a black and white sense as things
10:15are either good or bad. I look at things
10:17as it is what it is.
10:19And the human interpretation is
10:22effectively the lens upon which you
10:24consider what is good or bad.
10:27How how are you thinking about it?
10:29>> Well,
10:30that's a very good question. And I think
10:33what my view is, yes, of course, it
10:35comes with its own risks.
10:37First thing is what we need to think of
10:41how can you make
10:43people believe in the technology?
10:46>> [snorts]
10:47>> Is it a black box for them?
10:49If it
10:50how can people believe that this black
10:52box is
10:54uh not biased, it is giving the correct
10:56data. So,
10:58and the data is good data and it's it
11:01has got I always say
11:03data can be
11:05I can say, okay, I can go to one area of
11:07thing and say, I have collected the data
11:10of whole country.
11:11>> Mhm.
11:11>> But has I gone to the whole length and
11:14breadth of the country to collect my
11:15data about people and then I predict.
11:18So, that's danger is there based on the
11:20data. So, first step is what is the
11:23right set of data? Do you have that
11:25right set of data with you to make the
11:28artificial intelligence work? Because
11:31intelligence can come only based the
11:33information it has got.
11:34>> And that is something so often
11:36overlooked. And and I've noticed that
11:38always conversations with um
11:41numerous people in Silicon Valley, but
11:43so also here and at several
11:46universities.
11:47And especially within at an enterprise
11:50level.
11:50>> Yeah.
11:51>> The assumption that if you say oh AI
11:53will sort it.
11:55Well, that's a capability.
11:58Where does the data come from? Can you
12:00trust the data?
12:01>> Yes.
12:01>> And it's like making food. If you have
12:04really bad ingredients, the probability
12:07is
12:07>> Yeah.
12:08>> you're going to have a really bad
12:09outcome irrespective of what you're
12:10trying to cook.
12:11>> Yes. So, it's basically that the rightly
12:14said
12:15if the data is bad, the output has it
12:18will be bad. So, first thing
12:19people are
12:21thinking
12:23senior leaders, it has to be driven from
12:25them. They need to understand oh yes,
12:27you can it can do it can write a
12:29business case for you. It can write a
12:31good code for you. It can write um
12:34policy matter for you or
12:36some good dashboard for your
12:38um decision-making, but it can only do
12:42it if you have the right data at the
12:43back.
12:44So, they have to spend time in that
12:47building the background and building
12:49that data. Why Chat GPT could make a
12:52such an impact when they launched in
12:542022 was basically they had a 5 years of
12:58work at behind and then they showed up.
13:01Which nobody realized that oh they had
13:05done 5 years of things at the back and
13:07then that that's why the impact is so
13:09high. So, if you have to have that
13:10impact you need to do that
13:13background work number one. Second is
13:16you need to train the model.
13:19The training of the model is very very
13:20important that how I am going to make
13:23people believe in my model that yes,
13:25it's unbiased.
13:27And how the training has happened. And
13:29initially, your model may hallucinate.
13:32There has to be human in the loop. Or I
13:35always say, in rather than saying human
13:37in the loop, I always say human in the
13:38lead. Good.
13:40>> Mhm.
13:40>> For first few iterations, it's human in
13:44the lead who is leading that
13:46uh intelligent system to work.
13:49Machine can learn, but it has to be
13:51guided. It has to be trained.
13:54So, like a child or a small kid is
13:56there, it goes on learning the language,
13:59intricacies of that language. Initially,
14:01they are learning the alphabets and then
14:03joining the alphabets, joining for the
14:06word, and then learning the
14:08uh pronunciations, learning the grammar
14:11of it. So, it's the same here.
14:13And we forget that that we are not
14:15spending time into that part. And we
14:18think, yes, it can be a very good output
14:21coming out. So, that's the real thing
14:23where we need to think. And if you have
14:25to take the risk out of it, we have to
14:27think of
14:29how people think about it that yes,
14:33can people believe it? So, is it
14:35explainable?
14:36Is it responsible AI I am doing?
14:39Is it ethical?
14:40>> All of these comments, I've had several
14:42conversations this week, in fact, in in
14:45meetings.
14:47And there was
14:48>> [sighs]
14:48>> there was a few things that came up and
14:50the expression human in the loop
14:53>> Yeah.
14:53>> and
14:54having had a a discussion with with
14:57people that are not AI experts,
14:59>> Yes.
14:59>> they hear the phrases
15:01and and I asked one person, do you
15:05can you explain to me what human in the
15:08loop means? I mean, I have I I know
15:10myself, but I wanted to hear
15:11interpretations.
15:15And obviously, a human being there
15:17relating to making decisions.
15:20And
15:21you think there's human
15:23everything and a human decides there's
15:26rules based.
15:28So there may be automation and it could
15:31be AI automation but the rules are
15:34written by the humans that's simply
15:35enacting what a human decided and then
15:38there's autonomous AI decision making
15:41where you've no idea what's going on and
15:45it just made me think
15:47before
15:49this type of AI automated decision
15:52making
15:53what did we call human in the loop?
15:57Well it was ever that's what there was.
16:00Yeah that's what there was yeah.
16:02And in many ways that's the opposite
16:04to
16:06the term in a way. So now
16:09human in the loop is simply
16:12not necessarily a human being literally
16:17there deciding as such.
16:20But
16:22it's how the world was. So
16:25if we can reverse back and and distill
16:27and well
16:29what's the opposite of human in the
16:31loop? Automated decision making. And if
16:34it is automated decision making then
16:37what's on the line? What's the risk of
16:40not understanding what decisions being
16:42made. It's effectively losing the
16:45transparency
16:46of how you came to an outcome and can
16:49that affect other people and you don't
16:52know how where why or when. So in in
16:55effect it's it's
16:57all of those those things that you've
16:59just said is it ethical
17:02to not know
17:03why a decision was made which could be
17:07it was made what would appear a racially
17:10motivated decision
17:12and you think well yeah nobody wants
17:14that. Why was it doing that cuz it's
17:17only been trained on white people.
17:19>> Yes.
17:19>> And not others. So, it's it's almost
17:21like unpicking one bit, then another
17:24bit, then another bit. And when you pull
17:25the thread enough,
17:26>> Yeah.
17:27>> then it ends up as a cohesive
17:29conversation,
17:30>> Yeah.
17:30>> where you then value
17:32why should we understand what we're
17:33training? Why should we understand where
17:36the data comes from? How do we construct
17:38it? What are we comfortable
17:41with distributing out to AI? Is it
17:44rules-based? Is it full automation? Is
17:47it full
17:49AI decision-making? Where's our level of
17:52comfort? What's practical? What's not
17:54practical?
17:55>> Yeah.
17:55>> But understanding all of that and how
17:58they're interconnected,
17:59>> Yeah.
18:00>> I think is the education gap.
18:02>> Yeah.
18:02>> It's not hard to get.
18:04>> It's not hard.
18:05>> But is that being communicated and
18:07explained? Cuz but you know, when we're
18:08talking about it now,
18:09>> Yeah.
18:10>> it makes perfect sense. Yeah. But it's
18:12almost like um
18:13a lot of people are using buzzwords
18:15>> Yes.
18:16>> to say, "But you wouldn't understand."
18:17>> Yes.
18:18>> So, don't ask all the basic obvious
18:20questions. Cuz when you do,
18:21>> Yeah.
18:21>> we're empowered. We have agency.
18:23>> Yeah.
18:24It's basically a as you said, education.
18:26That's what I said, training.
18:28>> Yeah.
18:28>> So, it's the training of the model.
18:29Human in the lead, I always say I
18:32I avoid saying human in the loop.
18:34>> [laughter]
18:34>> Normally, it's basically because human
18:37in the lead was already there. You have
18:39to teach the machine. Okay, you do this,
18:41this, this, this, this steps.
18:44Right? And then you teach the model
18:46again and again that this is how it is
18:48done and this is the outcome it should
18:50get. And then you once you are
18:53uh satisfied enough, come Yeah, now this
18:56is now going to work around.
18:58Like training your electric vehicle. And
19:00there are so many cameras giving it
19:02input and then somebody telling
19:05it has to maneuver it.
19:07But that will fail if
19:10your traffic is
19:12chaotic traffic in some countries,
19:15right? So that will that will be a
19:17challenge.
19:18So if
19:19the traffic is disciplined, then
19:22electric vehicle will be good to drive.
19:24But if traffic is chaotic, then it's not
19:27going to work.
19:28So that's where it has to be
19:30how you are getting your data, where you
19:33are getting your data from.
19:35And just to give you an example, if I am
19:37to let say give analysis of
19:40number of votes or how the people have
19:42voted and where have they have voted.
19:45Now if I say that okay, in general
19:47elections, this is going to happen.
19:49Party ABC has to win because I have done
19:53this analysis. But do I give my data
19:56behind it? That I have gone to each and
19:58every constituency and got to the got my
20:01data or I just sit sat in London and got
20:05into few areas of London and then I am
20:07predicting for whole of UK.
20:09But I'm coming back to again to my back
20:12point, data. Then training the model on
20:15their data that how you analyze it, how
20:18can I trust the result it is giving me?
20:20Validation. Validation, yes. And can I
20:23trust the result it is delivering? And
20:25that's where human in the lead has to
20:27come. Or once you have trust enough of
20:30the system that is it can learn now
20:32itself.
20:33Because any AI system will work.
20:37Intelligence part will come later. First
20:40to the machine learning part has to
20:41happen.
20:43So machine learning
20:45can I trust that machine is learning in
20:47a way which I we want it to learn? And
20:50then I can trust because any system it
20:53can go on building the data, go on
20:55building learning from its decision,
20:57learning what it has given, and then
20:59saying, "Oh, yes, it is working now."
21:02Then I can trust that system. That now
21:05it has done thousands of decisions which
21:07are right and now I can trust that
21:09machine is learning correctly. It can
21:11give the next decisions right. So that
21:13machine learning part we have to spend
21:16time there. Intelligence can come at any
21:18point of time. That that is what is to
21:21be projected now outside.
21:23And that's the real challenge which
21:25people are forgetting that machine
21:27learning will require time, training,
21:30explainability, and
21:32I am I building the systems in ethical
21:35way? When I say ethical way is I'm using
21:38the data in the right way, in the right
21:40form. Second is is it sustainable? I'm
21:44using the technology which is
21:45sustainable technologies. Or I am doing
21:48not doing the right thing. I'm training
21:50the machine to be biased.
21:52So that's in my hand.
21:55Right? Whosoever it is. So they can
21:57train the machine to
21:59do the bias or it can be unbiased. So
22:01how you trust that black box? How you
22:04build that confidence in people to whom
22:07you are delivering that yes, you can
22:09rely on the system even if there is no
22:12human because we have tested it is
22:15unbiased. For that your explainability
22:18and responsible AI has to come up. That
22:20yes, people will trust because they know
22:23it's a transparency that they are using
22:25the data in this way.
22:26>> And just for a bit of context, mean
22:28context as a word itself as well I think
22:31is important. But I hear the the term AI
22:34as a broad stroke.
22:36>> Yes.
22:37>> [snorts]
22:37>> What we're referring to here with black
22:40box is more transformer based
22:42technology. Cuz there are other forms of
22:44AI. And
22:46the research that we've been doing more
22:48with the LGM large geo-temporal model
22:51that's
22:52that is by its function flowing data
22:55through
22:56making big systems more kind of
22:58I suppose graph rag oriented where it's
23:02context on scale, where you can trust
23:04it. There's different types of vision
23:07AI. There's numerous types of AI, but
23:10the the one in the popular public domain
23:12because of
23:14Anthropic and Open AI, tends to be the
23:17large language model transformer based
23:19ones.
23:20So, given they're the most prevalent,
23:22that's where the greatest risk is to to
23:26implementing without consideration
23:28because of that nuance
23:30functionality or or how you would
23:32describe how a transformer works. But,
23:36context in itself, I think is an
23:38important factor.
23:40Because if we train something here, and
23:42this is where it scattered um I think
23:44spans into
23:46uh international scaling.
23:49If we train something here,
23:52is it going to work
23:54in Taiwan or
23:56um Papua New Guinea?
23:59Well, probably, and not intentionally,
24:02probably it won't. I understand from
24:05health
24:06a lot of health AI, they would have
24:09concentrated based on the data that they
24:11had access to. But, there are nuances in
24:14different parts of the world.
24:16One of the biggest nuances is also the
24:18climatic conditions. I mean, if
24:21think if you brought
24:22born and brought up in Iceland
24:25as opposed to
24:27Bangalore, you're probably used to
24:29different temperatures. Or
24:31um St. Lucian the Caribbean. Naturally,
24:35the environment has a huge difference.
24:37Maybe you're from a major major city in
24:40a landlocked area where
24:43the the probability of respiratory
24:46disease because of pollution is higher.
24:49So,
24:50even health data
24:52is probably so geo-specific
24:56and dietary and all of these different
24:58factors. I know them in the
25:00Mediterranean. Remember seeing the TV
25:02commercials that they live longer there
25:03and they're healthier because of their
25:05diet with olives and olive oil.
25:09And
25:10there's probably some truth in it. I
25:12mean, I don't know the ins and outs of
25:13that.
25:14But of course there are nuances to the
25:16world. Um
25:19That climatological element, let's let's
25:22explore that because I know you have
25:23some some facts and data, some thoughts
25:26and views. And I think it's a very under
25:30under
25:31communicated [clears throat]
25:33discussion.
25:34>> Yeah.
25:35Yeah, you are right that any system
25:38which has been trained in let's say UK,
25:41US, or Europe, or some other country, it
25:44may not work in Asian country or
25:46Caribbean or somewhere else
25:49because the data it it was trained on
25:51doesn't have that those parameters in
25:53it.
25:55All right.
25:56So, we have to think of those things
25:58that a system which were
26:02trained in one country, it may not be
26:05working in the second country and may
26:06not give you the similar results. You
26:08have to retrain the model for that data
26:11to get that data.
26:13And
26:15there is another you might have seen
26:17compass which was
26:20US
26:21government's uh
26:23system for criminals record and it
26:26treats them
26:28giving them the rating on their whatever
26:31they have done the crime and the height,
26:34weight, color, and this.
26:36And it was highly biased
26:38towards one particular
26:40community.
26:42And the results are there and they have
26:43changed it now. Um and it's changed. You
26:46can
26:47search on the web. It's the research
26:49papers that have been was on that.
26:51It's highly
26:54It will give you a rating if you are
26:56certain color with low rating even if
26:57you have done the crime
26:59higher, the other color will get a
27:01higher rated criminality
27:05even though his crime was very minor.
27:08So it's it's it's a change, right? And
27:11coming to the climate part,
27:14all these systems
27:17uh AI machine learning, they were less
27:20energy consuming.
27:22But generative AI machine learning, they
27:24have large
27:26consuming powers.
27:29If I say
27:30that one Google search will give uh is
27:33around
27:34whatever carbon footprint is,
27:37and if you say the similar
27:40search in open AI or Gemini or any other
27:43tool LLM models, it may be 10 times
27:46carbon footprint the other one.
27:48>> Mhm.
27:49>> So that's
27:50And then when you are processing large
27:54amount of data, the data center which is
27:57at the back
27:58and the front end system, the GPU power
28:01required to process that data
28:04is you increase 10 times to 100 times
28:07depending on the complexity of the
28:09problem which you are solving.
28:11>> Mhm.
28:11>> So the carbon footprint is increasing.
28:14So how
28:15So ethical is not just I'm using the
28:18data ethically, but
28:20I am doing I am I am doing the right
28:22ethical things in technology.
28:25That I am
28:26seeing the climate change because it
28:28will be a huge impact on the climate
28:29change it will have.
28:31Yeah, all the data center energy
28:33consumption is going is
28:36multiplying many folds in all countries
28:38because of this.
28:40And that's why data centers people have
28:42opposed data centers establishment in
28:43US. You might have read that report. And
28:46they have gone to India or other places
28:48to establish because
28:50it requires lots of water and other
28:52things.
28:53You don't want that environment to be
28:56uh done bad.
28:58So, which means we need need to think of
29:01how we can use
29:03sustainable energy in these.
29:06Or
29:07can I put my data center in Iceland
29:10where it's naturally cool or places
29:12where it is cool already, put the data
29:14center I don't need more energy to cool
29:16them.
29:17So, can those things be explored
29:19exploited?
29:20>> I mean, a lot of this is common sense as
29:22well.
29:23>> I've been I've I've had this thought in
29:26whilst having discussion with people.
29:28The idea that if the place is really
29:29really cold already
29:31then
29:33surely
29:34it makes more sense. I I I did some work
29:36with
29:38the largest ice cube company in Europe
29:41and the cost
29:43of keeping ice cubes
29:46at a certain temperature in the winter
29:48to the summer, the energy cost was
29:50significantly more. I mean, it sounds
29:52obvious.
29:53>> Yes.
29:54>> But
29:56isn't a lot of this obvious? But for
29:58some re Well, again, I say for some
30:00reason, I think it's probably obvious
30:02why it's not wanting to be
30:06communicated so so deeply because
30:10the financial structure of how a lot of
30:12this works is selling tokens.
30:15So, sell more tokens
30:18then make more money. Although, weirdly,
30:20the math is um
30:22pretty much how much loss is there and
30:25the more tokens that are sold, the
30:27greater the losses. But the idea is to
30:29get
30:31market saturation first and foremost,
30:33then figure out how to become efficient
30:35and increase sp- um
30:38uh revenue per customer further down the
30:41road.
30:41>> Yeah.
30:42>> But here we are today
30:44the drive for more tokens and a token is
30:47just a fancy word for expending compute
30:49power.
30:50>> Yes.
30:50>> And compute power requires water, it
30:53uses energy.
30:56And
30:57yeah, if there's an kind of insatiable
31:00hunger more and more for it.
31:02But I'd I'd like to look at what is that
31:05being used for? For example, is it to
31:08create more dancing cat videos?
31:11Is Is there a practicality? Is there a
31:13logic to it? Is there true value being
31:16created from a lot of it?
31:17>> Yeah.
31:18>> Probably in most instances, if we're to
31:23look at the evidence of
31:24in the internet and social media
31:27probably not.
31:28>> No.
31:29Why you are
31:30putting technology in first place? You
31:32have to answer that question.
31:35Is it to replace people?
31:37Number one. Or is to augment people?
31:40>> Mhm.
31:42>> So, that question we must answer before
31:46we go into technology.
31:47I'm using generative AI this this this
31:50tool, but is it going to
31:53augment people? Can we use a technology
31:55to augment people? And it's not just to
31:59replace people because it's going to be
32:01cost efficient for me because
32:03because I can do That's fine. That
32:05That's also a goal. But in that in that
32:08goal, you can augment people to learn
32:10new technology and be more efficient in
32:13delivering things.
32:14The tokens which we are talking today
32:17they are going to be agents agentic
32:20computing. They are going to be agents
32:22will become so expensive in certain
32:25number of years to maintain them.
32:28It may be costly than a person human.
32:31Because human you have to write that and
32:33it becomes so heavy like for
32:35cybersecurity people are taking now that
32:38we can
32:39give it to agent and it will do that
32:42monitoring of the
32:44uh network and it can take action
32:46quickly and this and that. But
32:50it has to continuously learn and it will
32:52eventually it will become so big a
32:54system in itself.
32:55The running cost of that may be
32:57expensive.
32:59So, are we thinking in those terms? So,
33:01we are we
33:03Of course, we have to make profit out of
33:06technology. We have to make system
33:08efficient. Um we
33:11but never think that yes, it can replace
33:15people. I always think is augment people
33:18with the technology so that it can
33:20deliver more efficient results to you
33:22rather than just thinking in one
33:24direction.
33:25>> Also,
33:27human beings live specific lives. I The
33:29idea of generalized
33:32generalized and scale, yeah.
33:34On the surface,
33:36looks good.
33:37>> Yes.
33:38>> But the reality is when you have a job,
33:41I have a job, everyone has a job.
33:44You have your family, you have your
33:47relationships, your friends. We have our
33:49own.
33:51Generalized is actually an incompatible
33:54theory
33:55with the reality of existing as a human
33:58being. That's not to say you can't
34:01interact, be inclusive, hm, but you
34:04can't have 4.5 million friends that you
34:08go out for a drink with on a Friday
34:10night.
34:10>> Yeah.
34:11>> You can participate in things
34:15where there are a lot of people like
34:16going to a football match or a cricket
34:18match, but you're not speaking to
34:20everyone, you're not engaged with
34:21everyone cuz we're we're bound by the
34:23laws of the physics. So, unless we can
34:26kind of stretch
34:288,000 hours
34:30and fit that into one day.
34:33There's there's practical realities and
34:34also
34:36there are things that you like which are
34:38personal to you. There's
34:40the reason you're good at your job is
34:43the aggregation of your lifetime
34:44experience. It's the team that you work
34:46with, the dynamics of how it is. It's
34:49It's the uncommunicated
34:51information.
34:53Which if we're looking at an agentic
34:55driven large language model, it is
34:57predominantly based around words. And
35:00the amount of times people can be
35:02misinterpreted
35:05because they're not understanding the
35:06nuances and the context.
35:10I think it's almost arrogant and naive
35:12of humans to assume
35:15that
35:16AI can
35:18at least anytime in the near future
35:20become
35:22better at the human experience. And the
35:25human experience has probably become
35:26quite undervalued because the people
35:28that are trying to sell the tokens.
35:30Of course it's in their interest to
35:32convince everyone that this is better.
35:35But we've seen a number of companies
35:37worldwide
35:39get rid of a load of people,
35:41replace them with AI, yeah, and then
35:44understand the consequences to what they
35:47no longer have, lose a lot of their
35:49customer base, and employ the people
35:51back.
35:52And then
35:53hopefully come to the the logical
35:56conclusion which is not how can we
35:58replace people, but how can we improve
36:02people?
36:03>> Yeah.
36:03>> And how can people improve themselves on
36:06the metrics which matter? So,
36:10can you
36:11excel at your job
36:13more efficiently, more enjoyably? Can
36:16you communicate better? They're the
36:18logical things which again, it's not
36:21rocket science. You just It's being
36:24sensible, isn't it?
36:25>> Yeah.
36:26I think it's um
36:27what we are not thinking is
36:30all the
36:31artificial intelligence, fine.
36:34But where is the emotions in it? Where
36:36is the senses in it? Senses means that
36:39can it feel? Yes, with sensors you can
36:42create
36:44close to it, but still emotional
36:46behavior of
36:48you as a human being is different to
36:50mine. To the same situation we will
36:53react to differently.
36:54>> Mhm.
36:55>> Right. So how we are going to capture
36:58that and how
37:00can you think of that mental well health
37:03well-being
37:05these systems will be able to take up.
37:08I think we do not no one is talking
37:11about that.
37:12>> No.
37:13>> And that's a real challenge. I'm talking
37:16to a machine. Okay, I can go on talking
37:18to it or at the other end of the phone
37:21phone or a chatbot is a machine talking
37:23to me.
37:25But for bringing those emotions and
37:28cultural things into it, it's a huge
37:32task.
37:34Cuz one culture is completely different
37:36from the other culture. US is completely
37:38different to UK culture.
37:40>> Mhm.
37:41>> UK is different to Swiss or European.
37:44India, Caribbean, Middle East every
37:47culture has different nuances in it.
37:49>> And different values.
37:51>> Different values.
37:51>> share many.
37:52>> Yeah.
37:53>> But they equally they're all different
37:54values.
37:54>> Yeah.
37:55>> And that plays also into
37:59the formulation of sovereign AI. I
38:03I recently. My interpretation
38:06I welcome everyone to think of their own
38:08interpretation. Sovereign
38:11to me is what's on my phone. My me
38:15speaking to a version of me, more like a
38:17conscience.
38:19Not I can choose to share some images
38:21with some friends, but I have the
38:24choice. Effectively, the only person
38:26that controls my personal thing is me.
38:29As part of the company
38:31in the group, we have a team. So,
38:32sovereign is within the realm of the
38:34company.
38:35There are companies that operate in the
38:38UK, US, France, Germany, various other
38:41places. Now, they have to
38:43especially financial, they have to
38:44respect the SEC rules, um
38:47uh the FDA rules and all the equivalent
38:50and the countries and the GDPR and
38:53Californian privacy. So, but that's
38:56still maybe one company that has to be
38:59sovereign
39:00to the company, but respecting multiple
39:02jurisdictions, right? Then you've got a
39:04country. Obviously, a country is
39:06sovereign. It seems the country-wide one
39:09is the one that has become
39:14the most, let's say, enforced given
39:17recent events where
39:21people internationally understood what
39:23happens when it can be turned off.
39:26>> Turned off, yes.
39:27>> Then it and only since then have really
39:30in the public psyche have people
39:32realized what sovereign means, but I
39:34don't think it ends there.
39:36>> Yes.
39:36>> It keeps going and that goes back to
39:38what you talked about the beginning of
39:40the conversation to do with ethics.
39:42Uh let's explore that.
39:44>> Yeah, I think sovereignty is
39:47great, um
39:49>> [snorts]
39:49>> or the entropic closing switch and then
39:53everyone's
39:54um thinking of because if you are you
39:58thinking of
39:59you are you are using any foreign-based
40:02API or anything, basically your brain
40:05power is there. So, that if you want to
40:09get hold of that back, you need to own
40:12your data. You need to own your
40:14system.
40:15And that can be whether you have your
40:18servers within your jurisdiction
40:22and then you can access the other
40:23systems. And that's we you need to
40:26really think of
40:28that people have started now thinking
40:31Oh, we do I really need this to happen?
40:35And if if that is the case, then
40:38how I can
40:41think of that
40:43it's my data. I have the authority to
40:45use it. Nobody else should be able to
40:48dictate me the terms and conditions.
40:52And
40:53going from one jurisdiction to other as
40:55you rightly said with whichever business
40:57you are, you are going to have
41:01think of what is your strategy
41:04in
41:05short-term, long-term, medium-term, and
41:07long-term.
41:09And what is that
41:11decision you really want to take about
41:13it?
41:14What what I mean by that is
41:16that
41:18why data is exposed, to whom?
41:22And it should be available to people.
41:25It It comes to the same old traditional
41:28CIA as we call it. Confidentiality,
41:31integrity, and
41:34adaptability.
41:35Or accessibility, right? That it should
41:38be confidence enough in integrate means
41:41validated data is coming. And accessible
41:45to the right people at the right time.
41:49That is very important.
41:50>> Availability.
41:51>> Yeah, availability.
41:52>> Yeah.
41:52>> So, right it should be available to at
41:54the right people at the right time. If
41:56it is not, then you need to really think
41:59of
42:00what's your role model here, that who
42:03you are giving access to, and who is
42:05managing it.
42:07And for making
42:10overall
42:12accountability of your data. And that's
42:14where again coming to the transparency
42:16with be with your customer or client or
42:19citizens telling them, yes, your data is
42:22safe and this is within this
42:23jurisdiction and we have got hold of it.
42:26It's not going beyond these walls and
42:29these are the
42:30SLAs or these are the things which we
42:32have put in place.
42:35So, I think
42:36people are worried because they don't
42:39know what's happening
42:40to their data.
42:43>> Well, I think there's from what I've
42:45seen from several universities and a
42:48number of professors that I know,
42:51this country, as in the UK,
42:54>> Mhm.
42:54>> is really in prime position
42:57for the convergence of
43:01I think the inevitability
43:03that
43:04compute will become democratized.
43:08Because it has been very expensive and
43:11it has been data center led. But as it
43:15becomes possible
43:17to effectively have the compute Maybe we
43:19just look at history. Things have got
43:21smaller,
43:22cheaper,
43:23more powerful,
43:24>> Yeah.
43:24>> and more efficient.
43:25>> Yes.
43:26>> Now, we can say
43:28this is different, but effectively
43:31it's
43:32the all evidence shows the contrary. So,
43:35what is the logical
43:37place that we will end up if actually
43:41the compute we hold in our hands
43:43>> Mhm.
43:45>> is
43:46more than adequate and extremely capable
43:50of being the place where the compute
43:51happens. Because I look at it this way.
43:54We have a data center
43:56in even the UK. But data center in the
43:58UK, data center in the US. We have to
44:00take the information
44:02to the data center. The compute happens
44:04and it gives it back.
44:06Well, the safest, most private, most
44:08secure, the most sovereign it can be is
44:10if the data never needed to leave in the
44:12first place.
44:14We may still be
44:16a few months away from this, but it's
44:18definitely the trajectory of Apple and
44:20Google.
44:22So, that I think that could change the
44:24dynamics a lot, but it also opens up
44:28the next topic,
44:30which I think will be again really
44:33important to hear your perspective.
44:35Looping back to the guest that we had,
44:37who was the the CEO of the UK's largest
44:41quantum computer company.
44:45When quantum
44:46reaches the point,
44:48whether it's data center or wherever it
44:51is, but
44:52the capability that will bring
44:55will again change everything. Let's
44:58discuss quantum.
44:59>> Yes. Definitely, that's um
45:03with quantum computing
45:05>> [snorts]
45:06>> the horizon which we are thinking was
45:09around next 5 years, it should be
45:11available in labs, um in controlled
45:14environment, but because currently it is
45:16in very cryogenic environment, sub-zero
45:19temperatures, quantum computer can work.
45:23But in 10 years, we hope that technology
45:25is good enough, and that's where all the
45:28companies are working towards.
45:29>> Mhm.
45:31>> That it should be a kind of a laptop or
45:33something or desktop at least available.
45:37Leave that apart, but with quantum,
45:40the
45:42it will challenge everything
45:44which we are currently doing,
45:47uh from
45:49cryptography, because current
45:52RSA algorithm, which is highest
45:55encryption which is available to you is
45:572048 bits.
45:59>> Mhm.
46:00>> Well, that's the highest one.
46:03And current computer has been it will
46:05take years to solve it, but with quantum
46:08computers may take
46:10few hours only to break that encryption.
46:14And and that may be lesser than that
46:16also.
46:17And
46:18that's the real challenge.
46:20And second is
46:23quantum communication channels which
46:24people have started working then by
46:27working on 5G, 6Gs. China is already
46:30started working on quantum communication
46:31channels
46:33because then the data is more secure
46:35which we are transferring.
46:37And of course companies are using
46:40quantum
46:41to
46:42do simulations, solving complex
46:45problems, optimization problems
46:48like some of the banks like JP Morgan,
46:50Barclays, HSBC. They have started for
46:53the investment banks.
46:55>> Mhm.
46:56>> And then because then it can give you
46:57calculations
46:59speed which is normal computer cannot
47:02give you.
47:03Right? So that means
47:06And of course it gives us other
47:08possibility of drug simulations.
47:12Like for COVID, the drug came in 6
47:15months with simulations which we had at
47:17that point of time.
47:19But if quantum had been available, it
47:21could have taken hours only to give the
47:24same results.
47:25>> Mhm.
47:26>> The huge pharmaceutical AstraZeneca and
47:29other companies they have started using
47:30it.
47:32The algorithms have started, but they
47:33are used on classical computers because
47:36so you can't get that speed at at the
47:39moment, but still you can see that yes
47:41it is available it can do the work.
47:45The challenge current quantum has
47:48bringing is like currently the error
47:51rate is high.
47:52>> Mhm.
47:53>> And people are trying to do work on
47:56that. That's a big research area. Anyone
47:58can build good algorithms and that will
48:01be good
48:02to showcase and build your business into
48:04it.
48:06So, if you would reduce the error rates
48:08in short algorithm
48:11by certain percentage, it can be very
48:13good.
48:14And then
48:16the second challenge comes is, of
48:18course, the energy consumption.
48:21The current our normal laptops or
48:23computers are taking around
48:251 kW to 2 kW, but quantum's are taking
48:29100 kW or something like that. It's a
48:31huge because they need sub-zero
48:33temperature at the moment. Yeah. But
48:36maybe 5 years down the line that power
48:38will be reduced. So, there are huge two
48:40areas to work. One is decreasing the
48:43error rate basically optimizing those
48:45algorithms
48:47which are whichever are there for
48:49simulation, for error rate, for
48:51encryption, decryption, and
48:54of course, on the sustainability side of
48:56things, that how you can make the
48:57hardware more
48:59adaptable to the environment.
49:04>> I mean, it is a fascinating technology.
49:07And
49:08I think um
49:11James was telling me
49:13the big issue is writing the data.
49:17That's
49:18when that problem is really solved
49:22things begin to move
49:23>> Yeah.
49:24>> at incredible pace.
49:25>> And also
49:27the currently what uh
49:29I'm also part of that, that we are
49:32trained with digital world, zeros and
49:36ones.
49:36>> Yeah.
49:38>> Right?
49:39With quantum, you have to unlearn that a
49:42bit.
49:44Because now it can assume a state which
49:46is zero and one both.
49:48>> Yeah.
49:49>> So, that's a challenge and the digital
49:52gates, as we said, to build the
49:53technology, they are all different in
49:55digi- in zeros and ones. Digi- But, in
49:58quantum, the gate is completely
50:00different gates.
50:01>> Yeah.
50:02>> And you have to learn that. So, you have
50:04to unlearn The unlearning part will be
50:06more difficult. Learning new things is
50:09easier.
50:11If you have not learned that zeros and
50:12ones, because the
50:14If you see the our human life,
50:16it's it's not zeros and ones.
50:19It's It takes the state in between,
50:21also.
50:22>> Yeah.
50:22>> Right? We are not sad, we are not happy
50:26at one point of time. We are
50:28go through that phase, also.
50:30>> More fluid, isn't it?
50:31>> More fluid.
50:31>> Yeah.
50:32>> So, that's where quantum is building
50:34that state, giving you those stages in
50:36between that you can be this. And that's
50:39where it becoming powerful.
50:41>> So, it's going from two-dimensional to
50:42three-dimensional, as far as
50:43>> Multidimensional, you are saying.
50:46>> I think um
50:48what we have not discussed is uh
50:52metrics. How you are going to measure
50:53the success
50:56of any AI system you are implementing?
50:59What's the success criteria for you, for
51:02any organization?
51:04>> With this, it kind of goes back to what
51:05you've just said there, as well,
51:07which is where do we have to unlearn a
51:10lot? We're looking [clears throat] at
51:12things of how they've always been. Yeah,
51:14things are fundamentally changing. Yeah.
51:17And I
51:19I hear quite often people talk about
51:21benchmarks. But, whose benchmarks? To
51:24what degree?
51:26So,
51:27are we talking about the benchmarks of
51:295.
51:316 is better than 5.4, 5.7
51:35is better at specific calculations,
51:38looking at the utility of how things
51:40work,
51:41>> Mhm.
51:41>> looking at what it's you said success
51:44looks like.
51:45>> Yeah.
51:47>> Because things are new, people are
51:48accepting. Poor benchmarks is how we
51:50measure.
51:51>> Mhm.
51:52>> But is it? I mean, if if you want to be
51:54healthy, you don't look at is this a
51:56good apple or a bad apple I'm buying
51:58from the store.
52:00A good measure of your health is how do
52:02you feel after a prolonged period of
52:04time?
52:05>> Yeah.
52:05>> Do you have better blood pressure? You
52:07can look at certain metrics, but that
52:09doesn't mean you're happy.
52:11Maybe you want to also look at all the
52:15dimensions.
52:16We can see as human beings
52:19what good should look like.
52:20>> Yes.
52:21>> And what bad looks like, and what less
52:23bad, more good.
52:24>> Yeah.
52:24>> But that it it seems as a complete
52:26misalignment between the measurements
52:28and the metrics.
52:30Because
52:32people haven't had enough time to learn
52:35or to change their thinking and almost
52:37adopting
52:39or have adopted the metrics of the
52:42people selling the products. But I see
52:45that is wearing thin now, and people are
52:46questioning that.
52:47>> Yeah.
52:48I think the challenge which is coming to
52:51us
52:53um
52:55people are just
52:56thinking we are it can do this for us.
53:00It can write a business case for me. It
53:02can write policy for me. It can do a
53:05code for me. It can do a
53:07a cyber security hacking code for me on
53:09this.
53:10But
53:12have you really defined real what what's
53:14your your problem which you are trying
53:16to solve?
53:18What is the elephant in the room you're
53:20trying to solve, right?
53:22You are you answering that question
53:23yourself first?
53:25Is
53:28that if I had to think of
53:31that I have hundreds of problems
53:34what is that
53:35But what is that one main problem which
53:38I want to solve?
53:39And what's the outcome I want out of it?
53:44Are we really defining that path?
53:48And for defining that path, I have to go
53:50back.
53:51Okay, I want this to happen, but for
53:54this I would need
53:56this hardware, this is the data, this is
53:59the software, this is the algorithm,
54:01this is the model.
54:03Like in
54:04generative AI or AI or IoT,
54:08there are different models, different
54:10algorithm, different goal achievements.
54:13Unless I choose the right model,
54:16like for giving you example, for health
54:19care
54:19um
54:20um
54:21let's say there are x-rays, there are
54:23MRI scans, CT scans, and things like
54:26that.
54:28For that,
54:29what kind of image processing algorithm
54:32will work better
54:34rather than the normal um doing the
54:36portraits or photographs.
54:38>> Mhm.
54:39>> It's a completely different world
54:40scenarios.
54:41So, you have to choose the right
54:43algorithm there and the right model
54:46which can process that quickly and be
54:48more efficient rather than choosing a
54:50model which is not suited for that
54:52activity.
54:53>> Mhm.
54:53>> So, I have to spend time in doing that
54:56analysis
54:58to see that oh, this is my problem
55:01statement.
55:02For doing this work, I need to use this
55:05model, this algorithm, this data.
55:08And that becomes my work.
55:11So, unless I do that,
55:13I will not be able to get efficient
55:15results. I may get results,
55:18but they will not be as efficient as I
55:21want.
55:22>> You see, that's completely at odds with
55:25the narrative which is pushed. If we if
55:28we were to describe
55:29why have people taken on AI?
55:34And I'm talking mainstream, whether in
55:36the corporate world, perhaps in the the
55:39public sector,
55:41it's for efficiency,
55:43it's for speed,
55:45and
55:48>> Well, I mean, they're the main two,
55:49really, aren't they?
55:50>> But what is needed
55:52>> [laughter]
55:52>> is the extreme opposite to those two to
55:55actually get value. And of course,
55:57people have uh they saw the report, and
55:59it's been referred to on this podcast
56:01numerous times by guests, the complete
56:04inefficiency
56:05of and the lack of ROI for pilots.
56:08>> Yes.
56:08>> Because people have thrown things out
56:10there, they've not done the research,
56:13they've not thought it through, and then
56:15they've been surprised when no value was
56:17found.
56:19And it
56:20it's almost a form of insanity in a way.
56:23>> Yeah, I think what people have been
56:25thinking, oh, it can
56:28I will give it Yes, it has given some
56:31efficiencies to some people. Like, I can
56:33write a letter, I can write an email, I
56:35can write this very small basic jobs,
56:38which is fine.
56:40Right? It can refine the email for you,
56:42it can refine research paper for you, it
56:45can do this, it can do that, but it
56:47still hallucinates.
56:49If I say to my AI any chat
56:53model that write me a research paper on
56:55XYZ with references, with this, with
56:58that, it can write a beautiful paper
57:01with references, everything. But if you
57:03take the references, 50% of them may be
57:05fake references. They don't exist.
57:08>> And this happened in a court case a few
57:09times now, I believe.
57:11>> Yeah.
57:11>> Where it's referred to case law, and it
57:13just made it all up.
57:14>> Yes.
57:15>> I think you can't do that with humans'
57:17lives in
57:20in health care. You can't do it in
57:22regulated environments like finance.
57:25>> Yeah.
57:25>> You
57:27There's certain non-negotiables.
57:29>> Yes.
57:30>> Actually, what's more important, doing
57:32it quickly or getting it right?
57:36>> Getting it right.
57:37>> I I would agree.
57:38>> Yeah, so getting it right is
57:41uh
57:42preference that for any system which is
57:45going to affect people, if it's going to
57:47affect citizens or your clients or
57:49anything, it has to be
57:52right.
57:53Quick and dirty, if you are willing to
57:55say, quick and dirty can be for some
57:58analysis which you are using for your
57:59internal purpose. I just want
58:01to quickly give me this analysis. That's
58:04fine.
58:05You can use it for quick and dirty jobs,
58:07which is not going to affect any other
58:09thing, but just for your own analysis.
58:11But for
58:13getting your company
58:15into a business model, good product,
58:18good strategy, I any
58:21good services to your clients, then you
58:23have to do it the right way. And
58:25therefore, doing the right way,
58:28you have to go to the basics back to the
58:30basics.
58:31>> Yeah, that yes.
58:32>> Do I have got the data right?
58:35For any CRM solution also to work right,
58:37you have to have a database of customers
58:40so that it can just spit out the right
58:43dashboard for you out.
58:45>> Cuz
58:46especially some of these systems, they
58:47are the record of truth. Yes.
58:51Put something in where
58:54the record of truth is subject to no
58:57longer being true,
58:59then it has no utility.
59:01>> Yeah.
59:02>> In in its form and function. So, it's
59:04going back to the points and purpose of
59:05what are you trying to achieve? Then of
59:07course,
59:08in some aspects, and I know
59:11that it's been referred to recently a
59:12few times that the number one use of AI
59:16is for
59:18effectively chatbots for um emotional
59:22support.
59:24And someone mentioned to me last week,
59:26the ninth most popular
59:29use
59:30worldwide of AI is to get their daily
59:34astrology.
59:36>> Yeah, which
59:38>> you can kind of understand and there's
59:40nothing wrong with that.
59:42But
59:44on one side
59:46do you take the same thing and and say
59:50we'll accept the same
59:52outcomes is the same thing on the line.
59:55If if you ask her, "What's your star say
59:57today?" And you oh, "You're going to
59:59have a good day." Yeah, okay, maybe that
1:00:01has a psychological impact, but that's
1:00:04between you and you. Yeah, right? That's
1:00:07no different from you know, eating
1:00:09something which makes you happy.
1:00:10>> Yeah.
1:00:11>> It's it's a simple thing.
1:00:12It's not the same as implementing into
1:00:16potential government policy which could
1:00:17lead to affecting tens of thousands of
1:00:20people's lives or in a corporate where
1:00:23it could lead to millions in losses and
1:00:25subsequent
1:00:27impact to customers.
1:00:29It's
1:00:30>> It's like if you are using the system
1:00:32for in like healthcare, you have
1:00:34a healthcare I am training robots to do
1:00:37the surgery. I have to give it the right
1:00:39data that how it has to be done.
1:00:42Or if it has to interpret, okay, these
1:00:45are this is the MRI scan of this person.
1:00:49Can you tell that which disease this
1:00:52person may get or CT scans or this and
1:00:54blood test reports of this person for
1:00:56the last 5 years.
1:00:58If it has to interpret it and say that
1:01:01this person may be prone to a heart
1:01:03attack or a stroke or this or that,
1:01:08it has to have millions of data of
1:01:10patients in that area.
1:01:13It it will not be work that if I get
1:01:16millions of patients data from India and
1:01:18put it here and say this will work here.
1:01:20It will not Because it's completely
1:01:22different demographic and people have a
1:01:25different lifestyle, different eating
1:01:26habit, different culture, different
1:01:28environment, weather, everything is
1:01:30different. So, I have to build the data
1:01:32here so that it can give me the right
1:01:35aspect output from the art that oh, yes,
1:01:39I can see this blood test this and I
1:01:41have hundreds of blood tests and similar
1:01:43these reports and this person has got
1:01:46this this this disease in future. This
1:01:48may person is prone to this. Mm.
1:01:52>> So, so
1:01:54depending on where anyone is and they
1:01:56could be anywhere in the world. Yeah.
1:02:00To would to give them a little bit of
1:02:02bit of advice, whether they it's for
1:02:06personal use of AI, whether it's
1:02:09public sector or private sector.
1:02:12And irrespective from their eyes, their
1:02:14perspective of sovereign, anywhere in
1:02:16the world, but
1:02:18talking directly to them, what three
1:02:20little pieces of advice would you give
1:02:22them of how to think
1:02:26about
1:02:28using, adopting
1:02:31and working with AI?
1:02:36>> So,
1:02:37if
1:02:40the
1:02:40if it is three,
1:02:42let's see.
1:02:43You have you have you got with more.
1:02:45Uh first is what is your legacy data
1:02:48system and legacy data?
1:02:51Can that those systems integrate with
1:02:54the new systems?
1:02:57Can you solve that problem first of all?
1:03:02Because that is very important. Legacy
1:03:04data may not be the right
1:03:06I should not say right data. It may be
1:03:07the right data, but it may not be the
1:03:09data which can be used by the AI
1:03:12systems.
1:03:14So, deep [snorts] work on that. So,
1:03:16that's the first step.
1:03:18And if you are modern new company, you
1:03:22have growing,
1:03:24then the second challenge is that
1:03:28um
1:03:30do you have the right set of data with
1:03:33you?
1:03:34What problem you are trying to solve?
1:03:38And then
1:03:40um again, I will come into
1:03:44which area of work you are in and choose
1:03:47that right model, right algorithm, right
1:03:50goal,
1:03:51so that it can give you the eff-
1:03:54uh efficiency, um cost savings, and
1:03:58other benefits, return on investment,
1:04:00which you are thinking of.
1:04:04And the last one I would say,
1:04:07do you have the right skill set, right
1:04:10talent with you,
1:04:11who can make these systems work?
1:04:14>> Mhm.
1:04:15>> Because don't
1:04:18think of that, that how do we
1:04:21bring that capability within our current
1:04:24workforce,
1:04:25or we hire a new workforce?
1:04:27>> Mhm.
1:04:29>> [snorts]
1:04:29>> But the new workforce which you are
1:04:31bringing in, they may not have the
1:04:32functional knowledge,
1:04:34>> Mhm.
1:04:34>> which your company or our product is
1:04:37doing. They may have the knowledge of AI
1:04:39or technology, but they don't know how
1:04:41the system works.
1:04:43If I have to make, let's say,
1:04:46accounting system with the AI can do the
1:04:49first ledger and this and that,
1:04:51everything it will generate for me, uh
1:04:53picking everything from different
1:04:55systems, HR, finance,
1:04:57um transport, I'm bringing those
1:04:59together and it can generate balance
1:05:01sheets for me.
1:05:03But
1:05:04but does that person have person who is
1:05:07building the system uh the knowledge of
1:05:09accountancy?
1:05:11So, where is that functional knowledge?
1:05:12I need a doctor with me working to build
1:05:16a system, a telemedicine system.
1:05:19I will not be able to build as a
1:05:21technologist myself. I need to work with
1:05:23them.
1:05:25So, that's where people are forgetting
1:05:27where the right then to how you are
1:05:29going to build that talent.
1:05:31>> Mhm.
1:05:31Collaboration.
1:05:32>> Collaborations is the key.
1:05:34>> Yeah.
1:05:34>> So, we bring that collaboration
1:05:36together, bring people together into the
1:05:38room. So, that you can work on that
1:05:40problem.
1:05:41>> Mhm.
1:05:41>> So, thinking of that problem, who will
1:05:43be the right set of people? Even if I
1:05:45have to do the investment banking,
1:05:48right? I need to see, okay, I
1:05:51as an investment banker, I see
1:05:52Bloomberg, I see London Stock Exchange,
1:05:55New York Stock Exchange, Japanese and
1:05:56this and that. Then I take decision on
1:05:59investing or taking money out, putting
1:06:02it in there, putting it there, taking
1:06:04from this out or in this commodity or
1:06:07this
1:06:08market.
1:06:10So, somebody has to teach the machine.
1:06:12>> Yeah.
1:06:12>> How to do it?
1:06:14So, that knowledge, you have to
1:06:16collaborate with those people to build
1:06:17that knowledge. So, I mean, I think we
1:06:20are not spending time
1:06:22on train the right talent at the moment
1:06:26that
1:06:27do we have the right talent in the
1:06:29organization or building?
1:06:32And how we are going to retain them
1:06:33also?
1:06:34>> Yeah.
1:06:35Cuz it's subject matter expertise as a
1:06:38phrase.
1:06:38>> Yes.
1:06:39>> With the case of Oak North.
1:06:40>> thing, the subject matter expertise
1:06:42>> Yeah.
1:06:42>> with the people that understand AI. In
1:06:45many ways, the people that understand
1:06:48the new wave of AI are quite different
1:06:50to the people that understood
1:06:52infrastructure technology,
1:06:54productionization.
1:06:57>> Which are different to the people that
1:06:58are the users and people that understand
1:07:02the business cases. And in many ways,
1:07:04it's figuring out how to get these
1:07:06people to work together is as you
1:07:08rightly say it's um
1:07:11It's the challenge that people don't
1:07:13realize is the challenge.
1:07:14>> Yeah.
1:07:15>> It needs to be revealed, understood, and
1:07:18addressed.
1:07:19>> Yes.
1:07:19>> I think in many ways there are no
1:07:20shortcuts.
1:07:21>> There are no shortcuts. Like I worked on
1:07:25robotic arm during my college days and
1:07:28we built a complete robotic arm which
1:07:29could seven degrees of freedom and it
1:07:32can do all the stuff on going from one
1:07:35place to another doing things. But it
1:07:38has to be trained to do that activity
1:07:40again and again.
1:07:41And for objects smaller, bigger, this,
1:07:43that, it has certain size only that size
1:07:46up to that size it can pick and putting
1:07:48it there.
1:07:49So those things take time.
1:07:53Any robotic arm, robotic surgery
1:07:56doctor has to train that model that this
1:07:58is how it has to be done. This is how
1:08:00the cut has to be done. This is how it
1:08:01has to be done then stitched back and so
1:08:04on. So that those
1:08:06those things have to be built in and for
1:08:08that it requires
1:08:10time training and well as well as
1:08:14perseverance with the system that it
1:08:16becomes perfect.
1:08:18>> Yeah.
1:08:18>> Otherwise the system may not may work
1:08:20but it will not be good enough.
1:08:23>> Yeah.
1:08:24>> So you need to define what should good
1:08:25look like. What is your problem
1:08:27statement and then think of the process
1:08:30in between that how I reach from this
1:08:32place to this place doing the right
1:08:35things. And then and it can be
1:08:37sustainable
1:08:39>> Yeah.
1:08:39>> over the longer period of time and it's
1:08:41continuously learning auto you itself.
1:08:45But that is the success of any machine
1:08:47learning platform.
1:08:48>> Mhm.
1:08:48>> If you don't put that loop back
1:08:51any machine learning system will fail.
1:08:53It may work first time, 100 100 times,
1:08:561,000 times but it may fail
1:08:591,001 times.
1:09:00>> Yeah.
1:09:01>> Right? So that's the challenge. So, that
1:09:03unless it learns itself. So, who
1:09:06for that, you need resources, you need
1:09:09technology, you need servers, GPU power,
1:09:12and database growing.
1:09:13>> Yeah.
1:09:15And and it goes back to what you're
1:09:17saying before.
1:09:19See how things scale up, understand
1:09:22where things will go and what the
1:09:24consequences are.
1:09:25>> Yeah.
1:09:26>> Where you Where you begin is not
1:09:28necessarily where you'll end up. And
1:09:30unless you put consideration in now
1:09:33>> Yeah.
1:09:33>> and up front
1:09:35things are not going to work out as well
1:09:36further down the line. So,
1:09:38to to wrap up, your last thought.
1:09:42>> My last thoughts are
1:09:44it's a exciting times.
1:09:47For the I haven't seen technology moving
1:09:50from cassette recorders to
1:09:52mobile phones to AI machine learning.
1:09:55It's exciting time to be in.
1:09:57And we should not be
1:10:00fearful of the technology.
1:10:03Uh
1:10:03we have to embrace it.
1:10:05And I always says from the leaders, top
1:10:09leaders, if you have to say
1:10:12the technology should be as I said
1:10:13earlier also to is to augment people
1:10:16rather than replace people. If you work
1:10:18with my that mindset technology is going
1:10:21to give you more benefits
1:10:23and also bring more exciting solutions
1:10:27for you or exciting products for you.
1:10:30So, I think that's the strategy we
1:10:31should have.
1:10:33Um
1:10:34and of course, we I cannot deny for
1:10:37again again stressing on your
1:10:40data, algorithm, models, they have to be
1:10:43in the right place.
1:10:45From the hardware side as well as the
1:10:47software side to make the systems at
1:10:49work efficiently. Otherwise,
1:10:51um you will not be able to reap the
1:10:53benefits. And the return on an will not
1:10:56be there. And up 1 year down the line
1:10:58you will be saying I'm wasting my money
1:11:00over this.
1:11:00>> Mhm.
1:11:02So, back to basics and build out from
1:11:05there.
1:11:05>> Yes.
1:11:06>> Excellent. Well, Ravinder, it's been a
1:11:07pleasure.
1:11:08>> Thank you.
1:11:09>> Thanks again.

About this transcript

This page contains the full transcript of GAEA Talks - Human In The Lead, Not Human In The Loop with Cabinet Office's Dr Ravinder Singh by GAEA Talks, generated from the public captions YouTube serves with the video. The transcript has 10,431 words across 1,941 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.