Speaker: 00:00:12

What's your experience over time in terms of observing digital twins in that domain and

what do you think that means in terms of applying those concepts to this very complex

00:00:26,181 --> 00:00:28,863

biological being that is a human?

00:00:29,601 --> 00:00:41,611

I think uh this is actually very well on point with the question for world models

currently, because essentially we have the same problem here as with the world models.

00:00:41,611 --> 00:00:44,273

First of all, what's world model?

00:00:44,273 --> 00:00:46,475

World model is not an architecture.

00:00:46,475 --> 00:00:49,638

Sometimes people say solution is a world model.

00:00:49,638 --> 00:00:51,359

World model is not an architecture.

00:00:51,359 --> 00:00:58,245

It's uh not something that I can say, well, this is the mathematical formula of

00:00:58,950 --> 00:01:12,071

World model is a um to have either it's a neural network, either it's something else that

can predict the events and the effect of these events.

00:01:12,071 --> 00:01:13,753

So for example, you

00:01:13,969 --> 00:01:22,689

I don't know, like you drop something and then you hear the sound, it breaks, it jumps a

few feet away and so on.

00:01:22,689 --> 00:01:24,269

So this is world model.

00:01:24,269 --> 00:01:29,349

So predict the consequence of actions from your actions in as good as possible.

00:01:29,829 --> 00:01:37,909

Currently we have a switch to a world model, but a lot of world models, example, for

example, Sora 2 is supposed to be a world model.

00:01:38,629 --> 00:01:42,447

But that equates to a lot to a video model.

00:01:42,447 --> 00:01:52,242

that does not understand physics, uh rather learns from YouTube videos or whatever corpus

of videos it has there.

00:01:52,242 --> 00:01:54,594

again, it's a transformer.

00:01:54,594 --> 00:01:57,145

So world model can be a transformer.

00:01:57,145 --> 00:02:03,828

So predict the sequence of events without actually understanding the physics and the rules

behind this.

00:02:04,189 --> 00:02:11,773

There are other architectures, um for example, state space models that introduce

recurrency.

00:02:12,082 --> 00:02:16,662

But again, this also works to a certain extent.

00:02:16,662 --> 00:02:21,902

Then we have Yan Li Kun going into this direction.

00:02:21,902 --> 00:02:25,842

Like Fei-Fei Li has her startup on world models and digital.

00:02:25,926 --> 00:02:39,566

A twin is essentially something that needs a world model because if you're an engineer, if

you do something wrong with this digital twin, it should electrocute you with all the

00:02:39,566 --> 00:02:43,599

consequences from the machine, burn everything around there.

00:02:43,599 --> 00:02:46,081

If you drop something, it should react accordingly.

00:02:46,081 --> 00:02:48,443

Otherwise, it's not really a twin.

00:02:48,443 --> 00:02:51,446

It's not even a distant cousin anymore.

00:02:51,446 --> 00:03:04,416

And of course for humans, how difficult it is in the human body where, let's say, you

damage something, like how much impact it can have, and predicting these possible events.

00:03:04,416 --> 00:03:11,541

uh Currently, the world model will probably even struggle, like if you lose that much

blood, how much blood is left in you.

00:03:11,541 --> 00:03:13,862

They will even struggle to predict death.

00:03:13,862 --> 00:03:19,026

uh So this is something uh that...

00:03:19,026 --> 00:03:23,837

we are very far in digital twins because we don't have world models.

00:03:23,837 --> 00:03:27,017

I remember there was lot of hype for digital twins.

00:03:27,017 --> 00:03:31,537

In the automotive industry, everyone was building these cars.

00:03:32,017 --> 00:03:41,928

To test them, was an idea, for example, to do the crash tests with digital twins and

whatnot because crash tests are very expensive.

00:03:41,928 --> 00:03:44,612

I feel like without a good world model, that would be

00:03:44,612 --> 00:03:46,092

extremely difficult.

00:03:46,092 --> 00:03:47,112

will not be a twin.

00:03:47,112 --> 00:03:49,072

It will not be even a second cousin.

00:03:49,072 --> 00:03:55,572

It will be completely unrelated or a stranger digital rando.

00:03:55,932 --> 00:03:56,741

it's a

00:03:56,741 --> 00:04:09,366

so the, I come across this concept and I know enough, certainly not to the level of detail

that you do about the direction of what's coming.

00:04:09,366 --> 00:04:23,128

m But I know enough to recognize that that complexity of the human entity, because I've

been involved in clinical care and healthcare for

00:04:23,128 --> 00:04:28,872

40 years now, both as a practitioner and then selling technologies into the space.

00:04:30,074 --> 00:04:34,438

I knew enough that I went digital twins.

00:04:34,438 --> 00:04:37,600

That term is being thrown around a lot.

00:04:37,881 --> 00:04:39,622

What does it mean and why?

00:04:39,622 --> 00:04:43,105

So I started researching and I learned about the roots of digital twins.

00:04:43,105 --> 00:04:51,602

It's pretty quick and easy for anybody to be able to go and study that and understand it,

at least at a superficial level.

00:04:51,818 --> 00:05:00,175

And I very quickly then looked at some of the applications of it and I did a literature

search for digital twin in healthcare.

00:05:00,275 --> 00:05:15,398

And I came out where I decided I created a uh maturity model, a very simple four level

maturity model that is in an article I wrote recently that um I wish I could get wider

00:05:15,398 --> 00:05:20,670

uptake because I think it's a topic that needs to be exposed more, which is

00:05:20,670 --> 00:05:24,621

lots and lots of things are being called a digital twin, which aren't.

00:05:24,902 --> 00:05:36,467

If you apply the pure interpretation, like we just described, of what a digital twin

should be capable of and what Sam Altman and others allude that is going to be

00:05:36,467 --> 00:05:47,771

deliverable, you know, when they make their presentations and things, and then you break

it down to about what is actually being delivered and labeled a digital twin, it can be a

00:05:47,771 --> 00:05:48,912

dashboard.

00:05:48,912 --> 00:05:49,976

That's level one.

00:05:49,976 --> 00:05:55,416

And literally just here is another representation of the data we know about this patient.

00:05:55,416 --> 00:06:06,556

And it might be that it's interrogable, you know, I, you can ask questions of it, and then

you throw a chat bot or through some other mechanism, but it's still a dumb representation

00:06:06,556 --> 00:06:07,696

of data.

00:06:07,696 --> 00:06:12,296

Then you can apply some algorithmic prediction to it.

00:06:12,296 --> 00:06:19,396

And that would be level two, so that you try to create a predictive model out of the data

that's there.

00:06:19,416 --> 00:06:25,796

And maybe you are able to start to draw conclusions from that data.

00:06:25,796 --> 00:06:27,636

Well, that's just analytics.

00:06:27,656 --> 00:06:30,256

That's nothing particularly clever.

00:06:30,256 --> 00:06:36,376

And a lot of the models that are being applied are just applying those analytics

principles quicker.

00:06:36,556 --> 00:06:36,736

All right?

00:06:36,736 --> 00:06:45,547

They're basically using the compute and reasoning power to do it quicker than a human

would do with the BI tools that are out there.

00:06:45,547 --> 00:06:58,828

The third level is then it being updated in real time so that as the patient's condition

changes, that you're able to add new data in real time so that you're able to create a

00:06:58,828 --> 00:07:03,342

dynamic model that is of the patient that is actually changing.

00:07:03,342 --> 00:07:15,541

And it's only once you then synthesize that all to a reasoning against the body of

corporate, not a corpus of knowledge or professional knowledge.

00:07:15,563 --> 00:07:18,104

that you can get to that level four level.

00:07:18,405 --> 00:07:29,823

The big problem in terms of the evolution of it in healthcare right now is there are very

narrow applications because because the human body is so complicated, very narrow

00:07:29,823 --> 00:07:38,779

applications whereby even for those dashboards, they're dashboarding either a single body

system, i.e.

00:07:38,799 --> 00:07:44,683

the cardiac system or a single disease like diabetes.

00:07:45,131 --> 00:07:55,475

There is no way yet where anybody is trying to, even at a dashboard and a predictive

level, model an entire human being.

00:07:56,076 --> 00:08:08,641

Because of the complexity of all of the individual body systems, all of the individual

organs, all of the potential disease processes that exist out there, and the killer is

00:08:08,641 --> 00:08:09,881

genetics.

00:08:09,981 --> 00:08:13,773

The wonderful thing about the human body is it is constantly evolving.

00:08:26,226 --> 00:08:30,826

sometimes in good ways in terms of the way our body genetics change.

00:08:30,826 --> 00:08:33,906

Everybody used to think that your DNA was fixed.

00:08:33,906 --> 00:08:37,806

You were born, you inherited what you inherited from your two parents and that was it.

00:08:37,806 --> 00:08:42,266

No, genetics change over time.

00:08:42,366 --> 00:08:51,046

And so therefore that's why you need that third level of the real time inputs because what

your body state was five years ago is no longer your body state.

00:08:51,346 --> 00:08:53,402

For example, I've had a transplant.

00:08:53,402 --> 00:08:57,606

I now have cellular material in my body that belonged to somebody else.

00:08:57,606 --> 00:09:05,011

So therefore that actually then starts to affect the rest of my DNA, which is why I have

to take anti-rejection drugs and so on and so forth.

00:09:05,012 --> 00:09:17,421

So you look at all of that complexity, it means that the potential network of models and

data sources for all of this is enormous.

00:09:17,582 --> 00:09:23,086

And I think that that's one of the reasons why I try to sort of like, um you

00:09:23,314 --> 00:09:29,708

poke at this concept of, that's used from a marketing perspective of a medical digital

twin.

00:09:29,708 --> 00:09:33,683

Our solution provides a medical digital twin of the patient, blah, blah, blah, blah.

00:09:33,683 --> 00:09:39,051

And it's not one, it's incredibly narrow focus and two, it's probably a early maturity

level.

00:09:39,051 --> 00:09:42,571

It is not a level four application of medical digital twin.

00:09:42,571 --> 00:09:44,242

So there you go.

00:09:44,242 --> 00:09:45,803

That's my little advert for.

00:09:45,803 --> 00:09:51,878

for my reasoning and thinking and why I say I'm optimistic, I think that we will get

there.

00:09:51,878 --> 00:10:04,538

um Maybe not in my lifetime, but we will get there and then it will become like Star Trek

and just uh the computer will know everything and help the doctor be able to instantly

00:10:04,538 --> 00:10:06,330

diagnose things and treat them.

100

00:10:06,330 --> 00:10:14,376

So that's my optimism, but my skepticism is people throwing those terms around very

loosely now.

101

00:10:14,376 --> 00:10:15,763

102

00:10:15,763 --> 00:10:26,361

I think confuses a lot of executives, makes it very difficult for people to understand

what's really truly being delivered now, AI realist.

103

00:10:26,661 --> 00:10:31,024

And I think that that's one of the things that we have to try and do from there.

104

00:10:31,645 --> 00:10:37,319

But if you were given some advice, sorry, please react to what I just said.

105

00:10:37,774 --> 00:10:43,378

Yeah, it sounds very similar as the case with the Agentic AI.

106

00:10:43,378 --> 00:10:50,712

There are also different levels of automation uh from zero automation to completely

autonomous.

107

00:10:50,712 --> 00:11:00,989

People expect to have automation level completely autonomous where you just sit, do

nothing, and the AI agents uh equal employees.

108

00:11:01,110 --> 00:11:04,972

In reality, it's just a system prompt with a SharePoint connection.

109

00:11:06,131 --> 00:11:09,733

it's uh the level of automation is very, very low there.

110

00:11:09,733 --> 00:11:18,008

So it sounds like you described sounds like a very similar situation with this digital

twin where you think twin should actually be twinning something.

111

00:11:18,008 --> 00:11:26,224

uh in uh reality, like agents should have agency, but in reality, neither of them do this.

112

00:11:26,224 --> 00:11:42,194

Yeah, that's, you know, I kind of leaned into the cyborg centaur uh analogy for that in

terms of when I wrote an article recently, saying, uh because there's a lot of fear in

113

00:11:42,335 --> 00:11:56,304

medical and healthcare settings that if we design a lot of autonomous agents, you know,

that are basically taking clinical decisions in their own right,

114

00:11:56,580 --> 00:12:02,044

that without a human in the loop, then that could be incredibly dangerous.

115

00:12:02,044 --> 00:12:12,211

uh Because if you allow that to happen, know, it happening once in a test environment or a

research environment is, well, the model failed there.

116

00:12:12,211 --> 00:12:22,478

But if it happens persistently a thousand times before anybody notices, that could kill

people and many people.

117

00:12:22,478 --> 00:12:26,332

And I think that that's why there is such paranoia in healthcare.

118

00:12:26,332 --> 00:12:31,216

um about how autonomous do we allow agents to be.

119

00:12:31,216 --> 00:12:40,984

If you're processing invoices or processing bills or you're coding a medical record for

billing purposes, sure, I get it.

120

00:12:40,984 --> 00:12:43,886

It's very probabilistic.

121

00:12:43,886 --> 00:12:51,693

So therefore the variance is probably very low so that even if there was an error, it's

not necessarily gonna kill someone.

122

00:12:51,693 --> 00:12:53,084

Might you lose money?

123

00:12:53,084 --> 00:12:55,256

Might you bill too much money and get your claim denied?

124

00:12:55,256 --> 00:12:55,933

Yes.

125

00:12:55,933 --> 00:13:00,433

potentially, but those consequences have always been there with any kind of software

system.

126

00:13:00,433 --> 00:13:02,913

So yeah, I agree with you.

127

00:13:02,913 --> 00:13:17,853

I think that that intersection of agency and where agency should reside is such an

important philosophical concept and security concept for a lot of organizations, but

128

00:13:17,853 --> 00:13:19,711

especially so in medicine.

129

00:13:32,510 --> 00:13:37,297

So if we were to give, because I think you know that

130

00:13:37,297 --> 00:13:46,221

if they're still with us, a large proportion of our audience are uh sales and commercial

and business professionals.

131

00:13:46,221 --> 00:13:52,504

And they've got a day-to-day business of trying to get through what they're doing.

132

00:13:52,504 --> 00:14:04,949

obviously, you know, there's a lot of uh good that these tools can do in terms of

supporting marketing activities, day-to-day research activities and things like that.

133

00:14:04,963 --> 00:14:18,409

What are the kind of practical takeaways that you would recommend to people in those

business functions in terms of using these tools, how to think about adopting them and

134

00:14:18,409 --> 00:14:27,343

rolling them out and understanding how they work, thinking about the challenge of over

automating.

135

00:14:27,343 --> 00:14:35,115

For real day-to-day workers who are trying to sort of make sense, what practical advice

would you give?

136

00:14:35,287 --> 00:14:51,868

for everyone, maybe if you're, working marketing or if you work in sales, if you're a

decision maker who decides to buy something, I would just say spend like a day or two or

137

00:14:51,868 --> 00:14:56,622

three, you can go to the A.I.realist, make like one of the tutorials.

138

00:14:56,622 --> 00:15:02,150

For example, I have like the tutorial with Microsoft Copilot to build an agent or

139

00:15:02,150 --> 00:15:08,954

build a website yourself or go to YouTube, find some kind of other relevant tutorial to do

something.

140

00:15:08,954 --> 00:15:22,132

Just like before buying anything that you think will be an employee in your company as an

agent or before going to someone and selling it because I believe that many salespeople

141

00:15:22,132 --> 00:15:23,913

actually believe what they're saying.

142

00:15:23,913 --> 00:15:27,375

I don't think everyone just walks around and lies on purpose.

143

00:15:27,815 --> 00:15:28,656

I think they just...

144

00:15:28,656 --> 00:15:29,716

145

00:15:29,858 --> 00:15:34,171

read this, they hear other pitches and they repeat this.

146

00:15:34,171 --> 00:15:45,569

So spend two, three days, build an agent yourself, try to use it and you will see very

quickly where the limitations are, like how much autonomy And then think, do you want this

147

00:15:45,569 --> 00:15:48,951

thing to, for example, I don't know, like make your financial decisions?

148

00:15:48,951 --> 00:15:54,665

Would you give this thing your money to spend, invest, buy stuff?

149

00:15:54,665 --> 00:15:59,448

Would you want this thing to make your medical decisions?

150

00:15:59,544 --> 00:16:12,731

and uh would you want this to be an employee of your company in the sense with all the

agency and employee and decision power employee has and then if you really like you know

151

00:16:12,791 --> 00:16:25,988

hand on heart say yes well then uh congratulations you build the product that you probably

can sell and make millions from uh in these three days but most likely you will end up

152

00:16:25,988 --> 00:16:29,710

with the realization that actually this is a tool

153

00:16:29,744 --> 00:16:33,928

So it can help, but it's not an employee.

154

00:16:33,928 --> 00:16:41,946

And I'm never going to give this any decision power because this is just uh not

cognitively there.

155

00:16:42,192 --> 00:16:43,632

That's wonderful.

156

00:16:43,632 --> 00:16:45,032

Thank you for that.

157

00:16:45,032 --> 00:16:47,612

So I'm going to draw things to a close.

158

00:16:47,612 --> 00:16:54,132

Now you've been incredibly gracious with your time, but before we do, one last plug for AI

Realist.

159

00:16:54,592 --> 00:16:55,972

What next?

160

00:16:55,972 --> 00:17:03,656

What have you got on the slate for developing some new ideas and new articles on AI

Realist?

161

00:17:04,513 --> 00:17:06,814

I'm trying to live what I preach.

162

00:17:07,795 --> 00:17:10,437

I always implement myself.

163

00:17:10,678 --> 00:17:15,802

I test the new technology and uh then I write about it.

164

00:17:16,002 --> 00:17:18,865

Now I'm working on this.

165

00:17:18,865 --> 00:17:23,749

I think the next big step in AI for transformers will be continual learning.

166

00:17:23,749 --> 00:17:27,111

Now the models have uh

167

00:17:27,567 --> 00:17:33,287

cutoff date where they do not know past this date and they do not learn anything from you.

168

00:17:33,287 --> 00:17:37,967

Some people think, oh, the more my ragboard chats with me, the better it gets.

169

00:17:37,967 --> 00:17:39,667

No, it doesn't.

170

00:17:39,667 --> 00:17:41,487

It doesn't learn anything from you.

171

00:17:42,187 --> 00:17:50,547

now Google published something called Titan's Memories and they can actually update their

weights and kind of learn from the feedback.

172

00:17:50,547 --> 00:17:55,988

So now I'm trying to figure out how to plug this memory to open weight models.

173

00:17:55,988 --> 00:17:56,390

it's

174

00:17:56,390 --> 00:18:00,141

actually works quite well, at least from what I've built.

175

00:18:00,141 --> 00:18:02,602

So that's I'm going to publish soon.

176

00:18:02,602 --> 00:18:05,293

I already published some, I mean, it's still training.

177

00:18:05,293 --> 00:18:10,884

mean, the training, you see maybe that my eyes are looking that way because I look at the

training right now.

178

00:18:11,224 --> 00:18:17,886

So yeah, so this is what I'm going to talk a bit more about the continual training.

179

00:18:17,886 --> 00:18:22,457

I wrote an article about China winning the AI race.

180

00:18:22,457 --> 00:18:24,658

I still believe that will be the case.

181

00:18:25,000 --> 00:18:31,042

So I'm going to look more at what China is and it's not because I'm biased or anything

like this.

182

00:18:31,042 --> 00:18:33,182

It's objectively.

183

00:18:33,302 --> 00:18:34,763

Their models are efficient.

184

00:18:34,763 --> 00:18:37,293

have uh energy over resources.

185

00:18:37,293 --> 00:18:40,364

So they don't have a challenge to power centers.

186

00:18:40,364 --> 00:18:45,026

They now have lose this dependency on Nvidia GPUs with Huawei.

187

00:18:45,026 --> 00:18:53,680

They benefit from this research hive mind because they provide their open weight in their

research and everyone.

188

00:18:53,680 --> 00:18:58,666

everyone in the world kind of improves China's research, China's models now.

189

00:18:58,667 --> 00:19:00,048

So there.

190

00:19:00,048 --> 00:19:16,769

the US, think that is the challenge, you know, is that here in the US, there is such

paranoia, whether it is, whether it is correct or not, about, you know, the motivation of

191

00:19:16,769 --> 00:19:25,749

the Chinese technologists scientists, and the impact of the CCP, there is absolutely a

paranoia.

192

00:19:25,749 --> 00:19:36,192

uh And even in the current administration, that means that there are limitations in terms

of what people can do or what they can use or how they can interact with some of those

193

00:19:36,192 --> 00:19:37,393

technologies.

194

00:19:37,393 --> 00:19:50,397

that's why I think it's really important that we're able to learn uh from individuals like

yourself who are much more broad minded, much more open minded.

195

00:19:50,461 --> 00:19:52,804

as to what is out there.

196

00:19:52,804 --> 00:20:03,645

And I think that I would encourage you to keep publishing those articles because there

are, I think that there are people who don't, not because of censorship, but just because

197

00:20:03,645 --> 00:20:08,821

there is this implied distrust of a lot of what's going on there.

198

00:20:08,821 --> 00:20:12,815

Whereas my attitude is we should learn from everywhere.

199

00:20:13,861 --> 00:20:19,645

Even US researchers have to use those models because there is no alternative.

200

00:20:19,645 --> 00:20:26,069

Now that Yan Li Kun left Meta, they were the only ones.

201

00:20:26,069 --> 00:20:31,211

Llama was the only uh open-weight usable model.

202

00:20:32,012 --> 00:20:38,396

Top US researchers also use open-weight models from China because there is no alternative.

203

00:20:38,396 --> 00:20:39,997

Europe

204

00:20:39,997 --> 00:20:47,803

like startups, companies, they are completely dependent on open weight models for local

deployments because, there's nothing.

205

00:20:47,963 --> 00:20:48,793

What else?

206

00:20:48,793 --> 00:20:57,609

Mistral, I don't want to comment on Mistral maybe too much because I get a lot of

criticism for it.

207

00:20:57,690 --> 00:21:04,384

honestly, you cannot compare Mistral or even what this GPT OSS with.

208

00:21:04,968 --> 00:21:06,989

Chinese open weight models.

209

00:21:07,229 --> 00:21:09,870

They will be coming, they will be superior.

210

00:21:09,970 --> 00:21:14,252

One pro should reconsider this attitude to open weight models.

211

00:21:16,173 --> 00:21:18,144

mean, West needs open weight models.

212

00:21:18,144 --> 00:21:23,276

Otherwise, imagine, for example, tomorrow China just decides, are done with charity.

213

00:21:23,276 --> 00:21:26,388

ah We are not providing anymore.

214

00:21:26,388 --> 00:21:28,779

So how many products will collapse?

215

00:21:28,779 --> 00:21:31,440

216

00:21:31,904 --> 00:21:32,374

I agree.

217

00:21:32,374 --> 00:21:45,475

And I think that it's important and that's why I would encourage you gentle listeners and

watchers to follow Maria uh and follow the AI Realist and other publications.

218

00:21:45,475 --> 00:21:53,000

Keep your knowledge broad because that way I think it will improve all of us in terms of

trying to understand.

219

00:21:53,000 --> 00:21:55,687

uh Be optimistic.

220

00:21:55,783 --> 00:22:01,274

but also be skeptical about your use of some of these tools and what their capabilities

are.

221

00:22:01,274 --> 00:22:11,702

And just like with all forms of advertising and marketing, try to read between the lines

and understand underneath the promises that are made.

222

00:22:11,702 --> 00:22:20,248

So with that, Maria, I want to thank you very much for being so generous with your time

and for joining us here today.

223

00:22:20,269 --> 00:22:25,781

And I'm going to ask my audience to keep havering.