Speaker: 00:00:12

Hello and welcome to this episode of Haverin About.

00:00:15,885 --> 00:00:22,441

I am so delighted today to introduce you to another one of my favorite reads Maria

Sukareva,

00:00:22,441 --> 00:00:27,683

I wanted to talk with Maria because she goes by a brand of AI Realist.

00:00:27,823 --> 00:00:34,148

And that is really why I enjoy her writing, because she knows so much of the deep

00:00:34,148 --> 00:00:38,969

technical subject matter level, and yet she makes it very accessible.

00:00:38,969 --> 00:00:41,771

She's able to cross the boundaries in your writing.

00:00:41,771 --> 00:00:43,752

So I enjoy that very much, Maria.

00:00:43,752 --> 00:00:45,062

So welcome to you.

00:00:45,062 --> 00:00:47,664

Welcome to our viewers and listeners.

00:00:47,664 --> 00:00:54,168

And maybe Maria, I could start by asking you to just introduce yourself and explain a

little bit about your background.

00:00:54,312 --> 00:00:57,163

Thank you very much for such a nice introduction.

00:00:57,163 --> 00:01:08,194

I'm actually even a bit embarrassed now because of hearing so many nice words towards me

and towards the AI Realist That's really what makes my writing and my work rewarding.

00:01:08,194 --> 00:01:11,405

Yes, so towards the introduction of myself.

00:01:11,405 --> 00:01:15,736

Many years ago in 2003, I started as a linguist.

00:01:15,736 --> 00:01:18,438

like studying linguistics in a university.

00:01:18,438 --> 00:01:25,740

And while I really liked linguistics, my other major was translation and translation is

really not my cup of tea.

00:01:25,740 --> 00:01:40,546

I think it's a very complex job, very difficult job, but also it's a job that a uh bit

puts you in the shadow and is somewhat really not exactly my type of activity that I would

00:01:40,546 --> 00:01:42,569

see myself for years doing.

00:01:42,569 --> 00:01:50,900

I always actually liked computers and I liked math and artificial intelligence even back

then.

00:01:50,900 --> 00:01:54,260

Artificial intelligence existed for many, many, many decades before.

00:01:54,260 --> 00:01:58,660

mean, it was a slightly different type of artificial intelligence back then.

00:01:58,660 --> 00:02:05,260

lot of this ontologies kind of like neuro-symbolic stuff too was back then.

00:02:06,361 --> 00:02:09,132

neuro-linguistic programming and yes.

00:02:09,163 --> 00:02:14,736

Yeah, that's what I frequently get confused when I say I'm like an LP and they're like,

are you doing neuro-linguistic programming?

00:02:14,736 --> 00:02:16,427

No, it's like natural language processing.

00:02:16,427 --> 00:02:17,128

a different thing.

00:02:17,128 --> 00:02:25,893

yeah, so then I was basically, I got a scholarship from German government to study in a

German university in Saarland.

00:02:25,893 --> 00:02:33,445

And I moved to Germany and started studying computational linguistics, which, yep.

00:02:33,445 --> 00:02:34,178

you grow up?

00:02:34,178 --> 00:02:37,341

Where were you born and grew up in the first place?

00:02:37,341 --> 00:02:40,934

I grew up in the polar north of Russia.

00:02:40,934 --> 00:02:44,377

So on the coast of the Arctic Ocean, the city is called Arkhangelsk.

00:02:44,377 --> 00:02:53,203

It was built in 16th century as a place for punishment and suffering from the Tsar.

00:02:53,203 --> 00:03:00,090

So basically he was sending criminals there and people who just didn't like to live there

and suffer.

00:03:00,090 --> 00:03:13,275

And this city eventually grew into an industrial center of the North because there are

many resources, things that actually like an industrial country could profit from.

00:03:13,275 --> 00:03:19,637

So it became a fairly large polar city with many things happening there.

00:03:19,637 --> 00:03:22,490

But of course, it's extreme weather conditions.

00:03:22,490 --> 00:03:25,581

minus 40, minus 30 in winter.

00:03:25,581 --> 00:03:36,542

You have 52 days of vacations a year, paid transportation every two years, and many other

benefits just for living there.

00:03:36,542 --> 00:03:41,931

That's the type of living we are used to.

00:03:41,931 --> 00:03:43,565

visiting your family in St.

00:03:43,565 --> 00:03:44,908

Paul, Minnesota.

00:03:44,939 --> 00:03:47,561

Yes, that's absolutely right.

00:03:47,561 --> 00:03:48,532

I don't think St.

00:03:48,532 --> 00:03:51,163

Paul is even that cold.

00:03:52,595 --> 00:03:53,755

Yeah.

00:03:53,755 --> 00:03:57,598

So basically, then I did the computational linguistics.

00:03:57,598 --> 00:03:59,820

I graduated from this master's program.

00:03:59,820 --> 00:04:01,962

Then I started working in research.

00:04:01,962 --> 00:04:04,844

So basically, I started my PhD in Frankfurt.

00:04:04,844 --> 00:04:08,507

Unfortunately, never finished it, though it was very, very close.

00:04:08,507 --> 00:04:10,339

I had lots of publications, but...

00:04:10,339 --> 00:04:14,244

and life went a different way and I switched to industry.

00:04:14,244 --> 00:04:25,309

During my research career, I was working with machine translation and historical

languages, which many people actually found very confusing back then because, what machine

00:04:25,309 --> 00:04:29,391

translation can you have for middle low German and middle high German?

00:04:29,391 --> 00:04:32,042

I was working back then a lot

00:04:32,188 --> 00:04:35,461

with historical texts and historical texts are what?

00:04:35,461 --> 00:04:38,564

Well, there was only one text basically that existed everywhere.

00:04:38,564 --> 00:04:39,184

It was Bible.

00:04:39,184 --> 00:04:46,000

yeah, I was literally had a huge corpus of Bibles, like huge collection of Bibles.

00:04:46,000 --> 00:04:55,389

And I was doing machine translation from translating like from middle low German to high

German, so on like between different Bible versions.

00:04:55,389 --> 00:04:57,691

However, it might sound irrelevant.

00:04:57,691 --> 00:05:08,942

but in fact machine translation and all these approaches that we had, well, they are

basically were the founding of the theory behind large language models.

00:05:08,942 --> 00:05:10,784

we were doing language modeling.

00:05:10,784 --> 00:05:12,885

mean, language modeling comes from machine translation.

00:05:12,885 --> 00:05:15,608

So we started in N-gram language models.

00:05:15,608 --> 00:05:19,051

Then we moved towards neural language models.

00:05:19,051 --> 00:05:23,536

First, there were recurrent neural language models like LSTMs and so on.

00:05:23,536 --> 00:05:25,397

That's where attention mechanism appeared.

00:05:25,397 --> 00:05:27,670

So, you know, like the one that was famous.

00:05:27,670 --> 00:05:37,227

Then at some point someone, well, someone like the famous paper, attention is all you

need, realized that attention might actually be all you need.

00:05:37,227 --> 00:05:41,481

And what you don't need is the recurrency that was coming from those LSTMs.

00:05:41,481 --> 00:05:47,285

And they just kept the attention and that's what transformer models were.

00:05:47,605 --> 00:05:48,057

And...

00:05:48,057 --> 00:05:51,370

Yeah, and then we had the Transformers.

00:05:51,370 --> 00:05:53,792

They lived for a while.

00:05:53,792 --> 00:06:00,777

By that time I already moved to BMW, I was doing machine translation for an automotive

domain.

00:06:00,777 --> 00:06:07,102

So I was training these models because back then there were not many off-the-shelf

domain-specific machine translators.

00:06:07,102 --> 00:06:11,837

So I had a DGX station with 16 GPUs under my desk.

00:06:11,837 --> 00:06:15,108

Nobody really wanted to use it back then except for me.

00:06:15,108 --> 00:06:18,228

So I was lucky to have it all.

00:06:18,948 --> 00:06:19,228

Yes.

00:06:19,228 --> 00:06:24,979

It was kind of like funny in BMW because I was in the department, like NLP was very small.

00:06:24,979 --> 00:06:31,290

mean, NLP, like now everyone is an AI expert and everyone knows everything about large

language models and back then.

00:06:31,290 --> 00:06:37,921

And when I started in 2018, NLP group was just founded in BMW and there were four people.

00:06:37,921 --> 00:06:40,013

And four people is not enough to have a department.

00:06:40,013 --> 00:06:42,175

So they didn't know where to put us.

00:06:42,175 --> 00:06:48,521

So they put us into blockchain because whatever, guess blockchain was massive back then.

00:06:48,521 --> 00:06:52,375

It's like, it was like 22 people or something doing blockchain.

00:06:52,375 --> 00:06:58,700

And that's why we got the DGX station with Nvidia GPUs because blockchain obviously GPUs.

00:06:58,700 --> 00:07:03,884

So they bought the stations in BMW and the blockchain team actually like didn't do

anything with them.

00:07:03,884 --> 00:07:05,978

Like they didn't, they didn't have a project.

00:07:05,978 --> 00:07:08,129

So I was like, then I'll take them.

00:07:08,129 --> 00:07:15,382

yeah, I had in my car, they still, maybe I still have them somewhere, lying around the

blockchain and the NLP.

00:07:15,382 --> 00:07:20,827

And people, when they would see this card, they were like, oh, you're doing blockchain,

because everyone knew blockchain, nobody knew NLP.

00:07:20,827 --> 00:07:24,429

And I was like, I have no idea about blockchain, like, not a clue.

00:07:24,429 --> 00:07:28,209

And I really didn't know much about like what's happening in blockchain.

00:07:28,209 --> 00:07:31,640

I'm like, just like, don't ask me, but it says blockchain.

00:07:31,640 --> 00:07:34,211

Well, it's the other one, the NLP part.

100

00:07:34,211 --> 00:07:49,131

blockchain was one of those massively hyped technologies as going to be able to solve

everything from commerce to international financial transactions to security across

101

00:07:49,131 --> 00:07:50,422

networks, etc.

102

00:07:50,422 --> 00:07:54,404

And it was the biggest puff of smoke.

103

00:07:54,535 --> 00:07:57,096

I've seen in my career because you're correct.

104

00:07:57,096 --> 00:08:05,121

It was one of those things whereby it was so hot that everybody was talking about it, but

nobody understood what you could actually do with it.

105

00:08:05,121 --> 00:08:11,323

What was the problems it was actually in practice going to solve as opposed to

theoretically going to solve?

106

00:08:11,323 --> 00:08:22,541

So I have a lot of sympathy for you, it's, yeah, it sounds as though it enabled you to get

access at least to the technology that would allow you to start growing your.

107

00:08:22,541 --> 00:08:24,147

activity and work.

108

00:08:36,141 --> 00:08:39,865

Yeah, that was absolutely fun to have these boxes under your table.

109

00:08:39,865 --> 00:08:43,081

It was a luxury because now I don't have this.

110

00:08:43,081 --> 00:08:49,057

now we're all on the cloud and for comparison, I could run training for two weeks there.

111

00:08:49,057 --> 00:08:52,950

And it would train a massive machine translation model there.

112

00:08:52,950 --> 00:08:54,641

And basically it would cost nothing.

113

00:08:54,641 --> 00:08:59,795

mean, maybe some electricity, but for a corporation like BMW, it wouldn't matter.

114

00:08:59,795 --> 00:09:04,299

And obviously like nobody would be counting like how much electricity I burn in the

building.

115

00:09:04,299 --> 00:09:10,631

But now to have the training like this for two weeks would cost thousands on the cloud.

116

00:09:10,631 --> 00:09:20,622

And then I would simply not be able to do because I would actually have to explain why I

burned 10,000 euros on the cloud by training a model.

117

00:09:20,622 --> 00:09:31,282

this was really a luxury back then that is not really easily accessible nowadays because

we are heavily cloud

118

00:09:31,282 --> 00:09:43,290

a lot of people that, you know, when you look at companies like here in the US, there's a

number of those companies, whether it's buying capacity from Google, whether it's Nvidia

119

00:09:43,290 --> 00:09:51,334

selling their highest end processors, or there some of the specialty organizations that

are out there.

120

00:09:51,334 --> 00:09:52,005

I know that

121

00:09:52,005 --> 00:09:58,449

My organization has partnered with a number of them and you're correct, it's incredibly

expensive.

122

00:09:58,449 --> 00:10:01,952

And when you look at the amount of profit that Nvidia is making, interesting.

123

00:10:01,952 --> 00:10:12,469

And when you then look at the interest there is in developing the competitor chips, that

raw processing capability is the engine room that is making a lot of these hardware

124

00:10:12,469 --> 00:10:15,731

manufacturers, these chip manufacturers, we're all going bust.

125

00:10:15,731 --> 00:10:20,053

trying to develop CPUs for servers and computers.

126

00:10:20,053 --> 00:10:27,936

And then suddenly GPUs, it's become a license to print money for a lot of these hardware

organizations.

127

00:10:27,936 --> 00:10:32,549

So yeah, I think it's fascinating the way that that has changed from there.

128

00:10:32,549 --> 00:10:40,774

But you're working at BMW, clearly they didn't have a clue what to do with you or your

work.

129

00:10:40,774 --> 00:10:43,366

So what did you do next from there?

130

00:10:43,366 --> 00:10:47,718

Yeah, so well, in BMW, I was doing machine translation and some chatbots.

131

00:10:47,718 --> 00:10:49,880

uh Back then chatbots were also popular.

132

00:10:49,880 --> 00:10:51,442

And then I moved to Siemens.

133

00:10:51,442 --> 00:10:59,169

And in Siemens, they also decided to start NLP projects, like NLP initiatives.

134

00:10:59,169 --> 00:11:05,225

So I was also started with the usual NLP use cases, so like sentiment analysis.

135

00:11:05,225 --> 00:11:14,398

you know, like usual suspects like classification, but whatever we were doing then back

then with fine tuning transformers like BERTHA and doing them this.

136

00:11:14,398 --> 00:11:21,540

I think chatbot somehow avoided me back then, but it's so yeah.

137

00:11:21,540 --> 00:11:30,322

So, and actually nobody like really cared about me much back then when I started because

it was like just like NLP, some something.

138

00:11:30,322 --> 00:11:41,555

So I was in this dark corner of NLP doing some classification for financial domain, some

sentiment analysis and so on.

139

00:11:41,555 --> 00:11:46,822

And then suddenly, chat GPT happened and everyone decided that they need a chat bot.

140

00:11:46,822 --> 00:11:52,998

And it didn't happen overnight that we had an avalanche of AI experts.

141

00:11:52,998 --> 00:11:56,662

So the first two, three months, were actually like, so who's here?

142

00:11:56,662 --> 00:12:01,085

Actually like heard of this before and there were not many people.

143

00:12:01,238 --> 00:12:12,585

So uh that was interesting to come from this dark corner of NLP into the spotlight of

generative So at some point, yeah, we kind of like the NLP people, they got popular.

144

00:12:12,585 --> 00:12:14,258

145

00:12:14,258 --> 00:12:18,732

And the NLP people who come from that background, were immediate.

146

00:12:18,732 --> 00:12:23,075

mean, the story of language models, it wasn't new for us.

147

00:12:23,075 --> 00:12:24,226

We knew the models before.

148

00:12:24,226 --> 00:12:30,872

knew GPT-3, there was Ulythor, there was GPT-2, and there was a bunch of these generative

approaches before.

149

00:12:30,872 --> 00:12:39,839

And this open domain chatbots, the neural chatbots that kind of behave in a similar manner

as machine translation behaves, because that's basically what they do, right?

150

00:12:39,839 --> 00:12:43,286

They get an input and then they need to generate

151

00:12:43,286 --> 00:12:44,557

some kind of an output.

152

00:12:44,557 --> 00:12:52,257

That's a similar idea behind machine translation, only that it doesn't translate but

continues generating.

153

00:12:55,117 --> 00:12:58,477

We knew this and we knew all the problems with this.

154

00:12:59,077 --> 00:13:05,967

Immediately, we knew how to make this thing to tell you something mean, to spare it to

you, how to hack it.

155

00:13:05,967 --> 00:13:09,359

And so on, like it was like uh quite obvious.

156

00:13:09,359 --> 00:13:13,783

I was very skeptical in the beginning, like that it's going to go somewhere.

157

00:13:13,783 --> 00:13:17,405

And so when people started coming to me, I was like, yeah, that's good.

158

00:13:17,405 --> 00:13:23,566

Let's uh do this, this and this, but this, this, this will not be possible with these

models.

159

00:13:23,566 --> 00:13:28,948

uh But very quickly there was an avalanche of kind of...

160

00:13:28,948 --> 00:13:29,479

161

00:13:29,479 --> 00:13:36,358

AI experts who probably consulted the board first before becoming AI experts for their

education in the field.

162

00:13:38,765 --> 00:13:42,186

There are an enormous number of those people out there.

163

00:13:43,167 --> 00:13:50,711

find, because I write about AI, but I write about it from an applied perspective in terms

of, you know, like it's any other product.

164

00:13:50,711 --> 00:13:58,416

And I've shared with you, I'm one of those enthusiastic skeptics, you know, similar from

you, but you've got a technical background.

165

00:13:58,416 --> 00:14:01,429

I've got a business commercial and healthcare background.

166

00:14:01,429 --> 00:14:03,622

And I find it very funny.

167

00:14:03,622 --> 00:14:11,045

to see all of these people when you look back in their career and how they became instant

AI experts 18 months ago.

168

00:14:23,697 --> 00:14:27,789

And that's the wonderful thing I love about you is this has been your life.

169

00:14:27,789 --> 00:14:35,894

mean, the through path from your linguistics background into where you are today is

extremely clear.

170

00:14:35,894 --> 00:14:41,540

And the thing I love about your background is that through line makes complete

171

00:14:41,540 --> 00:14:49,637

intellectual sense as to why you understand those fundamental underpinnings of what these

transformers are doing.

172

00:14:49,637 --> 00:14:50,367

Because you're right.

173

00:14:50,367 --> 00:14:52,609

mean, they're basically, taking words.

174

00:14:52,609 --> 00:14:58,784

We dump a bunch of words into a context window and they're trying to sort of like look for

patterns, look for associations.

175

00:14:58,784 --> 00:15:02,499

And then the basis of that predicts something that you might want.

176

00:15:02,499 --> 00:15:07,885

And obviously the problem with that is it's, I don't know if you've ever heard this old

177

00:15:07,885 --> 00:15:16,793

joke about if you gave 10,000 monkeys, 10,000 typewriters, give them long enough and you

would eventually get the work of Shakespeare Yeah.

178

00:15:16,793 --> 00:15:17,573

Yeah.

179

00:15:17,573 --> 00:15:25,210

And so I feel like that all we've done is we've applied this massive computing power to

allow that to occur.

180

00:15:25,210 --> 00:15:30,083

So you get this predictive output with lots and lots of loops, you know, in there.

181

00:15:30,083 --> 00:15:35,449

You know, my son calls them Markov loops and because he studies more technically than I

am.

182

00:15:35,449 --> 00:15:36,811

And so I think that that

183

00:15:36,811 --> 00:15:48,538

linguistic background that you've got and that translation background that you've got in

terms of the fundamentals around translating from those various versions of Bibles is an

184

00:15:48,538 --> 00:16:00,525

amazing background compared to some of these IT people who have become AI experts only

because it's the latest flavor of technology and so therefore they've had to study really

185

00:16:00,525 --> 00:16:03,467

hard, very fast to try and keep up.

186

00:16:03,467 --> 00:16:07,087

So yeah, sorry, that was me flattering you even more.

187

00:16:07,087 --> 00:16:09,987

And I apologize for making you blush.

188

00:16:13,024 --> 00:16:13,844

That's okay.

189

00:16:13,844 --> 00:16:23,364

Well, let's pivot a little bit from your background because I know that you're now doing

some very serious strategy work in your work for Siemens.

190

00:16:23,364 --> 00:16:26,724

And we won't touch on that because you're not here to represent Siemens.

191

00:16:26,724 --> 00:16:29,343

We're talking about you and AI Realists.

192

00:16:29,343 --> 00:16:32,046

Oh, I need to tell you something.

193

00:16:33,089 --> 00:16:39,407

Yes, I'm a Siemens employee, but all views expressed in this podcast are solely my own.

194

00:16:39,407 --> 00:16:44,579

Yeah, this is...

195

00:16:44,579 --> 00:16:46,260

to you as the AI realist.

196

00:16:46,260 --> 00:16:48,554

And that's what we're talking about.

197

00:16:48,554 --> 00:16:53,210

that I'm supposed to say every time I'm doing something like this.

198

00:16:53,210 --> 00:16:54,222

Absolutely.

199

00:16:54,222 --> 00:17:00,685

All views of the contributor are her own and personal.

200

00:17:01,606 --> 00:17:02,206

Wonderful.

201

00:17:02,206 --> 00:17:07,019

Well, let's talk a little bit about what you see changing in AI right now.

202

00:17:06,992 --> 00:17:09,975

And before we do that, I'll do a little insert here.

203

00:17:09,975 --> 00:17:14,530

This is Maria's publication, the AI Realist on Substack.

204

00:17:14,530 --> 00:17:19,436

And I can strongly recommend that you go and look that up.

205

00:17:19,436 --> 00:17:24,890

Maria, do you mind if I put your LinkedIn connection as well for anybody who wants to look

up it, because it's public.

206

00:17:24,890 --> 00:17:36,320

And so I'll add in Maria's LinkedIn profile so that if you want to learn a little bit more

about what she's doing either professionally or more importantly, go subscribe to the AI

207

00:17:36,320 --> 00:17:36,901

Realist.

208

00:17:36,901 --> 00:17:38,953

It will be such a great read for you.

209

00:17:38,953 --> 00:17:42,537

Some of Maria's work is very technical, but some of it is...

210

00:17:42,537 --> 00:17:48,772

very insightful about just what's going on and how do you cut through the marketing fluff

and hype.

211

00:17:48,772 --> 00:17:52,535

So with that said, Maria, now I've done a small advert for you.

212

00:17:52,535 --> 00:18:07,606

What is that is changing right now in AI in terms of both at a commercial industrial

level, but also in what you write about, you write a lot about what's happening for the

213

00:18:07,606 --> 00:18:09,839

general utility user.

214

00:18:09,839 --> 00:18:13,843

of AI tools out there in the world.