0:10

All right, welcome back to Daytoday with Greg Michaelelsson. I'm here with my good friend Thomas Dinsmore. Uh we spent

0:16

uh a long time working together at data robot and he was our uh tell us tell us what you did. you were a competitive

0:22

like you were the guy who came up with the battle cards at at data robot to help us talk about competitors and all

0:28

that all that sort of stuff. Yeah. Yeah. That's what I did. Um Jeremy gave

0:34

me the title uh director of competitor annihilation but I guess I failed because all the

0:40

competitors are still in business. So no annihilation was accomplished. That

0:46

sounds very Jeremy-esque to be honest. Uh but uh tell tell us kind of how did

0:53

you um walk us through where you went after data robot and what you're up to these days.

0:59

I left data robot I think in December of 2021. I wanted to join Domino data uh is

1:07

funny. Um the reason I joined Domino I mean partly look I am now 72 years old.

1:15

I worked at 70 and in the last 10 years of my career, virtually every

1:22

opportunity I had came from networking, right? And so I knew people at Domino and and maybe I could have gone

1:29

somewhere else, but I'd rather work with people I know. So that's fair. So yeah, I knew some

1:34

people at Domino, but also funny thing um

1:40

and I don't know if this is still true, but at the time in 2020, I looked at some opportunities where data robot had

1:47

been competing with Domino and Domino won every one of them, right? Domino has a a different orientation

1:57

than data robot did. And again, my data may be a couple years out of date at

2:02

this point so far as data robots concerned, but and Domino is a company

2:08

that has enter literally enterprisecale

2:13

um data science customers, people with 500 data scientists um you know ve very

2:19

large scale customers and they have a platform that scales up to it. But also

2:25

where data robot focused on sort of efficiency, building our model faster and that's important. Um, Domino's

2:34

their stock and trade is governance, right? In other words, the the way Domino works is it's literally capturing

2:41

everything that all of your work on the on the platform so you can reproduce it,

2:47

right? And that has a couple different benefits. One is if you're starting a new project and someone says, "Hey, you

2:53

know, that looks an awful lot like this other project we did, you know, you go pull up the other project and you've

2:58

got, you know, all your assets, your project assets that are in front of you." Uh, but the other thing is if the

3:05

auditors show up, you can show them exactly and and there seven or eight

3:11

different aspects of kinds of assets that you generate when you're a data

3:17

scientist. So I mean you have your code, you have your data, an actual snapshot of the data as well as how you pulled

3:24

the data, you have the actual models you created, you have the various production

3:29

deployments of those models, etc., etc. In other words, there there are the different types of assets

3:36

and um there's a lot of vendors that claim to offer reproducibility because

3:42

they can do some of those things. But again, Domino's and the way Domino's is able to compete against companies like

3:48

Amazon or Google is is two things. One is that they're able to deliver a better story for

3:55

reproducibility, but there are also still an awful lot of enterprise scale data scientists that can't limit their

4:02

work to a single platform, right? So, and Domino has a hybrid platform. So you can

4:09

literally be running you can be building on Amazon and on Google and on your

4:15

onrem platform and all of that is integrated into a single instance with

4:21

visibility across all your projects and so that's that's that's how they appeal. Now again I'm not here to sell Domino. I

4:30

um but that's why I went there from data robot. I mean data robot was a fun company to work for. I love Jeremy. I

4:37

loved the company. um great people, people like yourself, um you know, all

4:42

the all the data scientists that data robot attracted uh just exceptional

4:48

people. And one reason I left is that, you know, it struck me at the best

4:53

people were leaving. And I think in any company, um I I left before the big

4:59

scandal broke with the Did you really? Yeah. With the the stock selling and all that

5:04

stuff. Yeah. And I bought my stock at 15, right? And

5:10

yes, so I I now have a sitting in my safe deposit box, you know, physical

5:16

certificates of data robot shares, as far as I'm concerned, are worth nothing, which I Well, they are at this point. Yeah.

5:22

Uh but um yeah and I and I got data robot will you know they they seem to be

5:29

finding a new niche in the marketplace but they have so much senior equity you

5:35

know there's a billion dollars of senior equity in the company unless the company manages to sell out for more than a

5:42

billion bucks all that common all those common shares that uh that that we

5:47

earned uh are worth nothing but you know you never know But yeah, I left before

5:53

all that hit the van in December 2021 and it was in I think June of 2022.

6:00

Was Jeremy still CEO when you left? Uh, no. Um, yeah.

6:05

So, I think I lasted until October. Jeremy left in April of 2021.

6:14

Yeah, that's Yeah, it was 2021. April 2021 you left. Yeah. Uh and then the company kept on

6:21

going. I left in December of 2021. At the time I worked I I had moved over

6:28

to the product team uh which was great great to work for. I mean, again, I have no I have no beef with data robot as a

6:36

company and as an employer, but I just wanted to go in new directions. And and again, it seemed like a lot of the best

6:43

people were were starting to leave, and that's to me always a good sign that uh it's time to time to, you know, pack

6:51

your bags. But then the uh the the article appeared in was it the Intercept? One of those uh

6:58

information, I think. Oh, yeah. the information they published the article in June of 2022.

7:05

Oh, what's his name? The CEO left in September.

7:10

Dan. Yeah. He was able to hang on for a couple months, you know, of scandal.

7:17

Meanwhile, the the you know, a lot of the members of the board, the entire audit committee quit. That that tells

7:23

you. Yeah. Yeah. Yeah. Yeah. But you know by all indications there's still a few

7:28

people that were there when I was there that are that are still there and you know there's fair fair bit of turnover most

7:35

of the executives have gone. Yeah. But so I mean I think the company may have hit bottom and it's on the resurge and

7:42

and the last Gartner uh Gartner scored him as a leader which they never did in in during Jeremy's tenure. Of course

7:49

Jeremy hated Gartner and they hated Jeremy. Maybe there's a Well, I mean, I don't have much love for

7:55

Gardner either. It's all pay to play that Gartner stuff. You know, you write him a big enough

8:01

check, they'll put you in whatever quadrant that you want. I think it's a little bit more subtle than that. It's not like a junk tooth

8:08

bags full of cash. But there is a built-in conservative bias because most

8:15

of Gartner's clients are like big IT departments, right? And so there's a

8:21

tendency to to favor the more conservative companies. That's why IBM

8:27

and SAS always score well on those things, right? Um, you know, but don't

8:33

get me started on Gartner. Um, yeah. Yeah. I'll just sound like an old man

8:40

ranting. Uh, yeah. It's a necessary evil. Maybe it's it's one of those things. And that

8:45

that's I mean I love Jeremy but um

8:51

it's like you it's Gartner is just part of the game you have to play as a CEO in

8:56

the you can't just thumb your nose at them um because they have influence and

9:04

yeah so um all you can do really is suck it up do your best and and try to spin

9:09

it or not participate at all right which is I I think the smarter move because

9:15

it's easier to message around we don't think you know we our product doesn't

9:20

fall into Gartner's predefined category than it is to say well we did our best but we're still a niche player you know

9:27

yes that's fair how do you think all this uh

9:33

this LLM uh genai stuff has impacted the autoML market that data robot used to

9:39

play in or tries to play in yeah well data robot's singing this song

9:44

of AI agents and um whenever I hear the word agents I go to sleep. I mean I I I

9:52

just this is old story. I've told this a million times but um back in 1996

10:00

Gartner was predicting that data warehouse agents would transform business. Right? And what were they

10:06

talking about? The data warehouse agents they were talking about were what we now know as stored procedures and table

10:11

functions. Right? So is there something in AI agents? Yeah, but it's like a it's

10:18

like a buzzword for something for a fairly specific technology that people

10:23

will understand, you know, and and get some value out of. Um I was speaking to

10:31

another CEO of a startup who I can't name, but he's in in in a category that

10:36

you might say is similar to data robot. and his organization is using AI agents

10:43

for um feature engineering and feature engineering and feature generation. I

10:48

would characterize it as proactive feature generation because the way they've set it up is

10:54

your agents are going out there looking for data and and finding ways to transform it before you build your

11:00

model. Um and so in that and that's a and you know does it work? I don't know, right? Can they sell it? I don't know,

11:08

right? But it sounds good. Um, so yeah, but and I get mixed signals from I do

11:14

try to keep in touch with people I know on the data science community and I get mixed signals. I mean, one one woman who

11:20

is a data robot who's now at a consulting firm says pretty much all her work is in predictive analytics and

11:27

another one who is a data robot says all her team's work is generative AI. So I

11:33

think it depends on where you are and what comp kind of company you're with. Generative AI is opening up new

11:40

use cases for data scientists that didn't exist before. So, you know, big

11:46

brand companies that do a lot of marketing and advertising are using Gen AI to create backgrounds and images and

11:52

text and content and all that content. That's the kind of stuff, you know, the the the marketing departments

11:58

were like the last people to use data science uh the way the way we define it

12:04

a data robot. But still there's there's there's still money in predictive applications. So

12:11

you know arguably there'll be a split. Now I I can think of and again I won't name names but um a data scientist that

12:18

you know from data robot top-notch fellow now all he does is is generative

12:23

AI and and he's really good at it. But the thing is he was he he like when

12:31

Google first published Transformers in what what was that 2017 2018 he was all

12:36

over it right so I mean there are people who have been all over the the the incremental developments that led up to

12:44

the the explosion of generative AI all along. you know, people we used to call them text miners, right? Or uh you know,

12:52

they were computer vision specialists, right? Well, and so they they like understood

12:58

that technology and they're the people that have the most credibility with generative AI today.

13:04

Do you use these models much in your day-to-day? You getting Gemini? I use Gemini. And why do I use

13:10

Gemini? It's not because I checked some leaderboard and they have the better. I use Gemini because it's embedded with

13:16

with Google Doc, Google Slides, which I use all the time. Um, I use Google

13:22

search. And for me, an important thing in in any AI assistant is its ability to

13:30

uh come up with sources, right, in order to validate what it's saying, right?

13:35

because I find with Google Gemini all the time it'll say some and I'll say whoops sorry it'll it'll say and uh

13:43

uh you know and I'll say no that's not true show me your sources and it goes

13:48

out and it comes back apologizes I apologize I you know you're right I couldn't you know you have to push it

13:54

but yeah so I use research when I write and um I use like I recently finished

14:02

like a 50,000word book and on what?

14:07

Uh it's about my son. Um Oh, I see. And um it's basically his his biography.

14:15

Uh and I took the complete text, copied it into

14:20

the the Gemini window and just said, "Please fact check this." In other

14:26

words, there there are things you can do with with AI assistance that you you never could do. I never uh one of one of

14:33

the one of the things I don't do is I don't ever let Gemini write anything for me. I mean, to me, that's just

14:43

I'm not going to say unethical. I I just think it's lame and and you know uh but

14:48

um I you know you can do things like you can

14:55

like in this text one of the things I wanted to do is I wanted to anonymize

15:00

um I didn't want to have use people's real names. So I asked Gemini to read

15:05

the entire 50,000word text and give me a list of all personal all the named

15:11

people in there. Right. it it can do that right so then I go back in and I

15:17

you know but just basic factchecking um you know I don't I don't you know again

15:23

there are things that are it's so much easier to use Gemini to search for information than it is to use Google

15:30

search yeah it's remarkable you just dump text in there and have it do whatever it is that you want it to do

15:37

I'm I'm being a retired person I've spent a lot of time in the garden and uh

15:42

I worked with Gemini to create my plan for the the coming season. Uh, you know,

15:48

it knows all the plants. It knows what their water requirements are, their sun requirements are, how fast they grow,

15:55

how tall they grow, all that good stuff. It's a great tool. I like it. It It is surprising how often it

16:00

apologizes for stuff, right? One of the things that So, my in-laws own a uh

16:06

bakery and we help them with soup. So they they like to serve soup and so

16:11

we'll ask for a recipe for, you know, uh corn chowder for for 50, you know, and it'll come up

16:18

with a recipe and we'll go, you know, that seems like an awful lot of chicken broth for that corn chowder recipe. Are

16:24

you sure that's the right amount? And it doesn't matter what amount. It's like, oh, you're right. This, you know, this was wrong kind of thing. It's very

16:31

uh eager to please. Yeah. Uh and so if you tell it it's wrong, it's going to completely agree with you.

16:37

Um so have you noticed this? have and and we all understand that

16:42

these things hallucinate and and thing is so do traditional models, right? If

16:48

you build a churn model or you know a risk model and you put in absurd inputs,

16:53

it'll still give you a score. It's just the score is no good, right? You have to build into your application the ability

17:00

to to validate and filter the inputs that are going in. The difficulty with

17:06

generative AI is that it's much more difficult to validate and filter the inputs that are going in and the outputs

17:12

that are coming out because it it just by the nature of it. So yeah, I mean

17:17

it's it's hard to please and and and the thing is you have to wonder it's like if

17:22

if if it's giving if it's so easy to give me wrong information, how can I trust it? The answer is no, you can't

17:29

trust it. All right? Don't we're trust but verify. Now I have the feeling that we are in um

17:38

let's see so I was reading a book the other day or an article about the way physics has proceeded like how

17:46

innovation in physics has happened over the last say 40 or 50 years and the the point of this article was that we're we

17:52

were in a blind alley that we went down and we made some progress in terms of like understanding the nature of the

17:58

universe and so on but somewhere along the way a bad choice has been made and it's led led to a dead end and that's

18:04

why there hasn't been any real innovation in the last you know few decades in in physics. I have the same

18:10

feeling that that's what's happened in the the Gen AI space that we've we've

18:15

got these models that are really good chat bots and that can do some pretty impressive tricks but that there's no

18:21

way to go from here to like AGI to like real real artificial intelligence

18:27

without like backtracking and changing some of those bad decisions along the way. Do do you think that that's true or

18:34

do you think that we're on the right track or you know how do you see how do you see this particular innovation in

18:40

the broader context of like trying to get to our machine overlords?

18:46

Anytime anyone mentions AGI, I go to sleep. It's just like with agents, right?

18:52

I I have a very simple attitude towards generative AI. I like it. I use it. It's

18:58

that simple. I mean I you know and and to me generative AI is just remarkable. I mean it's just remarkable that we can

19:06

do this. Um yeah I I know there are people out there Gary Marcus but my

19:12

opinion is Gary Marcus was the guy that watched the Wright brothers fly their plane and went out and and put all his

19:19

money in Zeppelin's right. I mean all all these people that say

19:25

that that are that are screaming that we'll never get to AGI. It's like, who cares, right? It's like, we're not going

19:31

to get, you know, are we going to get to autonomous cars? Well, my car that I

19:38

bought two years ago has all kinds of cool features that it didn't h that my

19:44

the previous car didn't have. All of which are related to. In other words, there's safety. It detects things and it

19:51

it you know in other words it there are things that we've developed on the way to autonomous driving that are actually

19:59

benefiting cars today. All right. Are we ever going to get to AGI? Well, who cares? It's like it's a little bit like

20:05

my father used to say, nobody cares about getting to the moon, but we at least we got Teflon, right?

20:12

Yeah. I mean, a lot of a lot of times the the the the main value of Uh first

20:19

of all, everybody has to have a goal, right? And if you know, and if for some people the goal is ai, okay, that's fine.

20:26

Whatever works for you, right? Um but a lot of the benefits don't come from

20:32

meeting the goal. They the benefits come from what you what you learned and developed along the way. Uh yeah, I mean

20:40

I I could probably go at length about that, but I won't.

20:47

No, I like the Teflon example. Certainly uh uh you know a recipe generator like

20:54

like Chad GBT is a really interesting uh sort of like byproduct of the pursuit of

21:00

building uh uh real artificial intelligence. That's fascinating. But if you know if chat GPT generated recipes

21:07

it would be like I don't know liver cake you know or

21:12

things that there are a lot of sort of constraints and boundaries on the way we think that we actually value right like

21:20

you know liver is actually good for you I don't know I'm not going to eat it right

21:27

um yeah so um

21:32

I again I'm just incredibly impressed with with generative AI as it is. Yeah,

21:38

I know there's I I keep hearing about risks, you know, all that that u the

21:43

controversy Grock created with undressing people and all that sort of stuff which took them I think a week to

21:48

fix. Um if they I mean again I I

21:54

I didn't hear about that. What what happened with Grock? Well, Grock introduced a feature where you

22:00

could upload a photo and Grock would return a photo of that same photo

22:06

undressed. Wow. Like a package featured doc.

22:12

Wow. That seems like a profoundly bad idea. Oh, that could be abused, right? you know. Um, and

22:20

I've seen people make claims that there this was like hugely awful and I can I

22:27

Yeah, I don't think I would want to have a picture of myself published undressed.

22:32

Um, but that said, I also know that Grock pretty quickly fixed that, right?

22:38

So now if you upload a picture first of all if you say you know you know

22:45

undress this person Grock Grock will say null right uh but you know they they

22:51

will they'll take a picture of K starr the prime minister and put him in a

22:57

bikini right in other words they it will it will do certain things and you know

23:04

is is it so in general is there potential abuse of course there is I mean I remember

23:12

couple years ago I mean chatting with some people on LinkedIn arguing that that you know as with every o other

23:18

innovation uh the most profitable uses for

23:24

generative AI will be will be you know war fraud and porn right and

23:33

I think only fans is one of successful startups in Silicon Valley I mean you know, so and they're not even using Gen

23:40

AI, but u or are they? I mean, I I I it may sound cynical, but

23:46

the history of science is largely the history of war and and uh and it's like

23:52

why did we invent the integral calculus to aim cannon? Right. Right.

23:59

Dar Darwin expedition on the Beagle was sent by the Royal Navy to find coing

24:04

stations. All right. He was just he was just cargo. But

24:10

do you think that we have I don't I don't think we've really figured out what our rights are yet or what rights

24:16

should be in terms of like is it is it is it just a bad idea for this like

24:22

undressing application that you talked about or is does that violate some sort of like human right that people should

24:29

or do have? Uh I'm I'm not a lawyer. I mean, I think most people would agree that an

24:36

application that takes a picture of you fully clothed and then recreates it of

24:43

you totally naked is is problematic, right? Is it? But I mean, we've had we've had the

24:49

capability of like putting your head on somebody else's body for a long time. Yeah. Like that that exists.

24:56

Yeah. And if you publish that, I think you are potentially at legal risk, not

25:01

necessarily of a crime or but of a but of a civil tort. Right? In other words,

25:07

if you take a picture of somebody you work with and, you know,

25:13

put that person's head on a naked body and then publish it and you damage their

25:18

reputation, that that's that's a lawsuit that I would happily invest in. Uh but again,

25:27

I'm not a lawyer, right? Um it doesn't have to be just like uh you

25:32

know, porn type stuff. It could be putting words in somebody's mouth, right? Uh or you know, something like that.

25:38

Those are all things that are very easy to do these days. There was a case in New Hampshire where

25:45

um it was right around the time of the the

25:50

Democratic primary in 2024,

25:56

right? So it was while President Biden was still considered a candidate

26:03

and um somebody put out a a video of

26:10

it showed Biden say talking and saying don't bother to vote in the primary

26:17

because I'm running and and there's no need for you to vote. In other words, it was and supposedly was like a voter

26:23

suppression thing and and that actually went to trial and the the person was

26:28

acquitted. Uh and the reason they were acquitted was it was it was it it was a

26:34

technical aspect of it had to do with the fact that that um

26:41

I think technically they published it after the voting had closed. Uh it wasn't again I don't remember the

26:48

details of it but again I know the person was acquitted. It was a case of faking somebody's

26:55

faking President Biden saying something. Uh and any anyone could do that. I mean,

27:02

there's some disincentives to do that among, you know, in other words, if you're if you're doing a political ad,

27:09

um I I I would think that people can pretty quickly smoke out

27:16

those fakes, right? In other words, if if you're if you're running for office and someone

27:23

publishes a video of you saying that, you know, something really outrageous,

27:28

well, the moment you come out and say, "I didn't say that." Right. That's, you know, in another the moment you deny it,

27:34

right? Fact is that there probably going to be some people out there that won't believe your denial. Right. Right.

27:40

But it's usually generally possible to figure out if something has been faked.

27:46

um you know you there for example I don't know about with images I know that

27:53

with text there's very good AI detectors um are there though

28:01

now the I've seen actors like uh actors and writers and comedians and so on

28:07

being up in arms about what AI can do and how it's sort of like uh well I mean

28:14

has the potential to kind of replace that whole industry in a in a pretty significant way. Um, do you see that

28:20

happening? Well, let's take I don't know George Clooney just to I mean George Clooney

28:26

has a property right in his image is is in other words if somebody uses AI to

28:32

clone George Clooney and then make a movie George firing George Clooney, George

28:38

Clooney can sue them, you know, into into oblivion. Um

28:44

that's certainly true for famous people. I a lot of the the concern I think comes

28:52

from more like bloggers, right? Just sort of people that aren't famous.

28:59

I mean there there is a I mean you you you automatically have a copyright in

29:05

anything you create. You don't in other words anything you write you you have a a copyright. So, and but one of the so

29:14

let's say chat GPT let's say open AI trains its models on

29:20

the text you wrote question is does open AI now owe you a

29:25

royalty um it's not been settled because open AI

29:31

will claim that under under copyright law there they

29:37

there is a a doctrine of fair use Right. If you It's like if you if you write a

29:44

book, I can quote of it and say uh you know or or I can write a satire about

29:51

it. All right. And I don't need your or I can I I can learn from it and generate more content of my own based on

29:57

it. Well, there are limits to the fair use doctrine, but there is there is a under

30:04

US law. It doesn't apply in in it doesn't apply in in England, I don't

30:10

think, and it doesn't apply in Europe, but under US law, the fair the fair and the UK and Europe have a different

30:17

version of a fairness doctrine. And they don't call it that. And again, I'm not a lawyer, but in the US, the

30:24

fairness doctrine has limits, but it's it's all to be limited. And I believe

30:29

there is a court case between the New York Times and Open AI. And how are the

30:34

courts going to come down? I don't know. I mean, they they could they could rule in favor of Open AI, they could rule in

30:40

favor of the New York Times. I I I think in some cases um there there are folks that argue that

30:47

Open AI didn't just violate their copyright, but they violated their terms of service. Um because like

30:55

no web scraping, this is hypothetical. Uh, I'm not saying this happened, but the the the argument

31:01

is that one person at OpenAI might have a subscription to the New York Times

31:06

and, you know, using that he now has access to the New York Times. He then trains a model. I'm not saying that

31:12

happened, right? Uh, I I think I've heard people argue that, but that would be not a violation of copyright. It

31:19

would be a violation of terms of service. I don't know. I mean, these things will get up to the Supreme Court

31:25

and they'll be settled, but they won't be settled for years. Um, and yeah,

31:31

and in the meantime, millions of pages of copy are getting generated by Chad GPT every day, and the internet is

31:38

becoming more and more dead than it ever has been. But

31:43

many of the concerns, for example, I have a blog, right? I never charge

31:48

anything for people to read it or use it. If chat GPT trains its model on my

31:55

blog, I don't I haven't suffered any damages,

32:00

right? I haven't and you could even make an argument that particularly to the

32:06

extent that um AI assistants link back to sources

32:12

that having a an AI model train on your content might actually help you, right?

32:19

In other words, you you may you may get more people linking back to it. But that that's the thing. In other words, the if

32:26

people are making a living off their content and their issues, again, I'm not

32:32

absolving any of the AI vendors of responsibility here. I'm just saying

32:39

it's a matter that's going to be settled in the courts and it's hard to predict how it will go.

32:46

But I mean the courts only have jurisdiction over the the the places where those courts are, right? So the US

32:52

Supreme Court only governs the US, right? And you know the EU only governs the EU and

32:57

and there's different different law governing content in other countries. Uh and yeah, so

33:05

yeah, but I mean at the end of the day, this is kind of a a worldwide phenomenon. Yeah.

33:11

So, it's not super clear how an individual country can can uh you know

33:16

govern that sort of stuff when it's when it's worldwide. Well, if I'm open AI uh

33:23

which I'm not, but if I were uh as with any multinational company, I need to

33:29

have an understanding of the laws of every country where I operate. Uh and

33:35

that means that uh I don't know if it's still true but for a while chat GPT wasn't available in

33:41

Europe at all. I mean is it now? I don't know. Right. But the point being is that that the the capabilities you deliver

33:50

and the level it affects both your your training, right? What what copy is fair

33:57

game for training? It might be different in one country than it is in in another country. But also what can you show what

34:05

can you deliver that has to be country specific and that's true it's like it doesn't you

34:11

know it's true in software if you're if you're marketing software you have to

34:17

market one way in Brazil and a different way in Chile and it's not yeah but I mean if you have if you have

34:23

200 million users or however many million users that that open AI has you

34:28

know there the people will demand what the people want you know what I So maybe open AAI has to like worst case

34:35

scenario the courts say hey you can't do that they relocate the company to Bermuda and then they've still got you

34:41

know they're still worth a hundred hundred trillion dollars or whatever the number is and they're still developing

34:46

you know this this software that everyone on the planet uses. Well I I I'm just going to make a

34:52

hypothetical. Let's suppose a country says that you cannot use um

34:59

generative AI to create pictures of animals

35:06

um let's just say right you know

35:11

because animals are sacred in this country or a particular kind of animal right so if you're a multinational

35:20

AI vendor you're either going to have to create and send create an exception. In

35:26

other words, a different instance of your software and your service that you serve up to users in that country or if

35:34

it's big enough, you might just apply that same rule globally. I mean one one of the um

35:43

one of the uh phenomena in the US textbook market like for school

35:49

textbooks is that California and Texas uh have centralized purchasing of of

35:56

textbooks. So if you're a textbook publisher, whatever California and Texas

36:01

want, you're going to you're going to roll out nationwide. Doesn't matter that in Massachusetts every town picks its

36:07

own textbook, right? you're not going to tailor your your your your publication to what they want in Lexington,

36:14

Massachusetts. You're going to tailor it to the standards of of so if the customers are big enough, then yeah,

36:20

they can have the influence, which to be I mean, not to diss anyone else, but the

36:26

United States market is still the biggest market for any of these uh um AI

36:32

suppliers. And so the the US market rules but we are seeing right now for

36:38

example um with in in the UK digital

36:43

services act right the the the UK is trying to shut down some US websites

36:51

that and and again what they can do is they can seize any assets that those

36:56

companies have in the UK but these are companies that have no assets in the UK right and so that that attempt to

37:04

regulate will likely be feudal. Um because again the UK can say we don't like what

37:09

you publish on your websites and they can they can try to block that in the UK

37:15

but they can't go to the US and say you have to shut down that website. Mhm. Yeah. Exactly. It feels to me like

37:22

the regulators in this case are are playing from a position of weakness where they the people want what the

37:27

people want and there's not really a lot that they can do about it. Yeah, I I suppose that's right. H

37:35

wild. Anyway, well, what's what's uh what's the next uh the next big thing for for Mr.

37:42

Dinsmore? What do you got on the horizon? Well, I'm working on uh last year I

37:47

started working on a history of AutoML and I got a certain point to it and I

37:53

don't know, I I kind of lost interest. not uh not that it's it's an uninteresting

38:00

topic, but to be honest, I'm more interested in the current and the future. So, I've been doing some

38:05

revisions to that recently. I hope to publish something soon. Um,

38:10

and I haven't decided whether I might just abandon the history and just do a current state of AutoML. One of the

38:17

challenges that I'm grappling with is I mean, AutoML really isn't a category

38:23

anymore. uh it's it's more of a feature and uh do I want to write about AutoML a

38:30

feature or I want to write more broadly about I I am I'm principally interested in I

38:37

mean some people call it data science some people call it predictive analytics

38:42

right I'm interested in that process of building useful models for enterprises

38:48

right you know credit score churn models whatever right that that whole process

38:53

process and I'm interested in platforms that make that more efficient, more

38:58

effective, etc., etc. And you call that is it automated machine learning? Is it

39:05

automated data science? Is it automated predictive analytics? Is it something else? I haven't decided. I will decide.

39:11

I have to before I can publish anything, but I want to sort of summarize the current state of commercial and

39:18

open-source offerings. Uh, that's probably one post right there. Uh

39:25

does does it seem to you that uh that things like claude code or chat GBT or

39:30

or Zerve or whatever are sort of making automated machine learning or automated

39:36

anything really sort of obsolete. So if I can if I can generate code at will that does exactly what I want do I need

39:43

an automated process? Well I think the answer is yes and no. Right? In other words, you can go to chat GPT and say,

39:50

"Write a write write the Python code for a churn model." And it'll write you the pi Python code for a churn model. Is

39:57

that code any good? Well, how do you know? You have to test it, right? And and if you think about what data robot

40:04

did or H2O does, and they run experiments. So, can you say to Chad GBT, um, you know, give me the best

40:13

churn model given my data? Um, it's possible that in the same sense

40:20

that a million monkeys typing at random will eventually write poetry. I mean, chat GPT might be able to produce that.

40:28

But it's more likely that what you need is some application that uh that sort of

40:35

manages the process and prompts a a large language model for the code

40:41

for the experiments. Right? In other words, write an experiment, run it, evaluate

40:49

it, and then write a new prompt, right? That is that is more precise that that

40:57

yields, in other words, you so and but is it going to be AI that manages that

41:03

process or is it going to be some application that people I don't know. I mean,

41:08

um I I don't think that exists today, but it's I know what people are talking

41:14

about when they talk about AI Asians. Um yeah. Uh

41:20

but I don't I don't want to give away too much of of what I'm thinking about writing, but I'm I'm a little skeptical

41:27

of the notion that chat GPT is just going to replace all your data scientists. Um,

41:33

partly because for the same reason Chad GPT isn't going to replace all your doctors,

41:38

which is that people go to doctors because they trust doctors. And

41:44

whereas it makes perfect sense for the doctors to use chat GPT or some

41:51

variation of chat GPT for the same problem exists in in in

41:57

business, right? In other words, the the chief executive doesn't know that this model is any good, right? She relies on

42:05

the chief data scientist who tells her this is a good model. Well, how does the

42:10

chief data scientist know? Well, it's not because the chief data scientist thinks chat GPT writes good models,

42:16

right? It's because the chief data scientist either knows how to prompt Chat GPT or because there's some process

42:24

of testing and measurement and so forth. Does that answer your question? I mean,

42:29

yes and no. Yeah, it does. I mean, it seems to me that there's an awful lot of stuff that's being created that is mainly

42:36

useless. Um, you know, like like a feature store or like weights and

42:43

biases. Like I feel like there's a lot of products out there that are sort of solutions in search of a problem and the

42:49

the stuff that's going to win is going to be the stuff that people actually use. And you know it seems to me that

42:55

that most machine learning is you know like anytime anybody talks about machine learning they go oh you know like

43:01

customer turnurn it's like an example of a problem that that is either too simplistic or not

43:08

relevant for most organizations out there. I would disagree about the two vendors you named because well feature

43:15

store well tecton is now data bricks I think but um it actually makes some

43:20

sense if you're a data scientist that rather than redefining a feature every

43:25

time you write a model to define it once lock it down uh and then

43:32

you reuse it that that makes that makes good sense right I mean it's yeah it seems to I mean certainly

43:37

the idea has caught on but is Anybody actually using it? I I don't know.

43:42

Well, Tecton, you know, the data bricks gave him a bit of money for uh is anyone

43:49

using it? I I think the answer is yes. Tecton added the footprint of customers.

43:56

Weights and biases. Uh I mean, nobody from data robot should throw throw uh

44:02

throw shade at Weights and Biases. They are they sold out. They they they they

44:08

date a robot, didn't um yeah, no question about that. And weights and bi Weights and bias

44:14

perfectly exemplifies the the market today in the sense that those guys were all over um generative AI before

44:22

generative AI was a thing, right? In other words, they uh open AI was a customer, weights and biases. Uh and so

44:30

they they kind of understood that long before you know it's like when when when

44:36

chap GPT launched you can actually see this if you go back and in the way back

44:41

machine you look at different vendor websites you know companies that have never heard of generative AI suddenly

44:48

are are like we're the generative AI vendor right uh but that wasn't true weights and biases understood that that

44:54

stuff a long time ago um Yeah,

45:00

they're doing fine now. Uh who bought up now? Uh uh is uh Core Logic Core. Um

45:08

yeah, right. Yeah. Yeah, those guys are all doing fine. Yeah, they they're doing fine, but I

45:13

mean how much money somebody makes and the actual real usefulness of their product is not always perfectly

45:18

correlated. I don't know. Yeah, I don't know. Yeah. I mean, supposedly we're entering

45:24

a a trough of disillusionment for generative AI. I don't know. Could be.

45:29

I'm not I'm not the least bit disillusioned. Are you? Not either. All right. But yeah, I mean,

45:34

the key question is, can anyone make money from it? And I don't mean, you know, I don't mean Nvidia. Of course,

45:41

Nvidia is making money. You know, you know, when when there's a gold rush, sell shovels. Uh but um the question

45:49

that people are trying to figure out in business like if you're a bank, you know, does Gen AI help you make money?

45:56

And and I my gut feel is yes, but it takes time to figure that out. Um I

46:01

published a study where they're saying, "Oh, this Genai stuff is nonsense. Nobody's making any money. I mean, the

46:08

ROI on these generative a projects is nothing." And you know, I worked with IT

46:14

organizations a long time. I never never ever knew any IT organization to measure

46:20

ROI on anything, right? You know, if you had to measure ROI on data warehouses,

46:25

they never would have built any of them, right? So, I I I'm I'm I'm

46:32

a little skeptical about the people that are saying, "Oh, there's no money in this stuff." Um because yeah, I mean

46:39

that the there's there's some obvious opportunities to make money with generative AI and and you know, we're

46:45

just beginning to discover them anyway. Yeah. Like what's what's the ROI on Google Docs, right? How do you even

46:50

start to calculate it? Yeah. But you know, can you operate without a word processor? No, you absolutely

46:56

can't. Yeah. I mean, I know what I mean, Google's just I mean, they're harvesting

47:02

everything I write, you know? Uh so the I mean what's the ROI on them putting

47:08

out a free uh uh free uh word processor? Uh well

47:14

pretty high. I mean they completely murdered Microsoft Office. Yeah. Anyway, um I've got to run in a

47:23

minute or two. Uh yeah. No, it was great talking to you. I really appreciate you jumping on. It's always always nice to

47:30

That's top of mind for you. No, no. Just uh just great to catch up and get your

47:35

thoughts on the things that are happening these days. Cool. And I didn't have a chance to ask you what your dome is, so maybe another

47:42

session. Yeah, you'll have to jump on. I'd love to hear your your competitive thoughts on the AI coding environments that are

47:49

out there for data science. All right, sir. Well, cool. All right, Thomas. Thanks. Really

47:55

appreciate it. Great seeing you. All right. Bye. Bye.