0:03

Welcome to Practical AI,

0:06

the podcast that makes artificial intelligence

0:08

practical, productive, and accessible to all.

0:10

If you like this show, you

0:12

will love The Change Log. It's

0:14

news on Mondays, deep technical interviews

0:16

on Wednesdays, and on Fridays, an

0:18

awesome talk show for your weekend

0:20

enjoyment. Find us by searching for

0:23

The Change Log wherever you get

0:25

your podcasts. Thanks to our

0:27

partners at fly .io. Launch your

0:29

AI apps in five minutes or

0:31

less. Learn how at fly .io. Welcome

0:45

to another fully connected episode

0:47

of the Practical AI

0:49

podcast. In these episodes

0:51

where it's just Chris and I,

0:53

no guest. We try to keep

0:55

you updated with some of the

0:57

things happening in the AI world.

0:59

Talk through some things that might

1:01

help you level up your machine

1:03

learning and AI game. So excited

1:05

to dig in with you today,

1:07

Chris. I'm joined as always by

1:09

my co -host Chris Benson, who is

1:11

a principal AI research engineer at

1:13

Lockheed Martin. And I'm Daniel Weitnach,

1:16

CEO of Prediction Guard. How

1:18

you doing, Chris? I'm doing good. I'm

1:21

looking forward to our conversation today.

1:23

It's a snowy day in Georgia and

1:25

we can talk a little generative

1:28

AI and talk about you wouldn't want

1:30

to use it it

1:32

was snowing in Georgia, that kind of

1:34

thing. In the theme of coldness

1:36

on today, which is also cold where

1:38

I'm at, talk about the cold side

1:40

of Gen AI or

1:42

actually what we had

1:44

talked about thinking through

1:46

where the bad use

1:48

cases for Gen AI or where

1:51

you shouldn't use Gen AI, five

1:53

or more bad use cases. Yeah.

1:55

And you know, the funny thing about it

1:57

is this is a topic that we have casually

1:59

talked about a whole bunch of times. And

2:01

we had not previously said, let's make it

2:03

an episode. But you know, one of

2:05

the, one of our, I think it may

2:07

be a little bit of a pet

2:10

peeve for not only us, but other people

2:12

I talked to in the AI space

2:14

is there are so many, you know, we're,

2:16

you know, huge hype within Gen AI

2:18

and people just want to use it for

2:20

everything that there could possibly be an

2:22

AI application for. And,

2:24

you know, there's so many places where

2:26

it doesn't necessarily produce the best

2:29

outcome for you. And we talk about

2:31

this casually all the time. So

2:33

glad that we're actually doing this in

2:35

the show today. Yeah, I

2:37

was creating some docs for

2:39

a customer of ours and

2:41

some training materials and I

2:43

have this section just labeled.

2:46

Here be dragons. So

2:49

yeah, there might be some hot takes

2:51

in here. I'm interested to hear what

2:53

your takes are. My first

2:56

one, so number one, bad

2:58

use of Gen AI or

3:00

maybe one that you want to

3:02

avoid at least for now

3:04

is maybe a hot take, but

3:06

I would say from my

3:08

perspective, completely autonomous agents of any

3:11

type are currently, you

3:13

know, well, who knows how

3:15

long this will be the case,

3:17

but currently and for some time,

3:19

generally a source of sadness for

3:22

people when they try to create

3:24

them. So what I mean by

3:26

autonomous agent would be an agent

3:28

or an automation that has no

3:30

human in the loop, just sort

3:32

of is running in the background

3:35

and you kind of hope that

3:37

it Does something for you so

3:39

it could be on the sales

3:41

side right? Oh, I'm gonna have

3:43

an agent do my whole sales

3:45

process for me and I'm just

3:48

gonna kind of sit back and

3:50

work on my product and the

3:52

agent's gonna make all of the

3:54

the sales for me or maybe

3:56

it's you know some sort of

3:58

internal admin process that you're automating

4:01

or You know even all the

4:03

way you know, into manufacturing with

4:05

automation and in planters, you know,

4:07

more industrial case, whatever you're thinking

4:09

of. My first one is completely

4:11

autonomous agents. What's your what's your

4:13

thought, Chris? Not only do I

4:16

think that's right. I'm smiling in

4:18

a big way because I'm going

4:20

to throw in something from the

4:22

side just to support that. Apparently,

4:24

there is a new show on

4:26

Netflix and I just read about

4:29

it last night. Netflix AI

4:31

is tough for me. And

4:33

the show is called Cassandra. And

4:35

it's about this. It's like a home

4:37

assistant robot, you know, with agency

4:39

in terms of doing lots of tasks,

4:41

but it goes apparently I have

4:43

not seen the show yet because I

4:46

just heard about it. But apparently

4:48

it gets very, very dark. And I'm

4:50

just like, when you're talking about

4:52

that just now, you know, in more

4:54

of a real world scenario, obviously, it

4:56

made me think of that. And

4:58

so, yeah, I agree. A completely

5:00

autonomous agent in this day and

5:02

age. with no guardrails around it,

5:04

and you're just saying, go at

5:06

it, generative AI, especially

5:09

if it's dealing with

5:11

anything that has any sort

5:13

of sensitivity or requires

5:15

a little bit of thoughtfulness

5:17

to it. Yeah, not

5:20

going there. Yeah. Well,

5:22

and I think even beyond the

5:24

kind of security privacy related things,

5:26

A lot of times I

5:28

just see people trying to do

5:30

this and it just doesn't really

5:32

work that well. Early days. Early

5:34

days. Yeah, it's early days. So

5:36

like when you have, and

5:39

for those that maybe have

5:41

or haven't listened to previous

5:43

episodes, when we're talking about

5:45

an agent, we mean you

5:47

give a task to some

5:49

sort of system. it

5:51

has the ability then to

5:53

generate queries maybe into other systems

5:55

like APIs or databases or

5:57

data stores or other things to

5:59

accomplish a certain task. And

6:02

it kind of loops over that

6:04

task until it reaches an objective, right?

6:07

And in the autonomous, fully kind of

6:09

autonomous case, you would

6:11

have, you know, just using

6:14

the sales example, because it's easy,

6:16

you know. you want an agent

6:18

to decide how to find

6:20

prospects for you on linkedin and

6:23

then you want to gather

6:25

a dossier about all of those

6:27

prospects and then you want

6:29

to initiate the contact and then

6:31

you want to pull off

6:33

some type of demo or call

6:35

and then you want to

6:37

close the deal and do the

6:39

contract arrangement right and just

6:41

sort of like determine how to

6:44

do every step of that

6:46

process basically relate. replacing a human

6:48

in their agency with the

6:50

autonomous agent. Now, I

6:52

think in that case, we

6:54

could say certain portions

6:56

of that can be very

6:59

interestingly addressed with AI

7:01

functionality. So doing the prospecting,

7:03

generating the dossiers, right? I

7:06

would consider those good use

7:08

cases if they're tied to a

7:11

you know maybe a sales professional

7:13

that's deciding how and when to

7:15

do those things in the imagination

7:17

it would be great to think

7:19

of just kind of letting that

7:21

run in the background and you

7:23

getting sales all the time but. It

7:26

just doesn't really work very well. There's

7:28

a lot of fragility in that type

7:30

of system when there's a lot of

7:32

that determination of objectives and determining how

7:34

to interact with systems and all of

7:37

these things that produces a lot of

7:39

errors, a lot of fragility. It's

7:41

much, much more productive, at

7:43

least currently for you to

7:45

have a tool that can

7:47

help your sales professionals. prospect

7:50

or a tool that can help

7:52

them create these dossiers and that

7:54

sort of thing. And

7:56

certainly tie in AI to that,

7:58

but not kind of this end -to

8:00

-end, completely autonomous

8:02

automation. I totally agree with

8:04

you. And I certainly, by the way,

8:06

just as a clarification from what I

8:08

said earlier, I was not meaning to

8:11

imply agents would typically have a robotic

8:13

body. Should I have confused anybody? There's

8:15

a lot of people exploring that. There

8:17

are. There are. just

8:19

one of the things to note in

8:22

terms of, you know, we're in this,

8:24

the rise of agents right now, it's

8:26

the hottest thing out there. But there

8:28

are, you know, it's interesting, there are

8:30

a lot of guardrail mechanisms that are

8:32

out there. I know in the industry

8:34

I work in and defense, there

8:36

are especially in things like you know, weapon

8:38

systems and stuff like that. The DOD has

8:41

guardrails around such things. So if you're listening

8:43

and aren't familiar with that, but are a

8:45

little bit worried about the world, it's,

8:47

fortunately, there are people thinking

8:49

along these lines. Yeah, and there

8:51

are, I would say, useful

8:53

agents at this. point, just not

8:56

kind of in that fully

8:58

autonomous kind of setting. So AI

9:00

systems that can connect to

9:02

multiple things and maybe are used,

9:04

triggered by a human to

9:06

do certain things, those are the

9:08

most successful that I've seen.

9:10

Absolutely. Number two from

9:12

me, Chris, so we've

9:15

got autonomous agents. Number

9:17

two for me was

9:19

time series forecasting or

9:21

really any sort of

9:23

prediction mechanism. Whether

9:25

that's predicting future

9:27

stock prices or

9:29

reasoning over series

9:31

of data, making

9:34

predictions, there's some

9:36

level of prediction that these

9:38

models can do somewhat well

9:41

in terms of maybe it's

9:43

things like general text classification.

9:47

Is this message spam or not spam

9:49

and you can give some examples

9:51

and you could get some reasonable output

9:53

from a model like that. That's

9:55

why I kind of honed in on

9:57

time series forecasting specifically because at

9:59

least far as I know, and I

10:01

know that there's research in this

10:03

area kind of using transformer models for

10:05

time series forecasting. But when

10:07

I think of Gen AI, I

10:09

think of I'm going to log

10:11

into chat GPT or I'm going

10:13

to use deep seek or one

10:15

of these models and. you know,

10:17

if you paste in a bunch

10:19

of time series data and try

10:21

to create a forecast just with

10:23

the gen AI model and nothing

10:26

else, then I think that's going

10:28

to end again in sadness for

10:30

you. It's not going to work

10:32

so well. Yeah, I think so.

10:34

I actually had that on my

10:36

list too in the form of

10:38

high stakes financial trading. High stakes

10:40

financial trading. Where do you want

10:42

to put your million dollars today

10:44

and see where it goes? Maybe

10:46

explore some of the possibilities

10:48

there, but I don't think I

10:51

would leave it to an

10:53

agent to forecast or make that

10:55

prediction on its own. Yeah,

10:57

I think people have shown basically

10:59

that these models definitely

11:01

don't have the kind

11:03

of world understanding, real world

11:05

grounding to make certain

11:08

reasoning or take certain steps

11:10

in reasoning to make

11:12

reasonable predictions, but also they're

11:14

really bad, generally really

11:17

bad with numbers. You

11:20

may be able to, even with a

11:22

vision model, paste in a graph of

11:24

a time series and say, what month

11:26

was my highest sales if it's a

11:28

graph of sales? A vision

11:31

model could reasonably return that

11:33

value to you. But then if

11:35

you say, well, now model

11:37

out my sales for the next

11:39

four quarters or something like

11:41

that, I think generally that's

11:43

not going to work so well.

11:45

I guess you could argue that. a

11:49

model could generate

11:51

code that might

11:53

use packages, forecasting

11:55

packages to actually

11:57

make a reasonable

11:59

forecast over certain

12:01

data. Then my

12:04

general question then would be, well,

12:06

that might be useful to generate your

12:08

code to do it, but really

12:10

it's not gen AI that's doing that.

12:12

It's the stats models in Python

12:14

or - That's right. profit

12:18

from from meta and that sort

12:20

of thing. Yeah, I mean in

12:22

just in case that confuses anyone,

12:24

you know, there's the generative AI

12:26

portion, you know, which can, you

12:28

know, is trained on a general

12:30

data set. And then there's these

12:32

models that it might be generating

12:34

code to access, which are designed

12:36

specifically for that function. So those

12:38

are two different things. Yeah, the

12:40

code that ends up being executed

12:42

is not having anything to do

12:44

with JANAI, basically. Yeah,

12:46

and maybe it would be worth highlighting in

12:48

each of these cases that we talk about,

12:50

Chris, some interesting tooling for

12:52

some of these things. You know,

12:54

in the autonomous agents case, certainly

12:56

workflows and automations can be created

12:59

and executed. You know, we had

13:01

Prefect on the show, which is

13:03

a workflow orchestrator that can be

13:05

monitored and handle retries and all

13:07

of that. That's a great thing

13:09

if you're looking at kind of

13:11

workflows and orchestration. Time

13:13

series forecasting. GoTo

13:16

has usually been Facebook

13:18

or Meta's profit package, which

13:21

makes certain things pretty easy,

13:23

but there's also many choices for

13:25

that as well. So take

13:27

a look through those things if

13:29

you're interested in the non -GenAI

13:31

side. Well

13:43

friends, AI is transforming how we

13:45

do business, but we need AI

13:47

solutions that are not only ambitious,

13:49

but practical and adaptable to. That's

13:51

where Domo's AI and data products

13:53

platform comes into play. It's built

13:55

for the challenges of today's AI

13:57

landscape. With Domo, you and your

13:59

team can channel AI and data

14:01

into innovative uses that deliver measurable

14:03

impact. While many companies focus on

14:06

near applications or single model solutions,

14:08

Domo's all -in -one platform is more

14:10

robust with trustworthy AI results

14:13

without having to overhaul your entire

14:15

data infrastructure, secure

14:17

AI agents that connect, prepare

14:19

and automate your workflows, helping

14:21

you and your team to gain

14:23

insights, receive alerts and act with ease

14:25

through guided apps tailored to your

14:27

role and the flexibility to choose which

14:29

AI models you want to use. So

14:32

DOMO goes beyond productivity. It's designed

14:34

to transform your processes, helping you

14:36

make smarter and faster decisions that

14:38

drive real growth. And it's all powered

14:40

by DOMO's trust, flexibility, and

14:42

years of expertise in data

14:44

and AI innovation. And

14:46

of course, the best companies rely on DOMO

14:48

to make smarter decisions. See how DOMO can

14:50

unlock your data's full potential. Learn

14:52

more at

14:55

ai .domo.com.

14:57

That's ai .domo.com.

15:04

All right, Chris, on to

15:06

number three. My third

15:08

one was do not

15:10

use Gen AI to do

15:12

complete code rewrites or

15:15

the complete development of your

15:17

applications, your software

15:19

applications. Thoughts? Oh,

15:21

I've tried that just playing

15:23

around. And I definitely

15:25

don't think that that's ready for

15:27

prime time, despite the fact that,

15:29

you know, as we sit here

15:32

and say this, there have been

15:34

quite a few CEO luminaries out

15:36

there who have been advocating that

15:38

over the last year or so. And

15:41

when I sit down

15:43

and try to do that,

15:45

I get varying results. And

15:48

it depends largely on how

15:50

mainstream a language is, for instance, on

15:53

how good it is. But

15:55

I haven't gotten anything that I

15:57

would say is a production

15:59

grade program fully functional through nothing

16:01

but generative AI, just toy

16:03

programs. Yes, without interaction. Right.

16:06

Right. Yeah. I

16:08

know this is advancing quickly,

16:10

so who knows how... this conversation

16:12

will be in a few

16:14

months, but I think we've been

16:16

talking about this for for

16:18

some time now and we've seen

16:20

things like Devin and cursor

16:22

and these sorts of things come

16:24

out which are pretty amazing

16:26

and Do a lot of really

16:28

interesting things but often don't

16:30

kind of provide that full, like

16:33

I'm going to prompt and

16:35

get a software application out of

16:37

it. There is, there's more

16:39

to it than that. So I

16:41

think sometimes people are maybe

16:43

a bit disillusioned and, you know,

16:45

a better way to think

16:48

about this or there are

16:50

amazing kind of agents and

16:52

toolings come out like the

16:54

Devin cursor, all hands, windsurf,

16:56

et cetera, that can provide

16:58

a huge acceleration. in

17:00

your code development, I

17:02

think, if you treat them

17:04

like code assistants and,

17:07

you know, maybe even junior

17:09

developers that you are

17:11

pairing with, right? So

17:13

it's not so much that I'm just

17:15

now a not complete non developer,

17:17

right? I have no technical skills and

17:19

I just say I want this

17:21

application. And it is generated for

17:24

me. That's really what I'm meaning when

17:26

I say kind of complete app development.

17:28

So, Gen AI, from my

17:30

perspective, is not capable of

17:32

that right now, or you should

17:34

not rely on it for

17:36

that right now. There may be

17:38

interesting demos and cases where

17:40

some form of that is shown,

17:42

but for the most part,

17:44

I think thinking of the technology

17:46

integrated into your code and

17:48

programming as an assistant. even

17:50

a highly functioning agent that

17:53

you compare with is a

17:55

good model. Maybe

17:58

it's a specialization of the

18:00

autonomous agent thing that I

18:02

mentioned before. I think you're

18:04

making really good points in

18:06

that. You can't just toss

18:08

it over the wall and

18:10

just say, here's an instruction,

18:13

do it all, and generate

18:15

kind of a complex set of

18:17

programs and stuff. I have

18:20

done tasking, small things very successfully,

18:22

but the scope of what

18:24

they were addressing was constrained. And

18:27

I think we are there for

18:29

things like that and doing small

18:31

bits, it's not uncommon for me

18:33

to generate. Many years

18:35

ago, I would write a VBA

18:37

code, Visual Basic for Applications for

18:39

Microsoft stuff. I don't much anymore,

18:42

and so now I can do

18:44

something like that if I happen

18:46

to be working for something in

18:48

office to do something, put something

18:50

together at work. But when

18:52

I'm actually coding up a large

18:55

project, it's very helpful to have different

18:57

tools on this, but I've not

18:59

found one yet that I was able

19:01

to discussfully do a significant coding

19:03

effort by itself, just tossing it over

19:05

the wall. So I agree with

19:07

you completely. It will be interesting to

19:09

see where we are a year

19:11

from now, two years from now. Yeah,

19:14

well, definitely, I would encourage people

19:16

to check out things like Windsurf and

19:18

Devon and all hands and cursor

19:20

and all of these things. Super cool.

19:22

Try them out. But

19:24

don't expect that if you're not

19:26

a programmer or have at

19:28

least some minimal level of skill

19:30

that you're going to create

19:32

a huge application or project with

19:34

all of its intricacies and

19:36

have that work and scale well.

19:38

Fair enough. All right, Chris, what

19:41

are we on? Number four for

19:43

me on the list of

19:45

don't do this with Gen

19:47

AI or bad Gen AI

19:49

use cases for me is

19:52

anything extremely high throughput load

19:54

latency. So of

19:56

course, small models and

19:58

very high throughput advances

20:00

have taken place with

20:02

Gen AI models, but

20:04

still, you know, if

20:06

you're doing quality assessment

20:08

of products coming

20:10

off of a actual scaled

20:12

up manufacturing line where

20:14

you have to do maybe

20:17

the assessment of each

20:19

of those products in a

20:21

fraction of a second. Really,

20:24

you don't want to

20:26

be reasoning over that data

20:28

with a Gen AI

20:30

model and take 10 seconds

20:32

to generate your quality

20:34

assessment for the product. It's

20:36

just not feasible. I

20:38

would agree with that. And I actually have a

20:40

subset that I'll throw in on that that

20:42

I think kind of fits in there, which would

20:44

be kind of like real time. applications

20:47

with critical outcomes. a

20:50

great way to phrase it. I think

20:52

that that's an area that you

20:54

may have generative AI as a

20:56

component in that mix, but you're

20:58

going to have to have some

21:00

guard grills around it, and you're

21:02

going to have to have some

21:05

specialized models to keep things on

21:07

track because in a real -time app

21:09

where things matter on the tail

21:11

end, great to

21:13

use, but you don't want to rely entirely on

21:15

that when it goes off the rails. You need some

21:17

way to catch it that doesn't take any time. I

21:20

think you make a couple great

21:22

points. Part of it is around the

21:24

latency, which I highlighted. These

21:26

models just don't operate fast enough,

21:28

and they don't operate in the

21:30

types of environments necessarily that you

21:32

need them to operate in for

21:34

these type of maybe edge use

21:36

cases as well in many cases. But

21:40

also, these models perform

21:42

or they do what they are

21:45

supposed to do most of the time,

21:47

right? But still, if

21:49

you train a computer vision

21:51

model, for example, to

21:53

do that manufacturing task, that

21:56

could run on CPU extremely

21:58

high throughput and have a

22:00

much higher accuracy than any

22:02

generalized vision model out there,

22:04

even that would need a

22:06

GPU to run, right? I

22:08

agree with that. Yeah, so

22:10

it's just not, what is

22:13

that, the separation between those

22:15

two cases is still just

22:17

really, really high in terms

22:19

of those kind of use

22:21

cases merging. Now,

22:23

I do think that in

22:25

a manufacturing scenario, right, there's

22:28

a great, or any of

22:30

these sort of other cases that you

22:32

might think of high throughput, critical

22:35

type of scenarios, JNAI is very useful.

22:37

maybe just not for that

22:39

high throughput load latency piece,

22:42

but certainly for staff at

22:44

the manufacturing facility that want

22:46

to look at and analyze

22:48

the data coming off of

22:50

the quality assessment system and

22:52

ask questions about, hey, I

22:55

see this alert, pull this

22:57

data for me to help me

22:59

understand what's going on. Or

23:01

are there any of these types of

23:03

events that have happened in the

23:05

past X time? And that query level

23:07

side via natural language can be

23:09

very powerful. for example, and there's many

23:11

other things that you could do

23:13

in those scenarios. I'll

23:16

extend this just a little bit.

23:19

As you know, my personal

23:21

passion is in autonomous platforms,

23:23

especially at massive scale, swarming,

23:25

things like that. When

23:28

you talk about that, one

23:30

of the areas where I think JNAI

23:32

does play is exactly the equivalent

23:34

of what you just said on the

23:36

manufacturing. That's having a human in

23:38

the loop or on the loop that's

23:40

able to interact and so you're

23:42

using Gen AI to actually be able

23:44

to enhance the communication between the

23:46

human who is in control or on

23:48

the loop and able to step

23:50

in and not but but not so

23:52

much in the other areas especially

23:54

considering that when you have lots of

23:56

vehicles and this could apply for

23:58

lots of different use cases both in

24:00

the commercial space and the military

24:02

space where you have a lot of

24:04

different platforms or vehicles in communication,

24:06

which requires high throughput. But yeah, I

24:08

think that the only space there

24:11

that is a big one is in

24:13

those interactions with the humans that

24:15

are involved in that for safety. Yeah,

24:17

for sure. Well, I have

24:19

one more, Chris. A last interesting

24:21

bad use case for Gen AI.

24:24

The one on my list

24:27

was anything outside of

24:29

the major languages of

24:31

the world. So anything

24:33

with any sort of linguistic

24:35

diversity or cultural diversity,

24:37

essentially the models of the

24:39

modern gen AI era

24:41

maybe work well in the

24:43

kind of top five

24:45

to 10 languages of the

24:47

world. But there's 7000

24:49

spoken languages in the world,

24:51

which means they basically

24:53

don't work for any of

24:55

the languages of the

24:57

world except for a couple.

25:00

Moreover, the

25:02

kind of cultural context

25:05

of the models is

25:07

driven by mostly what

25:09

has been gathered either

25:11

from the internet or

25:13

by Western tech companies, maybe

25:16

Chinese tech companies.

25:19

But there's certainly a

25:21

bias against certain cultural

25:24

context and languages and

25:26

you know, even if you

25:28

think about vision or

25:30

video models, I'm sure the

25:32

same is true, right?

25:34

Because just certain things aren't

25:36

represented there. So the

25:38

reality is that it would

25:40

be great if you

25:42

could, you know, land anywhere

25:44

in the world and

25:46

change your chat GPT or

25:48

whatever to help you

25:50

interact in X country in

25:52

Africa or Y country

25:54

in Asia and have that

25:56

work really well with

25:58

whatever languages you might encounter.

26:01

But I would say generally that's

26:03

not the case as of now.

26:05

I think so. I know

26:07

you haven't mentioned it yourself, but

26:09

longtime listeners who have been with

26:12

us for years will know that

26:14

you used to be in that

26:16

space in a former professional life

26:18

and know quite a bit about

26:20

this topic that you've just brought

26:22

up. I

26:24

agree. I don't

26:27

think that's changed substantially over

26:29

the last few years. Yeah,

26:31

and even simple things that don't have

26:33

a lot to do with, I mean,

26:36

it has to do with Gen AI,

26:38

but also has to do with the

26:40

tooling around it, right? In terms of

26:42

even other scripts in particular Arabic, you

26:44

know, for example, which of

26:46

course is a major language

26:48

of the world, which to

26:50

some degrees, you know, models

26:52

can do reasonably well at

26:54

at least some models. The

26:57

tooling around the Gen AI

26:59

ecosystem, right? Like, oh, I

27:01

want to download this chat.

27:03

SDK or this UI that

27:05

I can plug in a

27:07

custom model to is likely

27:09

not going to support kind

27:11

of right to left. Potentially,

27:13

there's going to be some

27:15

issues with the script and

27:18

other things. It's just another

27:20

highlight of this disparity that

27:22

exists. It exists and I

27:24

think is worth highlighting because mostly what

27:26

we're talking about here is language models. Really

27:29

language models that support a

27:31

very small amount of the languages

27:33

on on the planet. Yeah,

27:35

yeah But that's what I had

27:38

Chris any thoughts after going

27:40

through through the list of bad

27:42

I think you know there

27:44

I do have a few thoughts

27:46

there. I think one of

27:48

the things that I've noticed there

27:50

is that There are

27:52

kind of high risk and

27:54

high and like where you

27:56

have significant outcomes that can

27:58

affect people in a major

28:00

way and whether it be

28:02

financial or manufacturing or Yeah,

28:04

my industry would defense or

28:06

whatever, you know, you don't

28:08

want to put a general a

28:11

general generative AI model in charge

28:13

of doing things for which there

28:15

are no guard grills. I think

28:17

that that is a thing that

28:19

I have noticed across a lot,

28:21

and I could throw out a

28:23

couple of other areas where I

28:25

think that applies, like things like

28:27

high -stakes legal advice. Do

28:29

you have a great tooling within

28:31

things like chat, GBT, and the other

28:33

big language models for legal advice? Yeah,

28:36

but would you really want to you

28:39

know, literally put your life

28:41

savings at risk with things like

28:43

that, maybe not today at

28:45

least. You see a lot of

28:47

this, you see a lot

28:49

of AI pervading medical diagnosis. And

28:52

once again, I think there's a

28:54

very good use for those, but

28:56

probably not by itself, you know,

28:58

in isolation. So any of these

29:01

areas where you have a substantial

29:03

risk in the outcome in terms

29:05

of good and bad, You

29:07

probably want to have guardrails around it

29:09

across many, many different industries. And that's, I

29:11

think that's my takeaway. And, you

29:13

know, I think that things are continuing to

29:15

improve at a really, really rapid pace. And

29:17

we've said things and had, you know, two

29:19

months later, had the world change out from

29:21

under us. And that may happen again here

29:23

with some of these. But yeah, it's, we're

29:26

on the learning curve with these things and

29:28

they're getting better, but they're not all the

29:30

way there yet. Yeah. I think that's a

29:32

great way to summarize, Chris. Thanks

29:34

for chatting through the things with

29:36

me and we'll look forward to carrying

29:38

on the conversation very soon with

29:40

you. Sounds good. All

29:46

right, that is our

29:48

show for this week.

29:50

If you haven't checked

29:52

out our changelog newsletter,

29:55

head to changelog.com slash

29:57

news. There you'll find

29:59

29 reasons, yes, 29

30:01

reasons why you should

30:03

subscribe. I'll tell you reason

30:05

17, you might actually start

30:07

looking forward to Mondays. Sounds like

30:09

somebody's a case of the Mundus.

30:12

28 more reasons are waiting for

30:14

you at .com slash news.

30:16

Thanks again to our partners

30:18

at fly .io to breakmaster for

30:20

the beats to you for listening.

30:23

That is all for now, but we'll talk to you again

30:25

next time.