0:00

Welcome to the prompt engineering podcast,

0:03

where we teach you the art of writing effective

0:05

prompts for AI systems like chat,

0:07

GPT, mid journey, Dolly,

0:10

and more. Here's your host, Greg

0:12

Schwartz.

0:15

Welcome to a joint episode

0:17

of the

0:17

Prompt Engineering podcast and the How to

0:20

Talk AI podcast.

0:21

We've got some awesome guests, so go ahead

0:23

and introduce

0:24

yourselves, guys.

0:25

Yeah, I am Aaron, er, the co-founder

0:27

and c e o of Vital.

0:30

And I'm Felix Brand at

0:31

Vice President of Data Science.

0:34

And they have a terrific product

0:36

that they have a launch. Today we're gonna hear all

0:38

about it. I think it's something that would resonate

0:40

with everyone and anyone that's

0:43

been to the doctor and had questions about

0:45

what, what was being told

0:46

to them. Yes. I

0:47

already tested it after watching your talk. Cool.

0:50

Really? Yeah. So I have sleep apnea. Yeah.

0:52

I put in a long diagnosis

0:54

with a bunch of stuff that I'm like, okay, I think I know

0:57

what that is. Yeah. I don't know what the hell that is. Yeah.

0:59

And it was like, sleep apnea, obstructive.

1:02

Yeah.

1:02

And two other things. Yeah. Oh.

1:04

Fantastic. Okay. That's great. I,

1:06

I think like you said what person

1:08

hasn't seen a whole long

1:11

list of doctor's notes or even been

1:14

in a situation where you're maybe an

1:16

inpatient in the hospital and then the

1:19

doctor on rounds is coming by and telling you something

1:21

in a million miles a minute because he's got 20 other people to

1:23

see. But it's probably important

1:25

because it affects your own health and being and like,

1:28

you're Probably already out of it anyway, because you're

1:30

in the hospital. What a terrific way

1:32

to,

1:32

Provide something. Doctor's notes are really almost

1:35

like a foreign language. Yeah. As

1:37

I said in my talk, doctors don't say

1:40

nosebleed, they say epistasis. They

1:42

don't say, hey, your mom has had a stroke.

1:45

They say, oh, she's had a cerebral infarction.

1:48

They use all of these abbreviations. It's almost

1:50

impossible to... Understand.

1:52

And so we use the large language

1:55

model, As the core of

1:57

what we call our Doctor to Patient Translator,

2:00

and it's at vital. io slash

2:02

translate. It's free to the public, available

2:04

worldwide, literally as of today.

2:06

You're just catching me at a good time. And

2:10

we're happy to tell you, a bit about the

2:12

prompts, the classifiers, the free parts,

2:14

and all the things that we offer. To

2:17

make that possible, technically.

2:19

Yeah, that would be great. I would love to I would love to

2:21

delve into some of the the technical aspects.

2:23

Maybe this is a better question for Felix. Could you tell us a little

2:25

bit about how the model was going to be trained and

2:27

what data was used to be able to produce these

2:29

great

2:29

completions? Sure.

2:31

We've tried a number of different prompts because there are actually

2:33

a lot of different types of doctor's notes. And with

2:35

the public facing stuff, we know that we're going to get the

2:37

whole gamut from imaging all the way to discharging

2:45

stuff. We People, when they get their paper discharge instructions

2:48

upwards of 90% of them chuck them straight in

2:50

the bin as soon as they leave the hospital. And the

2:53

literature people understand their care, and

2:55

understand the follow up instructions

2:57

the doctors are giving them, their post care situation is way

3:00

lower. So we've looked

3:02

at different prompts for different situations

3:04

and then built a pre

3:06

model classifier, a pre LLM classifier,

3:08

also using a language model, a

3:12

small one deciding which of our various prompts should you have

3:15

to write some nodes and then we have

3:17

a whole bunch of post parsing, it comes out, we

3:20

take sections out of translation,

3:22

we plug those sections of the website,

3:25

maybe when you saw it, you could see that you

3:27

get like a very brief summary. And then

3:29

also a much more sort of technical breakdown.

3:32

Yes. So we're getting the LLM to pull out a lot of information

3:35

about what's in your justice note, but we want

3:37

to show you in like a digestible

3:39

summary first. Yeah. I think an important

3:41

piece of context is a lot of these doctors notes,

3:43

they're 10 or 15 pages long,

3:46

and they have 80% boilerplate.

3:49

Yeah. They have a, Hey, don't smoke. Or

3:52

I don't. Hey, here's COVID education. Okay. You're two years out

3:54

of date. And they put a lot of filler

3:56

in there. And this is actually just a fraction

3:59

of our primary business. Our primary business is

4:01

patient experience offer. It guides you through

4:04

an ER visit, or if you have to stay

4:06

overnight in the hospital, it explains your lab results,

4:08

how long you're going to wait. And then your

4:11

notes. Yeah. And because

4:13

we have experience with a million patients

4:16

a year using it, we know the structure of notes.

4:19

from all over the country. And so

4:21

we can pre parse, and instead

4:23

of a 10 or 15 page, we can

4:25

get it down to actually we only need to pass 3

4:27

or 4 pages into the LLM. That's

4:30

an important business and engineering consideration,

4:33

because cost and speed.

4:36

Also context window. If

4:38

you're doing, especially if you're using few shot

4:40

training with an LLM, which is a good

4:43

idea so that you know what output

4:45

you want to get. You'll blow through

4:48

your prompt, your future

4:50

shot, your data, and then your output, it has to fit into

4:52

a 4k window or a 16k window.

4:56

And so you need to do a few things to give

4:58

yourself as much profit as

5:00

possible. That makes

5:02

complete sense, but having the almost

5:05

sub prompts acting like little sub

5:07

agents themselves trained to

5:10

say just get rid of all the boilerplate

5:12

stuff that's not unique to

5:15

that patient's differential diagnosis.

5:17

Exactly. So deciding which part you're

5:19

going to do... Are more or less with

5:21

your own code or your own classifiers,

5:24

and then how much to send, especially if you're using a, like

5:26

a commercial l m. And we've used

5:28

both. Felix's got, Lama up and running

5:31

and too Yeah. Med

5:33

Palm lm, which is

5:35

medical specific, obviously. The

5:37

open ai, we can't actually use open

5:39

AI directly. You have to use it through Azure because

5:42

you. You need this to be hit. We're

5:45

in a regulated industry. Open AI will

5:48

not sign all of those things. You actually have

5:50

to like Work your way through Corporate

5:52

Microsoft. Yep, they'll determine

5:55

whether you're a worthwhile person or not, and

5:57

whether they're willing to take the risk, and

6:01

so if you put all of it, you can, with a sophisticated

6:03

prompt, put it all through WebMGT. You

6:06

can say, classify this. Is this a discharge

6:08

report? Is this a physical therapy

6:10

report? Or is this a hostile input?

6:13

By the way, you should always protect against hostile input.

6:15

Is this a non English input? Is this something

6:18

else entirely? So you want, and then...

6:21

In your prompt, you can say based on the classification,

6:23

then do this. But if you do

6:25

all that, your prompt starts to get very complicated

6:27

and very big. You can use that to prototype,

6:30

but when you go into production, this

6:32

is also very slow, it gets very expensive, you

6:34

run a classifier that's much simpler and

6:37

much quicker on top of it, and

6:39

then you don't have the expense, your prompt's shorter.

6:41

And then you can say, if it's this, go to this prompt.

6:43

If it's that, go to that prompt. You can

6:45

also templatize prompts. So if you

6:47

say, I want the output in Spanish,

6:50

you can put a variable in your prompt. So the

6:52

prompts, don't think of them as static strings.

6:55

Think of them as a programming language

6:57

that is frankly pseudocode, yeah?

7:00

One of the things that, this is a bit like medical

7:02

specific, but the part that's very important

7:04

to patients is the plan and assessment,

7:07

what the

7:07

doctor says you're supposed to do. Here's the

7:09

problem. In some hospitals it's called plan and assessment. In other

7:11

hospitals it's called assessment. In other hospitals it's called

7:15

plan. In other hospitals it's got like an abbreviation.

7:17

And with classic programming if I say

7:19

match panda and I give it pandas

7:22

with a plural, it's no

7:23

match. Or you got a space in your column

7:25

header. Exactly. But with an

7:27

L M, I can just be like, it's gonna be called

7:29

this, or probably this. It's got stuff

7:31

that kind of looks like this and like

7:33

it's good enough that if I explain it to you guys, you'd be

7:35

like, oh, okay. I know what you're looking for. That's

7:38

the power of LLMs is

7:40

you can give them. Vague pseudocode.

7:42

Yeah, and to me, that's mind blowing.

7:45

This guy actually knows a map of how that's

7:47

passed. So

7:48

real quick before we get into that, just for the

7:51

audience, part of what I do on my

7:53

podcast is like, What are all these

7:55

technical terms? Content window, number one.

7:57

It's literally how much stuff you're putting into the prompt,

7:59

but also how much it's filling out, and

8:02

if you do too much, it forgets the stuff outside

8:04

the prompt window. Sorry, the context window.

8:07

And so you have to be careful how long everything

8:09

is. That's what they're talking about when you're saying,

8:11

if I can pull pieces of the prompt out and only run

8:13

them separately, it's way better.

8:44

It's a key reason to innovate in your own models, because

8:46

for a long time you've been working with 4K

8:49

context window, and if you're doing

8:51

this few shot in context learning, as

8:53

Aaron says, you just run through it.

8:54

Yeah. And also, I'm

8:56

the CEO as well as, maybe

8:59

you can tell I have a bit of an engineering background, not as

9:01

good as this guy. I don't have the British

9:03

accent, which is, that's true. And

9:05

also, that adds

9:06

20 IQ points, right? Yes.

9:10

But as the CEO, I have to think through the economics,

9:12

right? If you were using GPT 4 and

9:14

you give it the 16K 32K window,

9:18

the maximum one, it's going to cost you,

9:20

if you fully fill that thing, it's going to cost

9:22

you about 48 cents per, translation

9:25

or transformation, right? Yeah. We have

9:27

a million patients on our platform. They have

9:29

about five nodes each. You do

9:31

the math on that and you're spending

9:33

5, 000 a day.

9:35

Yeah. If that's what you do. You don't need

9:37

to. You use smaller context windows,

9:40

or you use 3. 5 Turbo, or you run

9:42

Llama. Yeah. Or you use one

9:44

LLM to pre parse for a different LLM.

9:47

You can do, those are the tricks that like, practically

9:50

speaking, this is an immature

9:52

industry because you have to hand do

9:55

All of that.

9:56

And what's really interesting is, some of these problems

9:58

are really exciting and new. As Aaron says,

10:00

you're trying to pull out something that's

10:02

very undefined in free text document.

10:05

Okay. So that's you need some modern stuff to

10:07

do that. But some of these problems are pretty traditional.

10:09

Classifying a document and you've got, plenty

10:12

of examples. You don't need to go and use

10:14

your OpenAI LLM to do this

10:16

classification problem. We've been doing this for a long time. And

10:18

you can do them a lot cheaper.

10:20

Yeah, it's slow and expensive to use OpenAI,

10:22

or Google, or

10:25

NetApp for basic classifications. But it's

10:27

great for prototyping. So the key

10:29

insight, is work out the piece that you really

10:31

need the expensive tech for, and ensure

10:34

that you boil down the problem only to that, using

10:36

other pieces of technology

10:37

upstream. Yeah. So how do you handle,

10:40

Like the, if you have all these

10:42

prompts essentially acting as agents, and

10:44

you have to have this sequence occur. In

10:47

a specific order, how do you asynchronously

10:49

is there a specific layer that's doing the

10:51

handoff? Are they doing the, are they doing

10:54

a turnover at rounds? In between

10:57

synchronization, let

10:58

me get a little technical. So we use an

11:00

event. sourced architecture. So this

11:02

is outside of AI, which basically

11:04

means that we handle streaming data quite well. So

11:06

we have data that's streaming from

11:09

over 100 hospitals now, more or less

11:11

real time. It comes out of Cerner, Epic, whatever

11:13

the electronic medical record system is.

11:15

So a doctor writes a new note, finishes it, it hits

11:18

our system and

11:20

goes on to the parsed,

11:23

classified, cut up into little bits, and

11:26

then divvied out to the yeah,

11:28

you need to synchronize it so you have queues of

11:30

work. Those queues can back up. We just

11:33

launched this.

11:34

Unfortunately at this point, we had some

11:37

audio challenges. So the video will

11:39

continue. But going

11:41

forward, we're only able to use audio

11:43

from a much lower quality source.

11:46

So it's going to get kind of noisy from here.

11:48

I'm sorry about that. The rest of the interview

11:50

is definitely very interesting. But

11:53

it was a pretty noisy room.

11:55

I've been so busy with talking to people. For all

11:57

I know, the system is, got an hour wait

12:00

queue back up. But it won't

12:01

fall over. It will just queue up. It

12:03

took two tries, and it was about 45 seconds,

12:05

but it worked! That

12:06

means, eventually, that's actually, I'm

12:09

like, happy to hear that, not from your

12:11

experience, but it means that we're putting serious

12:13

load on this. It means that people are, this

12:15

is a good day in the history of Python. But

12:18

you have to have a robust architecture to handle

12:20

that and not get things out of order and handle server

12:23

restarts and all of that,

12:25

so that's a, it's a pretty

12:27

engineering response, but yeah, it

12:29

can be

12:29

handled. And to speak to Aaron's answer

12:32

earlier, this is something new that we're doing, but

12:34

we have, what, a good four products at

12:36

the moment? Yes. We have a patient experience

12:38

product, which is going to guide your experience through

12:40

the emergency room. Yeah. And we're doing a bunch of AI

12:43

there. We're predicting, how long are you going to wait for a bed?

12:45

How long are you going to wait until a doctor comes and sees

12:47

you? Yeah. What are the lab results that you're, that

12:49

are coming back, what do they really mean? for you. We've

12:52

got a product for care teams. We're

12:54

providing clinical decision support alerting.

12:57

Are you likely to get sepsis at some point in

13:00

your stay? How likely are you to be admitted? Like allowing

13:02

doctors to manage their workflows

13:05

using this kind of alerting system. We've got a

13:07

system which allows you to find follow

13:09

up care afterwards. And so basically,

13:12

we've been doing this for a long time. We've been doing

13:14

it, what are we, like six years now? Six years, yeah.

13:16

Yeah, we and we've been dealing with this huge pipe

13:18

of patient data for a long time. We're not new to this.

13:21

The event sourcing stuff, that's not for the LLM stuff. That's

13:23

running our systems. That's running our systems at a hundred

13:25

hospitals, a million patient visits. That's, that

13:28

stuff has been the

13:28

easy part for sure. That's right. So if, if this

13:30

sounds foreign or if you don't have a system like

13:33

that with the robust retry mechanism it'll

13:35

take you a couple of years of engineering to get to

13:37

that solid

13:39

system. That's some getting your hands dirty, just in

13:41

the mud. Yeah. Noting and debugging

13:43

just to get there.

13:44

Medical data, the messiest data I've

13:46

worked with

13:48

so far. That's a great, that's a great segue maybe into

13:50

can you tell us a little bit about the process that you

13:53

had to go through to have an LLM,

13:56

Handling HIPAA, HIPAA secure

13:58

patient data. Yeah. I know this is a big fear that

14:00

a lot of enterprise customers have. We

14:02

don't want our trade secrets to get out there.

14:04

We have legal,

14:07

proprietary, interactions with our clients.

14:10

Yeah. We're in a regulated industry, right?

14:12

This is, fortunately

14:15

or unfortunately, not new to me. I

14:17

was the founder of a company called Mint.

14:20

com. We took, The usernames

14:22

and passwords for 25 million

14:25

people and a hundred million bank accounts, including

14:27

me. Yeah, it was a long time ago. Including me!

14:29

That's right.

14:30

No. Yeah,

14:31

and have never had a security breach. At

14:34

least to my knowledge. I sold the company about

14:36

a decade ago. So we're used to dealing

14:38

with sensitive information. You

14:40

want outside penetration testing

14:43

outside audits. HIPAA and HITRUST

14:46

is even more thorough, is

14:48

routine outside security audits. Honestly,

14:51

it can sometimes be a pain to log into our own

14:53

systems requires multi factor fingerprints

14:56

and a drop of blood, but it

14:58

is very secure. You can

15:04

You cannot do this with

15:07

OpenAI. You have to go with,

15:10

Google will sign what's known as a

15:12

BAA, a Business Associates Agreement.

15:15

And it's part of the medical

15:18

chain of liability that says, hey,

15:20

we have the right insurance. If

15:22

we mess up, we have to legally report

15:24

it to you, and you have to report it back to the health

15:27

system. Here's our security practices,

15:29

and we have to look at those, and we have a whole

15:31

compliance office. To do all

15:33

of this.

15:34

And so you actually can't go in some

15:36

sense with the l and m startups.

15:38

Yeah. Microsoft Azure is a fantastic

15:41

choice to start out with. Google's

15:43

been aggressive once they saw what we were doing.

15:45

'cause this has been this has been out internally in our

15:47

products for two or three months.

15:50

And yeah, but they're also Google and Microsoft.

15:52

They know what they're doing when it comes to. security.

15:55

Honestly, when it comes to medical

15:57

information, it's all the people

15:59

who are still running local servers with.

16:02

Yeah, that's it. You want

16:04

to know why they have so many like

16:07

Malware attacks. They're on an old version

16:10

of Windows. They don't patch their stuff. And,

16:12

they may or may not be the,

16:14

the best IT people in the business. I absolutely

16:17

trust the security of AWS

16:20

and Microsoft and Google. Because they have too much

16:22

to lose as companies. We

16:24

have a super secure system.

16:26

And we trial it all

16:29

the time.

16:29

And obviously, our BAA includes

16:32

none of our data being used for training.

16:35

Of course. Yeah. Nice.

16:38

Speaking of the patient experience,

16:40

right? Yeah. If, is it a bespoke

16:42

interaction each time I log onto the app?

16:44

Yeah. Or does it keep my health

16:46

record, so to speak, so I can refer back

16:48

to the last time I used it? And

16:51

then, is that stored locally on

16:53

my device, or is it used, in any

16:55

sort of... process to make

16:56

the tool better. So our primary business

16:58

is a tool that guides you through your visit

17:01

at an ER or inpatient. And

17:03

that is visit based. So we know what your health

17:06

history is and we might show you a little bit of your past

17:08

visit, but it's meant to use at the time

17:10

that you're at the hospital or the emergency

17:12

room or having surgery or something like that.

17:15

And it's just walking you through that experience and

17:17

understand your lab results. These are the videos

17:19

you should watch so you can understand it. These are

17:21

the medications and what you need to know about

17:23

the side effects. We give you

17:26

access to that data for the couple

17:28

weeks following your visit, but we always

17:30

hand you off to the patient at least

17:32

for now. And I will be tight

17:34

lipped about whether you will ever have a full health

17:37

history. IE,

17:39

I've been pitched probably a dozen times on,

17:42

we're the mint for healthcare. And I was like, I

17:44

could do the mint. It's

17:46

vaguely familiar as a

17:48

business concept. I've done this before. So

17:50

nothing to announce today, it's in the back of my mind.

17:53

I'm sure people would, of

17:55

course it would resonate with

17:56

someone to be able to query. years

17:59

and years of interactions, and not to mention

18:01

the opportunities that if you

18:03

apply some machine learning over top of some

18:05

of that diagnostic opportunities

18:07

to catch

18:07

things early. Now it's like you're inside what

18:10

my long term business vision is. Theoretically,

18:12

I could calculate your health future

18:15

if I had a big enough data set. Yep. And

18:18

keep in mind that I... At

18:21

Vital, we now see 2% of all U.

18:23

S. Emergency medicines. Wow! For

18:25

a startup that's been around

18:27

for not that long, that's a pretty

18:30

good sample size. We can see

18:32

how diseases progress and,

18:34

that there's more of this type of fall in the winter than

18:36

there is in the summer, right?

18:39

I know there's entire industries, like the health

18:42

insurance

18:42

industry

18:43

that has it modeled on curves

18:45

exactly when you're gonna die based on, the

18:47

fact that you went skydiving once when

18:49

you

18:49

were 31. Sure. Yeah. I

18:52

could probably predict whether, Bird

18:54

and Lime are doing business well based

18:56

on the number of elbow injuries and wrist

18:58

fractures that we can

19:00

plot over time. That's unfortunately

19:03

not a joke. Wow.

19:04

Okay, then I have to ask, since Google

19:07

got rid of the, I forget what they called it, but the

19:09

flu predictor feature that they had for so long?

19:12

Is that something you guys might potentially product?

19:15

No, we won't use that

19:17

sort of stuff. It's really interesting and we probably could do

19:19

it internally. But And we did come

19:21

up with a COVID checker. We did come up with a COVID

19:24

checker that was used a million and a half times. Wow. Yeah.

19:26

We were the first one out before

19:28

Google, before Microsoft. We were, CDC

19:31

considered using us. I was literally on the phone

19:33

with the White House Task Force in the middle of the night developing

19:35

this thing. A million and a half uses

19:38

within the first month. We did the COVID checking for the state

19:40

of

19:40

Oregon, right? Yeah. The whole state. We

19:42

pivoted the whole company as soon as the pandemic started.

19:45

Yeah. Said, okay, we've got all this health data coming

19:47

in. We've got the data science chart. Let's try

19:49

and do something quickly with that. Yeah, nice.

19:51

But the sort of north star for

19:53

the company is what's right for the patient.

19:55

will it improve patient outcomes? I'm

19:58

really tired of most of healthcare.

20:00

I'm looking at you. Medicare Advantage. Who

20:03

is, frankly, just financial arbitrage.

20:06

They're basically like, Okay, so the government says

20:08

New York's a more expensive place. We'll pay 1, 400

20:11

a month for somebody over 65. Phoenix

20:14

is cheaper, so we'll pay you 1, 100.

20:17

And Medicare Advantage companies are just like, You

20:19

know what, we'll advertise in rich zip

20:21

codes to get healthy, wealthy people

20:23

and we'll leave the rest to the public system. They're

20:25

not improving patient outcomes. They're not

20:28

increasing utilization. They put up barriers

20:31

and blocks. Like you have to get a referral

20:33

from your primary care doctor. We

20:35

will do none of that. There are lots of ways to make

20:37

money in healthcare. Our investors sometimes

20:39

push us towards that. I

20:42

have fortunately had a successful start up.

20:44

I don't know, I'm not doing this, for the money primarily.

20:47

Just want to do what actually affects

20:50

patient health.

20:52

Yeah, we all have other things that we could be doing, there

20:55

are other ways to make money, but we, I've never

20:57

been in a more mission led company,

20:59

so thank you first.

21:00

Yeah, and it resonates with everybody,

21:03

even if they're healthy, we've got parents,

21:05

we've got grandparents, who like who wouldn't feel empowered

21:08

and able to help them out just with their

21:10

care make them feel a little more at ease during

21:12

a A time of struggle.

21:14

Completely. And actually I think one of the best use

21:16

cases for what we launched today, Vital.

21:18

io slash Translate, is

21:21

if you have an elderly parent or somebody

21:23

that you're caring for, especially

21:25

if they're elderly and they're a little confused

21:28

and they went to the doctor's office and they're like,

21:30

hey dad, what did ahhh.

21:34

Put their notes in there and see what actually

21:36

comes out. Diseases and issues

21:39

they actually have. Yeah. I

21:41

was talking I don't know the full story, but,

21:44

I had a friend whose sister died

21:46

basically because they didn't catch something

21:48

that was on page three or four. Because

21:51

humans can't scan text

21:54

that quickly. And you might have hundreds of pages

21:57

of medical history if you're a chronically

21:59

ill person. And sometimes

22:01

that history really matters. And doctors give

22:03

it like two or three minutes to maybe

22:05

scan through. AI does a way

22:07

better job of picking out. The stuff that

22:10

they might need. The fair comparison, this

22:12

is, listen, This, we've marked

22:14

it as 99. 4% safe

22:17

for animal adopters, independent people, employed

22:19

by the company. It's

22:22

not. Without risk. If 1 in 200

22:24

times, it'll miss something small. But

22:26

doctors miss something big. 1

22:29

in every 10 times. And so the stats

22:31

are actually much better for AI than they are for

22:33

humans, and that's the problem. And

22:36

when we, we have some great clinicians in

22:38

our team, when I talk to our

22:40

clinical staff, our advisory board

22:42

about the stuff that they really want to

22:45

see, all of them talk about patients paying

22:47

attention to and understanding their discharge instructions.

22:50

The value there is enormous. The value in terms

22:52

of long term care and in terms of immediate

22:54

outcomes is huge.

22:55

Nice. Were there

22:57

different specialties within

22:59

medicine that were a little more challenging?

23:03

We started out with medical imaging. Medical

23:06

imaging is nice because it's confined. CT

23:08

scans, x rays, MRIs and

23:11

then, what we released today I

23:13

don't know what people are going to put into

23:15

it. And so it has to be pretty robust

23:17

to doctor's notes, nurse's notes

23:20

lab results, all sorts of things. It's

23:30

time to wrap it up. Yeah. That's

23:33

a good time to wrap it up. We really

23:35

appreciate your time today to,

23:37

come talk to us guys. Such a product that I think

23:39

everyone can. You can benefit from learn

23:41

from other family members of these. And

23:43

just remind listeners at HTTTA The Rob's

23:46

Engineering Podcast.

23:48

Yeah, thank you. We really appreciate the

23:50

time. Yeah, Thank you for having us. Yeah, we never get

23:52

to talk about the nerdy tech stuff. Dude, we

23:55

can go even harder. Oh,

23:57

I'm gonna change the memory card for that, yeah, I

23:59

think I held off. I'm like, all right, tell us about your air handler

24:02

later. Yeah, I was like no,

24:04

we're not going that

24:04

deep. We're not going

24:05

that deep. Fantastic. Thank

24:07

you. Thank you

24:09

guys.

24:12

Thanks for coming to the prompt engineering podcasts

24:14

podcast dedicated helping

24:16

you be a better prompt engineer Episodes

24:19

are released every Wednesday I

24:21

also host weekly masterminds where

24:23

you can collaborate with me and 50 other people

24:26

live on zoom to improve

24:28

your prompts Join us

24:30

at promptengineeringmastermind. com

24:32

for the schedule of the upcoming masterminds. Finally,

24:36

please remember to like and subscribe.

24:39

If you're listening to the audio podcast, rate

24:41

us five stars. That helps us teach more

24:43

people. And if you're listening to

24:45

the podcast, you might want to join us

24:47

on YouTube so you can actually see the prompts.

24:50

You can do that by going to youtube. com

24:53

slash at prompt engineering

24:56

podcast. See

24:58

you next week.