0:00

So, Hey

0:05

listeners, welcome back to No Priors. This

0:07

episode marks a special milestone. Today is

0:09

our hundredth show. Thank you so much

0:11

for tuning in each week with me

0:13

and Alad. And it's been an exciting

0:15

last couple of weeks in AI, so

0:18

we have lots to talk about. Why

0:20

don't we start with the news of

0:22

the hour, or you know, really the

0:24

last month at this point, and Deep

0:26

Seek. Alad, what's your overall reaction? Deep

0:28

Seek is one of those things which is

0:30

about... really important in some ways and

0:32

then also kind of what you'd expect

0:35

would happen from a trendline perspective. And

0:37

I think there was a lot of

0:39

interest around Deep Seek for sort of

0:41

three reasons. Number one, it was a

0:44

state-of-the-art Chinese model that seemed really caught

0:46

up with a number of things on the

0:48

reasoning side and in other areas relative to

0:50

some of the Western models. And

0:52

it was open source. Number two, there was

0:54

a claim that it was done very cheaply.

0:56

So I think the paper talked about like a...

0:58

a five and a half million dollar run. It's

1:00

sort of the end. And then lastly, I think

1:03

there's this broader narrative of who's really behind

1:05

it and what's going on and some perception

1:07

of mystery which may or may not be real.

1:09

And as you kind of walk through each one

1:12

of those things, I think on the first one,

1:14

you know, state of the art open source model

1:16

with some recent capabilities. and they actually did some

1:18

really nice work. You read through the paper, there's

1:21

some novel techniques in RL that they worked on

1:23

and that, you know, I know some other labs

1:25

just starting to adapt. I think some other labs

1:27

would also come up with some similar things over

1:29

time, but I think it was clear they'd

1:32

done some real work there. On the cost

1:34

side, everybody that I've at least talked to

1:36

who's savvy to, it basically views every sort

1:38

of final run for a model of this type

1:40

to roughly be in that kind of dollar range.

1:43

you know, $5 to $10 million or something like

1:45

that. And really the question is how much

1:47

work when in behind that, before they distill

1:49

down this smaller model. And by census,

1:51

everybody thinks that they were spending hundreds

1:54

of millions of dollars on compute leading

1:56

up to this. And so from that perspective,

1:58

it wasn't really novel. I think that's

2:00

sort of 20% drop in invidious stock

2:02

and everything else that happened as news

2:05

of this model spread was a bit

2:07

unwarranted. And then the last one, which

2:09

is sort of speculation of what's going

2:11

on, is it really a hedge fund, is

2:14

something else happening, like, you know, thoughts

2:16

a little bit, well, speculative. There's all

2:18

sorts of reasons that it is exactly

2:20

what they say it is, and then

2:23

there's some circumstances in which you could

2:25

interpret things more broadly. So that's

2:27

kind of my read things more probably.

2:29

to it, but to your point, it's

2:31

also like what you might expect,

2:33

especially given historical precedent with like

2:35

GPT 335 and then chat GPTs.

2:37

So like Deep Seek v3, like

2:39

the base model, big AI model,

2:41

pre-trained on a lot of internet

2:43

data to predict next tokens, like

2:45

that was out in December. right

2:48

and invidious stock did not crash

2:50

based on that news. So I

2:52

think it's just interesting to recognize

2:54

that like people obviously do not

2:56

just want raw likelihood of next

2:58

word in the streaming way and

3:00

the the work of post training

3:02

and making it more useful for

3:04

human feedback or more specific data

3:06

like high quality examples of prompts

3:08

and responses just like we've seen

3:10

with the chat models like chat

3:12

gBT the instruction fine tuning that

3:14

made this such a breakthrough experience

3:16

like That really mattered. And then

3:18

the, as you said, the like

3:20

narrative violation, release of R1, reasoning

3:22

model, as a parallel model to

3:24

like opening eyes, oh, one. I

3:26

think that was also the breakthrough

3:28

moment in terms of people's understanding

3:30

of this. Well, it's also like

3:33

20 years of China, America, technology dominance

3:35

narrative, right? Yes. Like I think I

3:37

think it was also kind of this

3:40

like iced around US versus China. you

3:42

know, worse, you know, the West is far ahead.

3:44

And so, you know, will they ever catch up,

3:46

etc. And this kind of showed that Chinese models

3:48

can get really fast. But I do think the

3:50

cost thing was a huge component of it. And

3:52

again, I think cost me have been in some

3:54

sense misstated or misunderstood at least. It's not

3:57

clear to me that like final model runs

3:59

at scale are. in this price range,

4:01

but I think what you were

4:03

saying before I completely agree with,

4:05

experimentation tends to be a multiple,

4:07

like you need to have tooling

4:09

and data work and the experimentation

4:12

and the pre-training run and data

4:14

generation cost and post-training and inference,

4:16

right? I'm sure I'm missing something

4:18

here. It seems very unlikely that

4:20

there hasn't been a large multiple

4:23

of $6 million spent in total.

4:25

But I think there was also

4:27

a narrative violation here in that,

4:29

like, even at a multiple of

4:31

$6 million, it's not like a

4:33

multi-billion dollar entry price or a

4:35

Stargate-sized entry price to go compete.

4:37

And I think that is something

4:39

that really shook the market. That should

4:42

be expected, because if you look at

4:44

the cost of training a GPT-4 level

4:46

model today versus two years ago, It's

4:48

a massive drop-in cost. And if

4:50

you look at, for example, inference costs

4:52

for a GPT-4 level model, somebody in

4:55

my team kind of worked out, and

4:57

in the last 18 months we saw

4:59

a 180x decrease in cost per

5:01

token for equivalent level models.

5:04

180x, not 180%, 180 times. So

5:06

the cost collapse on these things

5:08

is already quite clear. That's true

5:10

in terms of training equivalent

5:12

models. That's true in terms of

5:14

inference. And so, again, I kind of

5:17

just view this. roughly on trend and

5:19

maybe it's a little bit better and they've

5:21

come up with some advanced techniques which you

5:23

know they absolutely have but it does

5:25

feel to me a little bit overstated from

5:27

the perspective of how radical it is. I

5:29

do think it's striking what they did and

5:31

it kind of pushes US open source forward as

5:33

well which I think will be really important. But

5:36

I think people need to really look at

5:38

the broader picture of these curves that

5:40

are already happening. Do you think it's

5:42

proof that models are commoditizing the fact

5:44

that they're the last 18 months. There's

5:46

this really great website called the

5:48

artificial analysis dot AI that actually

5:50

allows you to look at the various

5:52

models and their relative performance across a

5:55

variety of different benchmarks and the

5:57

people who run this actually do

5:59

the benchmark. themselves. They'll go ahead and

6:01

retest it. Processes take the paper at

6:03

face value. And you see that for

6:05

a variety of different areas, these models

6:07

have been growing closer and closer in

6:09

performance. And there's different aspects of reasoning

6:11

and knowledge, scientific reasoning and knowledge, quantitative

6:14

reasoning and math. coding, multilinguality, cost per

6:16

token relative to performance. And they kind

6:18

of graph this all out for you.

6:20

And they show you by provider, by

6:22

state of the art model, how do

6:24

things compare? And things are getting closer

6:26

over time versus more dispersed over time.

6:28

So I think in general, the trend

6:30

line is already in this direction where

6:32

it seems like a lot of people

6:34

have moved closer and closer to Valencia

6:36

than they were, say, 18 months ago

6:38

where I think there was enormous disparities.

6:40

And obviously there are certain areas where

6:42

different areas where different models are quite

6:44

a bit ahead still. But on the

6:46

average things we're starting to net out

6:48

a little bit more and that may

6:50

change right? Maybe somebody comes out with

6:52

an amazing breakthrough model and they leapfrog

6:54

everybody else for a while But it

6:56

does seem like the market has gotten

6:58

closer than it than it was even

7:00

just like a year ago What do

7:02

you think it was even just like

7:04

a year ago? What do you think

7:06

is the value? It was even just

7:08

like a year ago? It's just like

7:11

a year than it was even just

7:13

like a year ago a year ago.

7:15

What's even just like a year ago

7:17

a year ago a year ago a

7:19

year ago? just like a year ago?

7:21

just like a year ago. Just like

7:23

a year ago. Just like a year

7:25

ago just like a year ago. Just

7:27

like a year ago. Just like a

7:29

year ago. Just like a year ago.

7:31

Just like a year ago. Just like

7:33

a year ago. Just like a year

7:35

ago. Just like a year ago. Just

7:37

like a year ago. Just like a

7:39

year ago. Just like a year ago.

7:41

Just like a year ago. Just like

7:43

then having something is dramatically better to

7:45

make a difference. So that could be

7:47

data labeling, it could be artificial data

7:49

generation, it could be other aspects of

7:51

post-training. So I think there's lots of

7:53

things that you could start doing when

7:55

you have a really good model to

7:57

help you. It could be coding and

7:59

sort of coding tools, it could be

8:01

all sorts of things. You know, there

8:03

is an argument that some people make

8:06

that at some point, as you move

8:08

closer and closer to some forum lift

8:10

off. that the more state of the

8:12

art the model is the more it

8:14

bootstraps into the next model faster and

8:16

then it just accelerates for you and

8:18

you stay ahead I don't know if

8:20

that's true or not I'm just saying

8:22

that's that's something that some people speculate

8:24

on sometimes are there other things that

8:26

you can think of? No I think

8:28

one thing you mentioned maybe if I

8:30

just extend it is like kind of

8:32

underpriced or like not yet understood enough

8:34

to market as a theory, which is

8:36

the idea that if you have a

8:38

like high quality enough base model to

8:40

be doing synthetic data generation for like

8:42

a next generation of model, that is

8:44

actually like a big leveler. Right. And

8:46

if you believe that there will be

8:48

continued availability of like more and more

8:50

powerful base models, that that's a big

8:52

leveler of the playing field in terms

8:54

of having like, you know, self-improving models.

8:56

And so that's an interesting thing that

8:58

people have not really, really talked about.

9:01

There are different ways to have value

9:03

from being at the frontier. One of

9:05

the things that was really interesting to

9:07

me was that like the Deep Seek

9:09

mobile app became like, you know, cop

9:11

contender in the app store for a

9:13

little bit. like cheapest most capable model

9:15

in the market actually matters to consumers

9:17

and they can tell and that will

9:19

drive consumer adoption and that's what happened

9:21

and like that's why you need to

9:23

have the soda model to create these

9:25

new experiences and there's a competing view

9:27

which is just like well like this

9:29

whole drama is quite interesting and people

9:31

are trying it as much because like

9:33

they want to see what the leading

9:35

Chinese AI model is like if it's

9:37

as good as open AI an anthropic

9:39

and such I definitely believe that leading

9:41

capability can lead to not product that

9:43

draws consumer attention, but I think in

9:45

this case it's more the latter. Two

9:47

other things that kind of happened this

9:49

past week was on the open AI

9:51

side. One is they released deep research.

9:53

So speaking of really interesting advancements and

9:56

capabilities, and then secondly they announced Stargate,

9:58

which was, you know, a massive series

10:00

of investments across AI infrastructure that was

10:02

announced with Trump at the White House,

10:04

what are your views on those two

10:06

things that in some sense kind of

10:08

overlap in terms of open AI really

10:10

advancing different aspects of... state-of-the-art terms of

10:12

what's happening right now. Deep research is

10:14

a really cool product. I encourage everybody

10:16

to try it. The biggest deal to

10:18

me is that it immediately raises the

10:20

bar for a number of different types

10:22

of knowledge work where like a, you

10:24

know, where I might have hired a...

10:26

median intern or analyst before, I mean

10:28

we don't do that here, but like

10:30

where one could hire a median analyst

10:32

or intern, I'm going to immediately comp

10:34

a bunch of their work to what

10:36

you could do with deep research and

10:38

like your ability to do better with

10:40

deep research. And like the comp is

10:42

hard. I'd say it is a really

10:44

valuable product. I expect other people to

10:46

adopt this pattern too, but I think

10:48

it's a really novel innovation. Kudos to

10:50

the team. I would say I think

10:53

it is more useful, at least to

10:55

me, upon first blush. I'm sure they're

10:57

working on this. In domains, I understand

10:59

less to do surveying and to like

11:01

make sure I have a comprehensive view

11:03

and understand who the experts are versus

11:05

like in an area where I feel

11:07

like I have a lot of depth.

11:09

I take issue with its implicit authority

11:11

ranking and its ability to determine like

11:13

what ideas out there, what on the

11:15

web is good and not when it's

11:17

doing its search. from at least my

11:19

initial prompting and experimentation in domain. I'm

11:21

like, oh man, like you're really gonna

11:23

have to audit the outputs here. It

11:25

will orient you, but you can't take

11:27

as given like many of the claims

11:29

here. This is the AI form of

11:31

Murray, Murray Gelman Amnesia, which was coined

11:33

by the guy who wrote Jurassic Park.

11:35

I can't remember his name as pronounced

11:37

Gelman or Gelman. Murray Gelman was a

11:39

physicist who came up with quarks and

11:41

a few other things. He was a

11:43

Nobel Prize winner and was considered widely

11:45

brilliant. And it was named after him

11:48

by Michael Crichton, which was basically the

11:50

idea is if you're reading a page

11:52

in the New York Times about something

11:54

you really understand, and you're like, oh,

11:56

this is so dumb, and how could

11:58

they write this, and I don't believe

12:00

it. And then you just turn the

12:02

page and you look at something, you

12:04

look at something, you look at something,

12:06

you don't believe it. And then you

12:08

just turn the page and you look

12:10

at something, you don't know, and you

12:12

look at something, and you look at

12:14

something, or not know, and... You know,

12:16

if it's getting sort of expertise wrong

12:18

in a domain, I understand. Does that

12:20

mean they're also getting it wrong? And

12:22

domains, I don't understand, but of course

12:24

we never apply that as people. We

12:26

just assume, of course, it's right. domains

12:28

that we don't understand, which I think

12:30

is really interesting psychologically, but it also

12:32

has real implications in terms of how

12:34

people will use AI in the future

12:36

in general, because these things will become

12:38

the definitive source of a lot of

12:40

people's primary information. It's in some senses

12:43

really overlapping with some of the search

12:45

use cases in really deep ways. And

12:47

you have something where the sources traditionally

12:49

have been less evident. I know that

12:51

people are working on different ways to

12:53

surface, but the primary sources are for

12:55

some of these things. It does have

12:57

really interesting implications for how you think

12:59

about knowledge in the modern era as

13:01

you're using AI, especially as you're using

13:03

agents, so they just go and do

13:05

stuff and then report back and you

13:07

know what they did. So I think

13:09

it's a very interesting topic. I'm not

13:11

sure how you solve that from like

13:13

a UX perspective, or maybe it's like

13:15

somewhat unsolvable given, you know, it also

13:17

reflects like what is knowledge on the

13:19

web. It really does feel like a

13:21

really dangerous thing from sort of a

13:23

really dangerous thing from sort of a

13:25

propaganda and censorship censorship perspective and censorship

13:27

perspective. You know, social networks were kind

13:29

of view one of that, or maybe

13:31

certain aspects of the web were V1

13:33

and social networks were B2, and this

13:35

is kind of the big version, because

13:38

it's a mix of search. It's like

13:40

if you mix Google with Twitter, with

13:42

Facebook, with everything else that you're using,

13:44

with all the media outputs, or media

13:46

outlets, all into one single device that

13:48

you interrogate, that's kind of where these

13:50

AIs are going. And so the ability

13:52

to control the output of these things

13:54

is extremely powerful, but also very dangerous.

13:56

You know, that's why I'm kind of

13:58

happy that we're in this multi-AI world,

14:00

multi-company world. There's a way to offset

14:02

that, and that's where Obitsworth becomes incredibly

14:04

important if you worry about civil liberties.

14:06

What do you think about Stargate? Maybe

14:08

there's like a couple different implied questions

14:10

in like... Stargate, right? One is how

14:12

much does it matter in the race,

14:14

like to continue to have access to

14:16

the largest infrastructure? I'm going to skip

14:18

the question about like whether or not

14:20

it's real, like there's a lot of

14:22

money involved here. I think another question

14:24

is how deep are the capital markets

14:26

to continue funding this stuff? Maybe a

14:28

final one is like just the involvement

14:30

of different sovereigns or quasi-soft. in this,

14:32

like I don't know if I have

14:35

a strong opinion on the latter two,

14:37

the way I think about the dynamic

14:39

of like how much does the capital

14:41

matter and like the implied like how

14:43

like do we continue to see scaling

14:45

on pre-training be a dominant factor is

14:47

I think of it really as like

14:49

uncertainty rather than than risk. Right, like

14:51

if you think about capabilities as emergent

14:53

and people not being sure what sort

14:55

of algorithmic efficiencies counteract like the, you

14:57

know, improvements that will come from more

14:59

scale and the things you can do

15:01

to generate new data to improve in

15:03

other vectors and what we're going to

15:05

get out of test time scaling, like

15:07

I just think it's very hard to

15:09

predict, but I failed to see a

15:11

scenario where anybody trying to build AGI,

15:13

any of the large research labs wouldn't

15:15

want the biggest cluster that they could

15:17

have. if it was free, right, or

15:19

if the capital was available to them.

15:21

And that to me says more than

15:23

anything else, like, we're going to get

15:25

more out of pre-training. Is it going

15:27

to be as efficient? Like, I think

15:30

that's unlikely. We're like a little bit

15:32

delayed on this, but we'll just give

15:34

ourselves a free pass, given it's episode

15:36

100. Predictions for 2025. Happy New Year!

15:38

It's February, but it's like anything happening.

15:40

It's like the Larry David episode. Yeah,

15:42

basically there's some statute of limitations of

15:44

how late into the year you can

15:46

say happy new year. We're now a

15:48

month in, so of course, we're way

15:50

over that. We should probably say happy

15:52

Valentine's Day, even though we're like two

15:54

weeks early. No, a lot. You don't

15:56

like, what was the vibe for 2025

15:58

is you can just do things that

16:00

are likely to happen. First, the foundation

16:02

model market should at least partially consolidate.

16:04

And it may be in those sort

16:06

of ancillary areas, so that's image, video,

16:08

voice, you know, a few other areas

16:10

like that, maybe some secondary LLCs or

16:12

foundation models will also consolidate. So I

16:14

do think we're going to see a

16:16

lot of consolidation, particularly if the FTC

16:18

is a little bit more friendly than

16:20

the prior regime. We'll also see some

16:22

expansion of sort of new races in

16:25

physics, biology, and materials, and the like.

16:27

So I think that that will happen

16:29

alongside just general scaling of foundation models

16:31

will continue. And that includes reasoning and

16:33

that includes other things. So that's one

16:35

big area. I think the second area

16:37

is we're going to see. vertical AI

16:39

apps continue to work at scale. It's

16:41

hard to be for legal, that could

16:43

go on, and Sierra for customer success,

16:45

and a variety of folks for code

16:47

gen, for medical scribing, etc. So I

16:49

think it'll be the era of vertical

16:51

apps, and I think a subset of

16:53

those will start adding more and more

16:55

agentic things to them. You know, some

16:57

folks like cognition are doing that. Third

16:59

would be self-driving, will get a lot

17:01

of attention, Robo taxis, etc. Applied intuition

17:03

I think is kind of a dark

17:05

course to watch more generally on the

17:07

automotive stock. And then I guess fourth

17:09

is that some consumer things I think

17:11

will get really large scale experiments happening

17:13

in a way that hasn't happened until

17:15

now. So I'm starting to see consumer

17:17

startups. I'm starting to see more consumer

17:20

applications from incumbents. Like I actually think

17:22

we're we're going to see a little

17:24

bit of a resurgence in consumer. It

17:26

may take a while, but I think

17:28

that'll happen. And then lastly I think

17:30

there's things that we all know will

17:32

happen. But it may start to see

17:34

some interesting behavior on agents, maybe some

17:36

early robot stuff, you know, but it'll

17:38

be one of those things where it's

17:40

more going to be the glimmer of

17:42

how this thing will work versus the

17:44

whole thing. But I think some of

17:46

those developments will be very exciting. So

17:48

those would be my five predictions for

17:50

25. How about you? What do you

17:52

got? We agree on a number of

17:54

different things. I think the whole like

17:56

definition for agent is super fuzzy, but

17:58

if we just think of it as

18:00

like do multi-step tasks successfully in some

18:02

sort of end-user environment and take action

18:04

beyond just generate content, like we're already

18:06

seeing that I think we're to see

18:08

that more broad. as you were sort

18:10

of alluding to, but I think companies

18:12

get better and product companies or vertically

18:14

integrated companies, they get better at handling

18:17

failure cases and managing state intelligently. And

18:19

so we're already seeing that in security

18:21

and support and an SRE. And I

18:23

think that will continue to happen. This

18:25

already happened in CodeGen, as you were

18:27

sort of alluding to, but I think

18:29

companies doing co-pilot products will naturally extend

18:31

to agents. They'll just try to do

18:33

more, right, and take more on. I

18:35

think one of the inputs to broader

18:37

consumer experimentation as you. describe as just

18:39

like way more capable like small low

18:41

latency models. I don't think we have

18:43

like any monotonic movement toward compute at

18:45

the edge like I think when people

18:47

are like edge compute for the sake

18:49

of edge compute I'm like nobody cared.

18:51

Right, but if you can make that

18:53

transparent to the user and it's free,

18:55

then I think your ability to ship

18:57

things that are free is obviously unlocked

18:59

and I think that's cool. This is

19:01

something that will be a lot of

19:03

web apps, you know, so I don't

19:05

think it has to necessarily be on

19:07

device. Consumer products, I add a later

19:09

point, there won't be some, but I

19:12

also just think it's just going to

19:14

be things running on the internet that

19:16

just become part of your application stack

19:18

on your browser that will do really

19:20

interesting. Yeah, well stuff in the browser

19:22

can also use the GPU, but I

19:24

just think that the the ability to

19:26

run locally might be a big unlock

19:28

for them. I don't know if you

19:30

and I disagree on timeline. I think

19:32

we're going to see technical proof of

19:34

breakthroughs in robotics and in generalization this

19:36

year, though not deployments. I think one

19:38

thing that's like maybe miss. priced just

19:40

because it's very new is like people

19:42

don't really know how to think about

19:44

reasoning. I would claim that one thing

19:46

is as much improvement in reliability as

19:48

complexity of task. One like mistake that

19:50

entrepreneurs and investors make that I have

19:52

made is like you look at something

19:54

and it's not working and like the

19:56

issue is it is like a technical

19:58

issue and then you like assume it's

20:00

not going to work but I think

20:02

an AI you have to like keep

20:04

looking again and again and again because

20:07

stuff can begin to work really quickly.

20:09

Maybe one last one with that I'm

20:11

just, I've seen like small examples of

20:13

with our like embed program and also

20:15

broadly in the portfolio is because you

20:17

have this diffusion of innovation like not

20:19

just with customers but with the types

20:21

of entrepreneurs that go take something on

20:23

and we're like beyond just the tip

20:25

of the spear now and more and

20:27

more people like I can do stuff

20:29

with AI. I think we're going to

20:31

get more... smart data generation strategies for

20:33

different domains where you need like domain

20:35

knowledge as well as understanding of AI.

20:37

So examples here could be like biology

20:39

and material science. Like you needed the

20:41

set of scientists who are capable of

20:43

innovating on data capture, which might literally

20:45

be like a biotech innovation versus a

20:47

computer science innovation to understand the potential

20:49

of deep learning and that the bottleneck

20:51

was data and then the type of

20:53

data you were like looking for. And

20:55

I think that is happening. That's really

20:57

exciting. This may be the year where

20:59

we see something really interesting happen on

21:01

the health side. As an example, where

21:04

you need specialized data, but it's not

21:06

as hard as, you know, the atomic

21:08

world of, you know, biomolecular design or

21:10

something. Anything else we should talk about?

21:12

The facial hair? We catch, should I

21:14

bring it back? I liked the beard.

21:16

I like the beard and hat era.

21:18

Oh, interesting. I should go back to

21:20

that. The last question for today. We're

21:22

on episode 100. What do you think

21:24

the state of the world will be

21:26

relative to AI when we're at episode

21:28

200? I don't think we're part of

21:30

this anymore. I think it's just like

21:32

two agents going back and forth teaching

21:34

us stuff and like you and I

21:36

are no longer the hosts or the

21:38

choosers of topics were just like nodes

21:40

into the network. Will they be as

21:42

good-looking as us? Yeah, and they'll be

21:44

better. Computers. We'll see. Still like some

21:46

art more than submit or mid journey

21:48

art of those and be just beautiful,

21:50

beautiful things on there. Okay, episode 200,

21:52

that's like what? Well, it's almost two years, it's weekly, so.

21:54

I think we're either in the RLHF farm or we're like sitting

21:57

on a... in a visa -abundance. That's a prediction.

21:59

You heard it here first. Well,

22:01

hopefully I'll see you here 200 hopefully

22:03

I'll I think episode not as great.

22:05

a visa. I And all the listeners

22:07

too. Thanks is not right. Thanks great. Okay. And

22:09

all the Find us on Twitter guys. All

22:12

right. Thanks, Subscribe to our YouTube channel

22:14

if you want to see our

22:16

faces. Follow the show on Apple

22:18

YouTube channel Spotify, you or wherever you

22:20

listen. That way you get a

22:22

new episode every week. And sign

22:24

up for emails or find transcripts

22:26

for every episode at you get a new.com. week.