Full Transcript

·YouTLDR

How AI is changing Software Engineering: A Conversation with Gergely Orosz, @pragmaticengineer

26:285,568 words · ~28 min readEnglishTranscribed Apr 23, 2026
AI Summary

Large tech companies are weaponizing AI usage metrics through 'token maxing,' leading to performative automation while forward-thinking firms transition toward 'harness engineering' and internal agent orchestration platforms.

It exposes the cultural and operational shifts within Big Tech—from the gamification of AI metrics to the collapse of devops and product roles into the single software engineer.

Section summaries

0:00-1:00

Intro to Gergely and Token Maxing

optional

Introductory context about who Gergely is and the first mention of token maxing.

1:00-5:00

The Dark Side of AI Metrics

watch

Essential for understanding how Big Tech currently mismanages AI adoption and performance reviews.

5:00-11:00

Management Pressure vs. Actual Productivity

optional

Discusses leadership's push for AI and some specific studies on productivity results.

11:00-13:00

How to Get Good at AI

watch

Crucial insight on why theoretical knowledge fails and only experience works for agent workflows.

13:00-18:00

The Future of the Software Engineer Role

watch

Discusses the collapse of DevOps/Testing roles and the 'Mech Suit' manager metaphor.

18:00-23:00

Internal AI Infrastructure (Uber, Shopify)

watch

Highly relevant for platform engineers interested in MCP gateways and agentic architectures.

23:00-26:00

The Pragmatic Engineer Origin Story

skip

Personal history about the newsletter's growth; interesting but not related to AI technical strategy.

Key points

  • Token Maxing and Metric Perversion — Large firms like Meta and Salesforce use token expenditure leaderboards as a proxy for performance, leading engineers to 'token max'—running autonomous agents to build 'junk' or summarize docs they could easily read—to avoid the bottom 25% rank.
  • The Rise of Internal Agent Infrastructure — Companies like Uber and Airbnb are building custom internal agent gateways and MCP (Model Context Protocol) discovery systems instead of just buying off-the-shelf tools, often because internal codebases exceed current context windows.
  • Engineering as the 'Mech Suit' Manager — The role is shifting from manual coding to an orchestration role, similar to a tech lead or mentor, but for agents. Unlike managing people, managing agents provides an immediate feedback loop and lacks 'people drama.'
  • Trading Churn for Frontier Access — Shopify’s early adoption of Copilot involved enduring bugs and high churn in exchange for being 6-12 months ahead of the competition. This 'innovation recruitment' strategy justifies the high cost of current AI friction.
In Microsoft, there's a leaderboard... it's ridiculous how some people are just running autonomous agents to build junk honestly for the sake of having that number go up. Gergely Orosz
If you are at a large company and you're not already building an MCP gateway, what are you even doing? Gergely Orosz

AI-generated from the transcript. May contain errors.

0:00

Sure

0:17

game.

0:21

>> All right. I going to assume most of you

0:23

uh show of hands who subscribes to

0:25

Pragmatic Engineer. Oh my god.

0:27

>> Wow.

0:28

uh he is uh he needs no introduction.

0:32

Then let's get right into it. Um

0:35

what is token maxing and should everyone

0:37

here be doing it?

0:40

>> So I I heard about token maxing a week

0:43

ago or like week and a half ago first

0:45

and you know some people have been doing

0:46

it for longer and I tweeted about it I

0:49

think three days ago saying oh there's

0:50

this token maxing and again you see it

0:52

on social media and my DMs were blowing

0:55

up from from people at large companies.

0:57

I don't want to name names but like you

0:58

know Meta, Microsoft

1:01

uh some so some some other ones as well

1:03

like uh the likes of and and and so so

1:06

many more and the story is a little bit

1:08

different every at every company on why

1:12

people are doing it and whether they

1:14

like it or whether they think it's good.

1:15

But there's a few a few common themes.

1:17

One is token output at these larger

1:20

companies is measured in in some way.

1:23

There's like either a leaderboard or

1:24

there's a way to look up your your

1:26

peers. Salesforce, for example, you can

1:29

check the spend the the money spent that

1:32

every every person at the company did.

1:34

You can like search in a tool that

1:36

someone built and it shows how many

1:38

dollars they spent on on AI related

1:40

tokens. And you know, first there's this

1:43

number, then there's this uncertainty on

1:45

in the tech industry, right? We're kind

1:47

of hearing layoffs, like massive cuts at

1:49

the likes of block. And I mean there

1:52

like no matter how much tokens people

1:54

spend they were let go independent of

1:55

this but people start to think like does

1:57

is it part of performance evaluations or

2:00

promotions or all that and the answer is

2:03

kind of. So inside of meta I talk with

2:07

managers and in the performance

2:09

evaluation they have this data point

2:11

which is one of many data points right

2:12

the same way as as like diffs or impact

2:16

or or code reviews of how helpful this

2:19

person is but they do just like with any

2:21

data point they sometimes pull it in and

2:24

use it. So typically in just like any

2:26

data point it can be weaponized. So like

2:28

a low performer with low impact and a

2:30

low token count clearly not even trying.

2:32

So, and a high performer with high

2:34

impact and high token count. Clearly,

2:36

that's innovating and this must be doing

2:37

good. So, inside of these companies

2:39

specifically, I talked with a lot of

2:40

people at at Meta. And again, this is

2:42

not representative 100% of Meta, but

2:44

they had this leaderboard where people

2:45

showed up and they have like massive

2:47

amounts of tokens and a lot of engineers

2:49

got just scared, worried, so they

2:51

started to token max to try to generate

2:53

tokens. stories that I've heard first or

2:55

well secondhand from these people who

2:57

who who told me firsthand is for example

3:00

instead of reading the documentation I

3:02

will ask the agent to summarize it for

3:04

me and ask questions even though it

3:05

doesn't do a good job answering it but

3:07

my token count goes up people just want

3:10

to not be in the bottom 25% or bottom

3:12

50% for token count where these things

3:14

are measured inside of Microsoft again

3:16

there's a leaderboard and I'm talking

3:18

with people they're like it's ridiculous

3:20

like how some people are just running

3:22

autonomous agents to build junk honestly

3:24

for the sake of having that number go up

3:27

and and sometimes it gets ridiculous

3:29

because like inside of Meta they had

3:31

this leaderboard they got rid of it

3:32

after an article came out and it looked

3:34

amaz

3:36

like just just like closed it down. that

3:39

people are still token maxing by the way

3:40

because there's this this thinking that

3:41

it might have gone but you know we're

3:43

engineers and don't forget these are

3:44

highp paying jobs right that like you

3:45

don't really want to lose a job over

3:47

something stupid as like you didn't have

3:48

INF token count and that's how it feels

3:50

but inside Salesforce there's a target

3:52

of minimum spend per month like I think

3:55

it's like $175 between the things so

3:58

like people are like again you kind of

4:00

like you know beginning of the month

4:01

like just token max to get there so it's

4:03

it's it's weird and it started as a joke

4:05

earlier like a few months ago token

4:06

maxing was really just people like going

4:08

crazy and enjoying this thing and

4:09

building cool stuff. But it's kind of

4:11

turned into in a lot of companies I

4:13

think it's just a culturally weird

4:15

thing. So it's a weird time to be in cuz

4:18

I remember lines of code used to be when

4:19

when early uh developer productivity

4:22

tools came out like velocity and

4:24

pluralite flow. They kind of measured

4:26

lines of code and and number of QPRs and

4:29

we know that was stupid and people kind

4:30

of optimized for that at companies that

4:32

did it. But it's it's almost like what

4:34

now it's the top running companies like

4:36

Meta and Microsoft who are incentivizing

4:38

people just to do just stupid stuff

4:41

honestly.

4:42

>> Yeah, those are wild stories. And one of

4:44

the things you're clapping for that

4:48

deserves another full conversation. Uh

4:51

one of the things I like about talking

4:52

with you and subscribing to your

4:53

newsletter is that you basically kind of

4:55

anonymize all these stories from from

4:57

real incidents and real examples. Um why

5:02

is it that uh is is it still worth it

5:05

right with all the flaws uh you know

5:08

when you have good heart's law like what

5:10

whatever gets measured gets uh sort of

5:12

abused with all the flaws is it still

5:14

worth it you know is is is AI basically

5:18

still making us faster overall like the

5:20

cost of token maxing is still with all

5:23

these like really ridiculous examples is

5:25

it still net worth it

5:27

>> yeah so don't forget like the reason

5:28

token maxing is probably a thing is like

5:32

let's just go back to six months ago

5:34

where

5:37

I I I was at a I was at a CTO like

5:40

dinner conference whatever like a bunch

5:42

of CTO's gather CTO level people this

5:44

this was in Amsterdam and we had like

5:46

like a bunch of people and there we were

5:48

talking and and one of the CTO's like

5:50

the the the Amazon of the Netherlands

5:53

there there's a e-commerce company was

5:55

saying like hey like everyone like I

5:57

have a problem like engineers on my team

5:59

are really skeptical of AI and they're

6:01

not really using it. The AI tools, don't

6:03

forget this was before Opus 4.5 and

6:05

those models were were out. They were

6:06

not as as productive. We had uh we we

6:09

already had a cursor and and the like

6:11

and they subscribed. They're like

6:12

they're just not using it that much on

6:14

existing code bases, right? And and next

6:17

to them uh the head of the Dutch

6:21

National Bank said like, "Oh, we don't

6:22

have that problem. Our engineers are

6:24

using it because our our mission is to

6:25

regulate this thing. So, we need to

6:26

understand it." And they're kind of

6:27

motivated. And there was this time where

6:29

experienced engineers were kind of

6:31

holding off because if you had an

6:32

existing codebase and use AI cursor

6:35

whatever on it was mildly useful if that

6:40

even and these engineers were like why

6:42

should I use a tool if it doesn't help

6:44

me refactor it doesn't find the bug it

6:46

doesn't do what I need to do and

6:47

leadership saw they're not really using

6:49

it and they kept hearing you know the

6:51

likes of Antrophic for example was

6:53

already saying how they're writing a lot

6:54

of their code with with cloth code uh

6:57

and it just keeps increasing and

6:58

andropics, you know, like revenue is

7:00

going up like this. So those leaders are

7:02

kind of they might be confusing

7:04

correlation and and and you know, like

7:06

which one comes first, but they're like,

7:08

well, we should be using it more because

7:11

probably good things will happen and

7:12

thus bad things will happen if we don't

7:14

use it. So the whole targeting and

7:17

measuring things, it actually came from

7:19

leadership wanting, we want our

7:21

engineers to use faking AI. I don't care

7:23

what it is. And it it was a bit of a

7:24

push like we know this is bad but it's

7:26

it's better than them using it. Best

7:28

example is Coinbase where uh Brian

7:31

Armstrong the CEO just like fired an

7:34

engineer or he sent an email saying

7:36

everyone like needs to get on board and

7:37

use AI tools and whoever doesn't use it

7:39

in a week I'll have a conversation with

7:41

them and then I think a week later on

7:42

Saturday he fired an engineer and you

7:44

know like this again high paying job

7:46

like we're talking base salary like

7:47

three 400k,000

7:49

per per year uh and then both equity and

7:52

everything on top of it like they got

7:53

the message everyone just started to

7:54

just you know like use it and you back

7:57

to your question. So on on one there

7:59

there's a push and look I feel it's a

8:02

little bit like this is going to be

8:03

controversial but have you ever wor

8:06

wonder wondered why big tech loves to do

8:08

lead code style interviews algorithmical

8:10

interviews which have nothing to do with

8:12

the job and and we know it's the case

8:14

and there's a lot of criticism for this

8:16

and they've been doing this since since

8:18

like 20 years but here's the thing it

8:20

selects for a specific type of person.

8:22

It selects for the person who's smart

8:24

and willing to put up with absolute

8:26

[ __ ] to get the job.

8:29

And this person, you know, they will

8:31

study two months pre AI, two months or

8:34

three months of lead code, which again

8:36

makes no sense on the job, but you do

8:37

it. You get in there and this person

8:39

will be putting to put up with [ __ ]

8:41

that makes absolute no sense to keep the

8:43

job. So token maxing happens at large

8:47

companies and people are putting up with

8:49

this BS. And look, a lot of them are

8:50

smart and they will make the most of it.

8:52

some of them will build cool stuff. Um

8:55

it's it's the reality I think of big

8:57

tech. So we're in this weird place where

8:58

big tech is a bit weirder than startups

9:00

where you know no one cares about

9:01

tokenaxing. They care about like just

9:02

building stuff and you know use whatever

9:04

makes sense. Don't people will care

9:06

about the cost.

9:07

>> Yeah.

9:08

>> But going back to your question like

9:09

like you know like is is it making us

9:10

productive as as a whole like

9:12

individually it's it certainly is and as

9:14

teams we're kind of like a bit question

9:16

mark because we should be moving faster

9:17

and there are a few companies that do.

9:19

Entrophic is a good example, but a bunch

9:20

of companies are like not it's it's it

9:22

seems it's hard to retrofit all this AI

9:24

into like the way we have been working.

9:26

>> Yeah. Uh one of my favorite studies from

9:29

last year was the meter study where they

9:31

uh did a blind test of uh people and

9:36

their expectations of productivity,

9:38

right? And basically the the end result

9:40

was they felt 20% more productive, but

9:43

their demonstrated results was actually

9:45

they were 20% less productive on

9:47

average. Yes. But that that study was

9:49

very interesting because they

9:50

>> it was very small sample size.

9:51

>> It was 30 people and there was one

9:53

outlier uh who actually was way more

9:56

>> Anthony we we interviewed him on the

9:57

pod. Yeah.

9:58

>> Yeah. Yeah. So he was the one productive

10:00

AI engineer

10:03

but anyway so uh actually my theory is

10:05

that uh something that I've seen on my

10:07

team is that I've been enabling coding

10:09

agents for the rest of my team who are

10:10

non techchnical right and uh you as the

10:13

engineer may not be more much that much

10:15

more productive because and you can be

10:17

more productive if you uh attend AIE but

10:20

uh if you actually enable your

10:22

non-coding uh your your non-coding co

10:25

collaborators to code actually they are

10:27

more productive because they don't have

10:28

to wait for you right and that's that

10:30

like unlock of like oh suddenly you have

10:32

serverless developers basically uh and I

10:35

think I think that's that organizational

10:37

coding thing is different than studying

10:39

pull request level productivity for the

10:41

individual developer

10:42

>> yeah and and the thing that still I

10:44

still remember to this date I I talked

10:46

with Simon Willis I think in 2024 so two

10:49

years after Chad GPT came out and he was

10:52

Simon Wilson top commenter on hacker

10:54

news or he's he's

10:55

>> that's his that's not his title man top

10:57

commenter on hacker What the [ __ ]

11:00

>> No, he's

11:01

>> creative, Django, top blogger. Yeah. Uh,

11:04

prompt injections. Uh, yeah.

11:06

>> Yeah. He's actually not talk. I'm sure

11:07

he's the most submitted block cuz he

11:08

blocks so much like like and he's

11:11

>> but he told me back then he said like

11:13

this thing AI is is just so hard to to

11:17

get good at. He's like there's no

11:18

manual. And he's like, I've been doing

11:20

it back then for two years and I'm still

11:22

I'm still figuring out what works and

11:24

what doesn't. I keep changing my

11:25

workflows. And I think that's something

11:27

that is a bit hard for us. Two things

11:30

about AI that for any of us engineers is

11:32

hard to understand. One is it just takes

11:34

a long time to get good at it and you

11:36

need to keep doing it. And the second

11:38

thing is understanding the theory will

11:41

not make you better at using the tools

11:43

which is an absolute mind [ __ ] honestly

11:46

because we're so used to you know you

11:48

understand how the compiler works, how

11:49

assembly works. Okay, you will now be

11:51

more efficient if you want to write

11:52

low-level code because you know how it

11:54

works. But what with these things I mean

11:55

you can of course it's helpful to

11:57

understand how how the the architecture

12:00

underlying works attention the different

12:03

the the different probability sets etc

12:06

etc but it will not help you get a sense

12:08

for how you can use it and then once you

12:10

figure out how you can be more

12:12

productive if you're if you're inside of

12:13

a team again it kind of breaks and you

12:15

have to relearn again but but the more

12:18

effort you put into it it like it's

12:20

clear that it's it's working it's

12:21

helpful and I think it it's the teams

12:23

I'm seeing and getting more value out of

12:25

it. Low ego, open to learning, open to

12:28

leaving your priors behind. The word

12:30

priors I have not used forever and I

12:34

feel we're in this stage where like just

12:35

just leave your priors behind. Just have

12:37

an open mind like don't leave your

12:38

experience behind but you know be open

12:41

to it.

12:42

>> Yeah. Zooming out a little bit. How is

12:44

the role of the software engineer

12:45

changing?

12:48

>> I think it's always this was always

12:51

coming but AI is just just speeding it

12:53

up. uh even before AI a few

12:57

it's interesting I see like startups in

12:59

many ways venture funded startups are

13:01

kind of front running what the industry

13:02

will be catching up because venture

13:04

funed startups are about fast growth um

13:06

doing

13:08

mo moving fast with smaller teams

13:10

because smaller teams mean smaller comps

13:12

even preai so a lot a lot of these

13:14

venture funed startups start to expect a

13:16

lot wider range of roles from engineers

13:19

for example devops as a whole inside VC

13:22

funded companies from the mid210s every

13:25

engineer was kind of like responsible

13:27

for the code they deployed but like more

13:28

traditional companies they had more

13:29

money more sorry more less pressure they

13:31

kind of have dedicated devops teams and

13:33

some of those things so in in the

13:36

industry like the software engineer is

13:37

now becoming like the kind of the tester

13:39

role has collapsed into software

13:41

engineer we most companies don't have

13:43

dedicated testers very very few do

13:45

devops collapse into here uh and now

13:48

we're starting to have the product role

13:49

also starting to come so a lot of

13:51

companies even like in 2022 before AI

13:53

starts to hire for product engineers

13:55

that's happening faster and I think the

13:58

the last push that AI is doing is even

13:59

for early career engineers there's a lot

14:02

more seniority expected or or senior

14:04

like things planning about things

14:06

knowing about the business so I I I

14:09

think the role is expectations are are

14:11

higher teams are also getting smaller

14:13

everywhere I talked with someone at John

14:15

Deere 200 person uh 200 year old company

14:18

sorry uh you know like they do tractors

14:20

and and all all that stuff and and

14:23

inside of that company, one of their

14:25

their VP of engineering was telling me

14:27

how they're actually seeing that their

14:28

two pizza teams are now just one pizza

14:30

teams inside of that company. It's the

14:32

reality partially because of these

14:33

tools.

14:34

>> So, my joke used to be I am a one pizza

14:36

team because I eat a lot of pizza, but

14:37

uh depends how much pizza you eat.

14:39

>> Uh there's so I'm sorry to interrupt. I

14:42

don't know if I cut you off in some

14:44

critical point. Uh there's a comment

14:45

saying I've heard it twice even among

14:47

this audience where a lot of people are

14:50

saying that oh uh you're no longer an

14:51

engineer everyone's an engineering

14:52

manager now and you've been an

14:54

engineering manager and I wonder if you

14:56

agree with that or if you have a

14:58

different take you know because

14:59

basically you're the the the common

15:01

analogy is that you're no longer a

15:02

software engineer you're just managing

15:04

engineering agents right yeah if you've

15:07

been a manager before that is an

15:08

absolute [ __ ]

15:11

so so here here's the thing the like

15:14

Yes, you are a manager without all the

15:17

things that no one wants to become a

15:18

manager for the the when you become an

15:20

engineering manager. Hands up if you are

15:22

or have been an engineering manager,

15:24

right? Hands up if you actually if

15:25

you've not been and you want to be one

15:26

>> about 15 20%.

15:29

>> All right, you come and talk to me

15:30

afterwards. I I'll tell there's a hand

15:33

up there. I'll talk you out of it. So,

15:37

so what you think you become an

15:39

engineering manager to like help

15:40

people's career, maybe have higher

15:42

salary, higher impact, all you know

15:44

there can be a lot of dynamics but the

15:45

reality is is is you you become more

15:48

removed from the product and you have to

15:50

deal with people problems and the thing

15:52

with with agents is you don't have to

15:54

deal with people drama, people problems,

15:56

conflict between your team. I mean

15:58

unless the next generation of agents

16:00

starts to fight with each other. I think

16:01

that'll be something but you actually

16:03

you you do have to orchestrate but it's

16:05

more like a tech lead role or or or

16:07

experienced engineer where where you're

16:09

like mentoring uh mentoring engineers

16:11

but you don't have the people

16:12

management. You don't need to worry

16:13

about the personal problems. So it's

16:15

actually a lot more kind of empowering.

16:17

And I was talking with uh the podcast

16:19

was was just out yesterday with with DHH

16:22

uh creator of Ruby on Rails who said,

16:24

you know, people told him like, okay,

16:26

it's it's like managing things and he's

16:27

not excited about managing agents, but

16:29

it feels it's more like a mech suit

16:31

where you have like you can do seven

16:32

things at once, you can do a lot faster

16:34

and you're in control and that's more

16:36

what it feels like. So there's

16:37

orchestration, yes, but it's very

16:39

different to management. And also the

16:41

the really really bad thing or honestly

16:43

shitty thing about management if if you

16:44

make it into management which makes it

16:45

hard also rewarding later when you you

16:48

tell yourself at least this thing is you

16:51

start a project with all these people

16:52

under you you know congratulations

16:54

you've got 10 people wonderful and you

16:56

start a project and in 6 months you will

16:58

see some results of the decision that

17:00

you made with agents it's just so much

17:02

faster so the the feedback loop is

17:04

faster so I I think it's it's not much

17:07

of it except for the orchestration and

17:08

and and for that everyone's going to

17:09

have their own flavor. Some people will

17:11

will have the tendency to like run

17:13

multiple agents and they're good at this

17:14

or we good at it. Some people just do

17:15

like two agents. Michelle Hashimoto, I

17:17

interviewed him. He has two agents. He

17:19

always has one agent running. No, he has

17:22

one background agent that he doesn't.

17:23

That's it. He's like two is enough for

17:24

me. Great.

17:25

>> Yeah. Yeah. Uh we're figuring out the

17:27

patterns. Um uh I want to hit you on

17:31

large tech infra.

17:34

uh this is something that I think both

17:35

of us are very excited by by uh good

17:38

infra which is a very niche uh interest

17:41

what are you seeing

17:43

>> it's wild to see how much of the so I

17:47

said that from externally a lot of

17:49

companies a lot of big tech companies

17:50

especially the ones are spending a bunch

17:52

on AI and have platforms and all that

17:54

you're not seeing too much like more

17:56

come out like Uber is a good example I'm

17:58

not seeing too many more features come

17:59

out of Uber or new products launcher and

18:01

they're like but what's going on they

18:03

are really investing in AI but when you

18:05

look inside there's a whole lot of buzz

18:07

they are rebuilding their complete IM

18:09

infra you know they're and I'm not

18:11

talking about they're buying cursor or

18:13

or cloud code or all that they're doing

18:15

that as well but they're completely

18:17

they're building their own own custom

18:18

background coding agents that is

18:20

integrated into their monor repo they

18:22

are are having uh their own MCP gateway

18:26

that is is now integrated into service

18:28

discovery their on call tooling is being

18:31

retoled their internal code review

18:32

system is like like categorizing based

18:35

on risk. They are like and Uber is one

18:38

example but all everyone else Airbnb

18:41

intercom meta Microsoft even midsize

18:44

companies are just building so much

18:46

internal improp and I was asking to

18:48

myself like why on one end this feels

18:51

like such a waste but when I worked at

18:52

Uber for four years I realized they

18:54

spend so much on on internal platform

18:56

there's two reasons one is honestly it's

18:58

a it's a lowrisk way to get good with AI

19:03

uh to be hands-on and these companies

19:04

want to be hands-on but maybe you

19:06

shouldn't start with shipping AI

19:07

features no one wants into your

19:08

codebase. Second of all, because these

19:12

these companies have such so much code

19:14

that never fit in a context window, by

19:16

building custom solutions and just basic

19:18

basic wagons, that kind of stuff, they

19:20

will have better results than

19:21

off-the-shelf vendors. So, they already

19:23

have a win. And number three, honestly,

19:25

is anything that has AI in it gets

19:27

funded. So, there's this joke of if

19:29

you're in the developer platform team

19:30

and you're asking for more headcount,

19:31

like good luck with that. Oh, developer

19:33

platform. Oh, but say that you want to

19:36

get two extra headc count for agent

19:38

experience. Done. H. So, so there's that

19:41

part as well. But, but all

19:42

>> agent experience is just a CLI

19:45

>> pretty much. But all these come inside

19:47

there's so much buzz and so much work.

19:49

Everyone's building their own custom

19:50

system. So, I'm kind of wondering how

19:52

long this will take, but I think for

19:53

next year this is going to happen. So,

19:54

if you either have friends or if you're

19:56

work if you're working at a company,

19:57

you'll see. But talk with with friends

19:58

at other large companies and you will

20:00

probably see you are all building the

20:01

same thing. If you're at a large company

20:02

and you're not already building an MCP

20:04

gateway, what are you even doing?

20:07

>> Yeah. Um, actually a lot of these topics

20:10

are exactly the things I cured for

20:12

tomorrow. Uh, it's just fantastic to

20:14

have you as the closing keynote for

20:15

today because uh it's it's like a

20:17

appetizer for tomorrow. We have talks

20:19

about MCP gateway and all these sort of

20:22

AI architecture and infra things and I

20:24

do think like uh infra like

20:27

taking AI infra seriously as a company

20:30

is uh very mis not that well un

20:33

understood and right now you just kind

20:35

of learn by example from people because

20:36

there's not really like a textbook or

20:38

anything like about it. So the way I

20:40

think about this because again from if

20:42

you just kind of step out and we love to

20:44

criticize big tech of how they're

20:45

wasting money here and there and by the

20:47

way we love to criticize Google and I'm

20:49

kind of thinking to myself like hang on

20:50

what if Google ex actually executed well

20:53

like do we want that and you know they

20:55

would kill all the startups but but what

20:57

they're doing makes makes sense and

20:59

Shopify is an example where I'm like huh

21:01

I'm starting to get why it makes sense

21:03

to do all this stuff. So Shopify in 2021

21:06

they were the first company to have

21:07

access to a GitHub copilot. What

21:10

happened is the the head of engineering

21:12

fartoir heard about GitHub copilot being

21:15

developed internally inside of GitHub

21:17

and he pinged Thomas Dunca the CEO of

21:19

GitHub at the time and said hey Thomas I

21:21

heard you guys are doing C-pilot and

21:22

he's like yeah we are it's internal.

21:24

He's like I I'd like to get access to

21:26

it. He's like yeah but it's not for

21:27

sale. He's like no no no you don't

21:29

understand. I I didn't ask if it's for

21:30

sale. we would like to roll it out to

21:32

all of Shopify and in return we will

21:34

give you feedback for 3,000 people for

21:36

you know as honest feedback all the time

21:38

and so they got it a year before it was

21:40

out anywhere and they incurred a lot of

21:42

churn. It wasn't that great initially

21:44

and and they went through all of this

21:45

stuff and then Shopify was the first

21:47

company to on board to like a bunch of

21:49

other tools and they gave unlimited

21:50

budget and they're spending so much time

21:53

ironing out bugs. But the reason they're

21:55

doing it, this is what like made me

21:57

click is they are trading off churn and

22:00

expense and spending a lot more money to

22:03

be at the forefront of this. They are a

22:06

few months ahead or six months ahead of

22:08

their competition and for them it's

22:10

worth it. It's not worth it for anyone

22:11

else, right? If you're if you're at a

22:12

company where your business is like

22:13

something something physical and you

22:15

don't care like yeah just just wait out

22:16

it it'll come. But for a lot of us in

22:18

the tech industry this turn is worth it.

22:20

Plus what Farhan told me is like because

22:23

he actually told me he's kind of worried

22:24

about the cost now. But he was like look

22:26

like it's still worth it because if it

22:28

would look silly if I said you cannot

22:30

have these tools how would I hire the

22:32

best?

22:33

>> So it's it's innovation recruitment and

22:35

it kind of makes sense when you think

22:37

about it. And the weird thing everyone

22:38

is doing it at the same time. So it

22:39

looks silly but it it's rational.

22:41

>> Uh my next podcast is with Mikuel Parkin

22:44

the CTO of Shopify and uh the sheer

22:46

amount of machine learning that they do

22:47

and infra that they set up for their

22:49

customers makes me want to be a

22:50

customer. You know that's that's like

22:52

the best uh endorsement I can give. Um

22:54

I'm going to get meta a little bit and

22:56

talk about pragmatic engineer. Uh you

22:57

and I kind of startedish in COVID. Uh

23:01

you just left Uber. Uh how has it been

23:03

growing? What what are the main stats

23:05

that you're proud of that uh you'd like

23:07

to share with the world? Yeah. So I I

23:09

started pragmatic engineer I I I a joke

23:12

that if it wasn't for co I I would

23:13

probably never have started the this

23:15

thing because what happened with co is

23:16

uh Uber had layoffs and most of the tech

23:19

industry was doing great but Uber was

23:20

not and my team uh was hit by layoffs

23:23

and then we we had to disperse the

23:24

remaining people at other teams because

23:26

our mission no longer made sense and it

23:27

was just like a the morale was low my

23:30

morale was low so I was like let me take

23:31

a break. I wanted to write some books.

23:32

Swix was writing his book the the coding

23:34

career.

23:35

>> Yeah some of you have read it. I've met

23:37

some of you.

23:37

>> Yeah. and and that that's how we met

23:38

there and then uh my plan was to write a

23:40

book and then start start up some

23:42

startup something something platform

23:44

engineer control c control V from what

23:46

Uber was doing inside and that's

23:47

actually almost all Uber su Uber

23:50

startups it's it's amazing temporal is

23:52

is is from there

23:54

>> if I by the way if I did not start AI

23:56

engineer I would have started platform

23:58

engineer

23:58

>> that that would have been the industry

24:00

conference

24:01

>> yeah love it and then I start I started

24:04

the pragmatic engineer uh a year after I

24:06

left Uber It was just an experiment. Um,

24:09

I figured no one Substack was taking

24:11

off. No one was writing about software

24:13

engineering in-depth and I just acted

24:15

all confident saying pretended that I I

24:17

knew what I was doing. The first article

24:18

was about Uber's platform and program

24:20

split that no one had written about

24:22

publicly before and it's a it's a free

24:23

article. You can you can now check it

24:25

out. Uh, and it was like when you feel

24:28

product market fit, that's what I felt

24:29

almost immediately. The first week

24:31

before I published anything, just a

24:33

confident Twitter post, I had 100 people

24:36

pay upfront $100 for the whole year,

24:39

which I was like, whoa, I have published

24:40

anything. In six weeks, I was at a,000

24:43

people paying for this thing that didn't

24:45

exist before, which was my old Uber base

24:48

salary back back in Amsterdam. And it

24:50

just kept going up. So like I I figured

24:52

like when you find product market fit,

24:53

this is like outside of like there's

24:55

this rule like if you find product

24:56

market fit, just keep doing what you're

24:58

doing. So for me, I just kept writing

24:59

that one article. I got all these

25:01

interview requests, collaborations,

25:02

podcast. I just said no to all of them

25:04

because I knew the most important thing

25:06

was to do what makes it successful,

25:08

which is that one article. And later it

25:10

turned into two articles. And for two

25:12

years, this is all I did, just two

25:14

articles. And after two years, I looked

25:15

up and I was like, huh, like this is

25:18

actually working. People like doing it.

25:19

I like doing it. There's a future in

25:21

that. And that's when I decided I

25:23

actually want to turn this into a

25:24

business that I don't burn out because

25:26

for two years every vacation I went to I

25:28

was working 50 60 hours. I was always

25:30

thinking I was writing I I couldn't

25:32

really let go. So I started to grow the

25:34

team a little bit. Uh I I Ellen Bird the

25:37

first secondary researcher Ellen she's

25:39

ex x ex uh

25:40

>> she's here right?

25:41

>> Ellen's not here. Um Jessica is who who

25:44

just joined uh later.

25:46

>> Yeah.

25:46

>> And then uh so now it was two of us. Uh,

25:49

and I started a podcast year and a half

25:51

ago because I talked with so many

25:52

people. I figured it was a bit of a

25:54

shame to to not have it. So, the

25:57

primatic engineer became the number one

25:59

paid technology newsletter about four

26:01

months after starting. It stayed there

26:03

for three years. Now, semi analysis has

26:05

>> Dylan versus uh you guys. Um, yeah. No,

26:08

congrats on your success. Uh I think you

26:10

you're also a leading tech voice in

26:13

Europe which I think you're sort of

26:14

proudly sort of uh upholding that over

26:17

here which I would really wanted to

26:18

feature. Thank you for your support for

26:20

AIE. And uh everyone thank you Good.

26:22

Awesome.

26:24

Thanks, man.

More transcripts

Explore other videos transcribed with YouTLDR.

Get the TLDR of any YouTube video

Transcribe, summarize, and repurpose videos in 125+ languages — free, no signup required.

Try YouTLDR Free