Full Transcript

·YouTLDR

How I Interview AI Engineers (And the Projects That'd Impress Me)

17:40EnglishTranscribed Jul 26, 2026
0:00

All right, let me set the context of the

0:01

video. This video we're going to go

0:02

through six seven projects in applied AI

0:05

specifically related to agents that you

0:07

should probably build if you're trying

0:08

to target companies like, you know, I

0:10

don't know, like an emergent any India

0:12

based or US based Genie AI startup. For

0:14

the last one month we've done a bunch of

0:15

offline classes on applied AI on topics

0:17

like context engineering, writing

0:19

emails, building an agent from scratch.

0:21

And it's becoming very obvious

0:23

there are more companies right now in

0:24

this specific niche and almost all of

0:26

them are going to ask the standard set

0:27

of questions. So, in this video I'm

0:29

going to take you through how I am

0:30

interviewing someone here at Super 30

0:32

for a recent company that's raised $15

0:33

million. What are the kind of topics

0:35

that people are learning over here and

0:36

around seven eight projects that I feel

0:37

if you build from scratch and really

0:39

well, then intuitively it'll be easier

0:41

for you to crack these interviews.

0:42

Without any further ado, let's get right

0:43

into it. Here are five high-level topics

0:46

that we've covered in the last one month

0:47

at Super 30. Whichever one gets the most

0:49

up votes in the comments is the one that

0:51

we'll release on the channel as well.

0:52

Specifically, these include memory as in

0:54

how does your agent remember who you

0:55

are, the things that it has done in the

0:57

past and how does it learn from it.

0:58

Firecracker, which is a virtualization

1:00

technology that Lambda uses as well. If

1:02

you've ever used AWS Lambda for spinning

1:04

up sandboxes, if you've heard of E2B,

1:06

E2B also uses Firecracker under the

1:07

hood. And this class covers how you can

1:09

create your own Firecracker clusters

1:10

from scratch. Context engineering, this

1:12

one's taken by me. It specifically

1:13

covers what techniques do you use as

1:15

your context grows

1:17

to make sure you don't get context rot,

1:18

to make sure you're not spending a lot

1:20

of money doing a specific task using an

1:22

agent and to make sure you're able to

1:23

manage the context of the agent. If the

1:25

model that you're using is something

1:26

like Opus 4.8 that has a context window

1:28

of 1 million, what do you do when your

1:29

long running tasks reaches around 700

1:31

800k tokens? How do you summarize

1:32

context? How do you compact context?

1:34

Evals and RL environments, how do you

1:36

write tests or evals for your agents?

1:38

How do you make sure they're actually

1:39

improving over time? How do you know if

1:41

model one is better than model two? How

1:42

do you build something like sweep ends

1:43

from scratch? And what do you do if you

1:45

join a data lab where your job is to

1:47

create RL environments or data that is

1:49

used to train state-of-the-art models.

1:51

And lastly, a video on cloud agents. How

1:52

does something like Devin work or how

1:54

does remote control in cloud go to work?

1:56

Slightly more of a full stack question.

1:58

This is a set of classes on the last one

1:59

month at Super 30. Let us know in the

2:01

comments which one you'd like on

2:02

YouTube. With that, let's move to the

2:04

interview process. Um, recently a

2:06

company's hiring through us. They've

2:08

raised $15 million recently. They're

2:09

building some sort of coding agent. And

2:11

before we refer someone there, I wanted

2:12

to interview them personally. So, this

2:14

is the set of questions that I have

2:15

prepared for them. I have to take this

2:16

interview today. Um, and the question is

2:18

pretty simple. If you're building

2:19

something like Devin where you have a

2:20

screen like this, a user comes, they

2:22

they're asked to, you know, fix a

2:23

specific issue on their GitHub

2:24

repository, and the agent starts over

2:26

here. You're able to see the timeline of

2:27

everything that happened, including the

2:29

agent starting a terminal, the agent

2:31

editing a file, the agent building some

2:33

process, the agent reading files,

2:34

writing files, etc.

2:36

Considering you have to build a system

2:37

like this, here are a few interview

2:38

questions that I thought of. Number one,

2:39

um, what all services do you need, which

2:41

is like the simplest question here. What

2:42

would the architecture look like? How

2:44

many backends do you need? How many

2:45

frontends do you need? What does the

2:46

infra layer look like? How does the

2:47

infra layer scale up if you have 100

2:49

people at the same time, you know,

2:51

talking to the agent? Um, probably need

2:52

100 sandboxes running, um, where the

2:54

agent is running, where the task is

2:56

being fixed, where the final commit to

2:57

GitHub happens from. How do you spin

2:58

them up? How do you spin them down?

2:59

That's the only infra-heavy challenge

3:01

over here. Other than that, you're

3:02

probably going to depend on an OpenAI

3:04

for the LLM calls. You don't have to

3:05

scale that up. Frontend and backend is

3:07

going to be fairly simple. So, the real

3:09

crux of what we're trying to get at is

3:10

does the person understand, uh, for a

3:12

project like this, you need sandboxes,

3:14

and how do you scale them up and scale

3:15

them down? Number two, what happens if

3:17

your server that's running your agent

3:18

crashes? This is an interesting one

3:19

because,

3:21

um,

3:21

not usually in full-stack apps do you

3:23

see long-running tasks, but

3:25

agents, and specifically, you know,

3:26

coding agents that are doing a very

3:27

heavy task can run for 30 minutes to

3:30

hours, a really long time. Um, so, what

3:31

happens in case your infrastructure does

3:33

go down? How do you resume a

3:34

conversation, um, or resume something

3:36

that the agent was doing from a specific

3:38

checkpoint?

3:39

There are many answers here, you know,

3:40

you basically need durability across

3:42

sessions. If a sandbox where the agent

3:44

is running, or if the backend where the

3:46

agent is running crashes, um,

3:47

whenever it restarts, or, you know, some

3:49

other part of your infrastructure needs

3:51

to start to pick up from there uh,

3:53

and continue the conversation with the

3:54

agent. Good thing about agents is

3:56

there's a lot of back and forth that

3:57

happens between

3:58

your agent and the LLM and hence if at

4:00

some point the agent does crash, if

4:03

you've backed up the message history

4:04

between the agent and the LLM,

4:06

you can always resume the conversation

4:07

from a different part of your

4:08

infrastructure. Um

4:10

Good read at this point is Temporal. If

4:11

you've heard of what Temporal does and

4:13

you know how they provide durability for

4:15

use cases like these.

4:16

And if you talk about that during the

4:17

interview, it'll probably be a good

4:18

sign. Number three, how do you do

4:20

context management? There are 100

4:21

strategies over here. Um

4:23

I think most newer companies would not

4:25

really reach a scale where they need

4:27

context management or context

4:28

engineering at all. But nonetheless,

4:30

every new company like Devin that is

4:32

coming or you know if you're building a

4:33

coding agent yourself,

4:35

the real thing to optimize for is what

4:36

happens when the conversation becomes

4:37

really long. For a shorter conversation,

4:39

most models perform fairly well. It is

4:40

only when you're giving it a very big

4:42

feature or a very big task or an issue

4:44

that's very gnarly is when the

4:45

conversation is going to be really long.

4:46

It's going to be a long-running task and

4:48

hence compacting the conversation from

4:49

time to time or summarizing it. If

4:51

compaction isn't working, um are like

4:52

standard techniques that you need to

4:53

know of. Here is a video by the folks at

4:55

Manifold where they talk through what

4:56

they do and what Claude does for context

4:58

engineering. Number four, how do you

4:59

evaluate your agents? This is probably

5:01

something you've never thought of in a

5:02

side project, right? If you're building

5:03

something like a Lovable, um

5:05

you never really think of is your model

5:06

getting better? Is your agent getting

5:08

better over time? How do you know if

5:09

your agent is actually improving? Before

5:10

you do a new release of an agent or if

5:12

you introduce a new model, how do you

5:14

objectively, not subjectively,

5:16

objectively know that the output has

5:17

gotten better? There are a bunch of ways

5:18

to do it. Um slightly harder for a use

5:20

case like Lovable, but for you know

5:22

other use cases something like a Devin,

5:24

it's not too hard. There are standard

5:25

ways that companies use right now to

5:27

write evals. If you've seen SweBench or

5:29

the newer benchmarks that have come now,

5:31

they have a deterministic way of telling

5:32

if a model or a model harness is

5:34

performing good or bad, you know.

5:36

And is it improving over time? I think

5:37

every company in some form or fashion

5:39

eventually is going to write their own

5:40

evals on their data to see if, you know,

5:43

as new model releases are coming, are

5:45

their engineers becoming more efficient?

5:47

Whichever parts their app AI is using,

5:49

um is it getting better over time or

5:50

not? There are again a bunch of

5:51

techniques over here. I urge you to go

5:53

through the paper of SweepBench or, you

5:54

know, look at the TerminalBench and what

5:55

they do, how do they work? I think this

5:56

is like a very niche skill right now. Um

5:59

but eventually this is like what a lot

6:00

of us will be doing, um writing a lot of

6:02

human tests RL environments to make sure

6:04

our agents are getting better over time

6:06

and if not, improving them. And lastly,

6:07

what metrics do you observe as your

6:09

application grows? So, the thing is we

6:11

already had observability before AI,

6:14

right? where we had platforms like New

6:15

Relic, Data Dog, um Prometheus, Grafana,

6:18

um various stacks to make sure your

6:20

application is running. Um LLMs are

6:22

slightly different that way. There There

6:23

are different set of challenges if

6:24

you're building agents. How do you

6:26

observe them? Um Specifically, things

6:28

like is the agent failing at some point?

6:30

If yes, what went wrong? Um Are your

6:31

users happy or not happy with the final

6:33

output? Be able to arrive at the final

6:34

output? Seeing the actual reasoning

6:36

traces of what led to the model

6:38

actually, you know, giving a specific

6:40

output. And now there are like a bunch

6:41

of companies that solve specifically for

6:43

LLM observability and all the older

6:45

companies like Data Dog are also coming

6:46

into this market. One specific company

6:48

that solves observability for agents or

6:50

LLMs is

6:52

And I feel companies that have hit scale

6:54

need to worry about this, you know, a

6:55

little too much. Just like a company, a

6:57

web two company that hit scale, um at

6:59

some point needs to worry about

7:00

observability and monitoring and making

7:02

sure alarm bells ring if, uh

7:04

you know, the back end is down.

7:05

Something similar over here. If your

7:06

agent is suddenly not performing if a

7:07

specific, uh infra provider is down for

7:10

whatever reason, you know, lack of GPU

7:11

right now. You should auto switch to a

7:13

different one. You should know what went

7:14

wrong. you should have visibility as to,

7:16

you know, what caused the agent to fail.

7:18

That's like a high level list of

7:20

interview questions that I had for

7:21

someone who's specifically interviewing

7:22

for a company like Devin. Irrespective

7:24

though, I think a lot of these questions

7:25

can be applied to other companies as

7:27

well. Any company that's using AI or,

7:29

you know, building some sort of an

7:30

agent.

7:31

These are like good set of interview

7:32

questions for users because I think

7:34

today everyone understands what an agent

7:35

is, what a model is, how do they work

7:36

together. The real question is, you

7:37

know, you understand how can you improve

7:39

the model, how can you improve the

7:40

agent, how can you observe the agent? If

7:41

at all, you need every agent to have its

7:43

own sandbox, how do you scale that? With

7:45

that, I'm going to move to a bunch of

7:46

projects. Some of these projects we're

7:48

building in Super Trendy as well. Number

7:49

one, simplest one, terminal agents. Um

7:51

and wait, building a terminal agent is

7:53

very easy. If you've seen the code base

7:54

of Pi, uh

7:56

which is this thing right here, it's

7:58

very readable. If you understand

7:59

TypeScript, you can read this code base.

8:01

It's probably less than, you know, 2,000

8:02

lines of code all in all. Um

8:04

So, building one is easy. Making sure

8:07

your terminal agent is performing at par

8:09

with Claude code uh using the same model

8:12

is the harder bit. So, what I'd really

8:13

urge you to do, and you know, I urge

8:14

everyone here to do as well, is not just

8:16

build this because that's the easy part.

8:18

Actually, benchmark it against other

8:20

agents and see if it's performing better

8:21

or worse. And if it is performing worse,

8:23

uh which it probably will, what

8:25

techniques can you use to rank up over

8:26

that? You know, there could be a bunch

8:27

of extra hacks that you could try for

8:29

now to just reach up the leaderboard of

8:31

Terminal Bench. You know, hardcode some

8:33

use cases, for example, you know, if

8:34

you're ever giving a Codeforces contest,

8:35

you a lot of times

8:36

hardcode a few use cases just so you

8:38

can, you know, pass more tests. You

8:40

know, it's not the final solution, but

8:41

you know, you you'll be able to pass

8:42

more tests if you add a bunch of edge

8:43

cases. Something similar. Just to see

8:45

what might be the right direction to

8:46

take to improve a model. A good example

8:48

of this is Pi by default does not have

8:50

sub-agent orchestration. Claude code

8:51

does. Uh

8:53

Codex does. Yet, Pi performs, you know,

8:54

almost equally well uh on the Terminal

8:57

Bench. That is one of many techniques

8:58

that might nudge the agent in the right

9:01

direction and make the agent arrive at

9:02

the final answer. And when you do that

9:04

is when you have a good terminal agent.

9:06

Project number two, Hermes Claude bot.

9:08

Both of them are again agents. Uh they

9:10

require a little more than building a

9:12

simple terminal agent. Uh they require a

9:14

lot of memory. They require a lot of

9:16

integrations to things like WhatsApp,

9:17

Telegram, Slack. Um

9:19

At some point, I would consider this to

9:20

be a little bit of a full-stack

9:22

challenge. Um it's actually just

9:23

connecting with a bunch of services and

9:25

letting the AI go free and take

9:26

autonomous decisions over time. The good

9:27

thing about most projects here is that

9:29

they're all open source. Um So, if you

9:31

do have AI, you can actually understand

9:32

their architecture fairly well. That's

9:33

how I've understood, you know, a bunch

9:35

of these architectures. It's harder to

9:36

dive into the code base of these now

9:39

because a lot of these are like by coded

9:40

and very verbose. But what's easier is

9:41

to make an LM summarize what the project

9:43

does and go from there. By is probably

9:45

the only open source agent that I was

9:46

actually able to read the code in one go

9:48

and understand.

9:49

Number three, Slack plus AI. This is a

9:51

new set of companies that are coming.

9:53

I don't know if you guys know Jack

9:54

Dorsey who's the, you know, ex-founder

9:56

of Twitter and now the founder of Block,

9:58

the famous company that laid off 40% of

10:00

its folks.

10:01

They recently released this project

10:02

called Buzz, again open source and then

10:03

you might have heard of Prompt QL. A

10:06

batchmate of mine works there now which

10:07

is why I realized, you know, they exist

10:08

as a company and it's a little bit of a

10:10

speculative use case but one that might

10:13

become very big. It's basically a

10:14

workspace, a more AI native workspace

10:17

similar to Slack. In Slack you only

10:19

have, you know, humans as members. Here

10:21

you can have agents as members.

10:23

You can tag agents and then they'll do

10:25

go their own thing. You can add, you

10:26

know, different set of agents. One could

10:27

be a front-end expertise agent, one

10:29

could be a essay writing agent so on and

10:30

so forth. You can tag them, they'll do

10:31

their thing. You can have a more fluid

10:33

conversation on an interface like Slack

10:35

compared to, you know, orchestrating

10:36

agents through a terminal. The code base

10:38

of Buzz is in Rust. So if you're

10:41

interested in Rust at all,

10:42

this is like simple Rust that you can

10:43

read. If you do want to contribute Rust

10:46

yet,

10:47

contribute to, you know, an AI project,

10:49

might be one of the good projects to

10:50

look at. Super set or T3 Code. I don't

10:52

know if you've seen these. T3 Code is by

10:54

Theo, another YouTuber.

10:55

It's basically a desktop app or a mobile

10:58

app where you can spawn agents, um,

11:01

coding agents. But the good thing about

11:02

this is you can actually select

11:04

different coding harnesses

11:07

for specific tasks. For example, in a

11:08

single UI I can ask Claude Code to solve

11:10

issue number one, Codex to solve issue

11:12

number two, Devin to solve issue number

11:13

three.

11:14

This is, I mean, primarily useful for if

11:16

you have subscriptions to all three and

11:18

you need a single interface to talk to

11:20

all of three of them. The interesting

11:21

thing there is how the architecture of

11:23

Super set is very different from the

11:24

architecture of T3 Code. If you use

11:26

Super set, it actually spawns the Devin

11:28

CLI, the Claude Code CLI, you know, in

11:31

various tabs and you sort of use them

11:32

versus T3 code

11:34

actually starts the underlying agent um

11:36

and hacks into the agent and you know,

11:38

the messages that you see are all T3 UI

11:41

coded compared to using the actual

11:43

terminal output from Claude code. There

11:45

are pros and cons to both of these

11:46

approaches. Basically, they have very

11:47

different architectures. The reason I

11:48

know this is because the next video is

11:49

going to be building something like this

11:50

from scratch. And the T3 code

11:51

architecture is slightly better than the

11:53

other one for many reasons, you know,

11:54

I'll get to that in the next video.

11:56

Both of them are open source. Super set

11:58

is a YC backed company. T3.code is doing

11:59

really well. I think it's growing really

12:00

well in terms of number of users.

12:02

Primarily again, built on top of coding

12:04

agents. T3.code will still teach you a

12:06

few things on, you know,

12:07

agents and what are the kind of messages

12:09

that Claude code provides you or the

12:11

Claude code agent, you know, what is the

12:13

kind of back and forth that happens

12:14

between the Claude code agent and the

12:16

LLM.

12:17

That is something you can learn if

12:18

you're building it the T3 code way. If

12:19

you're building it the Super set way,

12:20

it's mostly a full stack challenge, not

12:21

really a, you know, an agent or any AI

12:23

challenge at all. This is a project I

12:24

saw yesterday. I don't think this will

12:26

ever pick up steam. I think this is like

12:27

not the right way to learn, at least for

12:29

me. Um what it does is, if you want to

12:30

learn anything, for example, you know,

12:31

full stack, you can ask it to create a

12:33

learning track for you.

12:35

And then the nice thing that I mean, the

12:38

thing that I like generally is

12:39

generative UIs, it basically creates a

12:42

course like this. For example, I asked

12:43

it yesterday, you know, I not need to

12:44

learn attention is all you need or, you

12:45

know,

12:46

the attention or the transformer

12:47

architecture. This generated this course

12:49

for me, which, you know, looks very

12:50

close to if you in in 2020 wanted to

12:52

create a course like this, you would

12:53

have to go through a lot of effort to

12:55

gather all this data, write all of these

12:56

things down versus it was able to create

12:58

these slides in one go. So, it created a

12:59

structure like this. If I click on, you

13:01

know, one of these, I can actually look

13:04

at a bunch of slides. I can talk to the

13:07

agent over here and you know, it'll

13:08

guide me through it, which is and maybe

13:10

at some point we sort of work like this

13:12

or, you know, we learn like this. I

13:13

think right now it's a little I would

13:15

love to wait have a personalized

13:17

learning path for me. Problem is in I

13:19

personally feel for me, I need like a

13:21

lot of accountability, a human sort of

13:22

guiding me, which is why I tried this

13:24

yesterday and you know, did not really

13:25

go through with it.

13:26

Nonetheless, generative UI is generally

13:28

a cool you know, how do you make the

13:30

user build their own UI? Compared to you

13:31

know, static set of courses that exist

13:33

that you can buy. So, unsure if it's

13:35

like a great product or will it ever get

13:38

PMF?

13:39

Great technical thing to build.

13:40

Generally generative UI you know, you

13:42

can also build something like a AI based

13:44

slides plus mentimeter. I think that's

13:45

also like a nice project. Come over

13:46

there with a project an admin can come

13:49

to the platform, select a topic for

13:51

example, transformers and attention

13:53

followed by a 10 question or 20 question

13:54

quiz

13:55

that they can distribute to a bunch of

13:57

users and then you know, it's basically

13:58

a bunch of slides for a real class

14:01

followed by a bunch of MCQ, something

14:02

like that. Cool. Model router. This one

14:05

is interesting, new, slightly nuanced

14:07

and you know, probably there's no right

14:09

or right wrong approach for this right

14:10

now.

14:11

What is happening right now is in a lot

14:13

of companies, you know, the cloud costs

14:16

are really higher. I don't know if you

14:17

guys saw the Uber announcement, you

14:18

know, Uber turned down AI because people

14:19

were just burning through the credits.

14:21

And this is like a necessary evil, this

14:22

will happen when you

14:24

have a very big company and you give

14:25

engineers access to infinite credits.

14:27

They're probably not going to use it

14:29

very efficiently. Hence, it probably

14:31

will make sense at some point that all

14:33

companies give their developers access

14:35

to something like cloud code, but the

14:37

underlying model is decided by the

14:39

company based on the complexity of the

14:40

task. So, you there will be some sort of

14:42

a model router and when I say model

14:43

router, I don't mean something like open

14:45

router. I literally mean a router that

14:47

is smart enough to know whether the

14:49

incoming request requires fable or will

14:51

it work with GPT 5.5 and or will it work

14:53

with GPT 4 or will it work better with

14:55

open source 0.8.

14:56

A layer that is aware around which model

14:59

the request should be routed to based on

15:01

the complexity of the task, based on how

15:03

much credits does this user have

15:04

and maybe even block requests that are

15:06

you know, personal requests for the

15:07

user. If there's a developer that a

15:09

company has given cloud code access to,

15:11

they should not be using it for their

15:12

personal projects. So, you know, things

15:13

like these. A model router that is aware

15:15

of the incoming questions or requests

15:17

and is either able to block or route the

15:19

request to the right model. As I said,

15:20

speculative up in the air sort of a

15:22

thing.

15:22

No easy way to solve this right now. So,

15:24

whoever does solve this is going to save

15:26

companies a lot of money and make a lot

15:27

of money in the process. How do you

15:29

solve it is still a big question mark.

15:30

There are a few techniques that a few

15:31

companies are trying. Start thinking

15:33

from first principles. There are a few

15:34

approaches that I have in my mind as

15:35

well. Not fully a full-stack challenge,

15:38

not fully an AI challenge, something in

15:39

the middle. And lastly, benchmarking for

15:41

a specific use case. I think there's a

15:42

lot you can learn if you're building

15:43

benchmarks like three events.

15:45

It's not easy writing these e-vals.

15:47

Yesterday, we had a cohort class where

15:50

it was supposed to be a 2-hour class but

15:51

it extended to 3 hours because we were

15:53

trying to build e-vals for dub.sh. And

15:55

in those 3 hours, we were barely able to

15:57

even start the tests locally. Um

16:00

that tells you something. If you ever

16:01

join one of these data labeling

16:03

companies, you will be given access to a

16:04

private code base that you have to write

16:06

a lot of issues, tests, and e-vals for.

16:08

Um

16:08

Setting up the code base itself was such

16:10

a challenge yesterday that I realized,

16:11

you know, writing a thousand e-vals for

16:13

it is going to be really hard. Making

16:15

sure they're actually representing the

16:17

ability of a model to perform well on

16:20

that code base is going to be hard. I

16:21

was telling this in the code yesterday

16:22

that, you know, up until last cohort,

16:24

you know, the cool thing was to set up

16:25

dub.sh locally. And now the cool thing

16:27

is not setting up locally because, you

16:29

know, for a fact, most of the issues in

16:30

dub.sh are going to be solved by an

16:32

agent or, you know, some sort of an AI.

16:34

Hence, the cool thing now is can you

16:35

build a set of e-vals that, you know,

16:37

tell which agent, which model is better

16:39

at solving issues in dub.sh rather than

16:42

solving them yourself. And then as meta

16:43

or dystopian as it sounds, it does

16:45

actually seem like the right direction.

16:46

If you write a lot of e-vals and data on

16:48

how issues are solved in dub.sh, for

16:51

example, then it is probably going to

16:52

solve issues better compared to a human.

16:54

I feel at some point companies and

16:55

engineers in companies are going to

16:56

spend time not just writing code

16:59

but actually writing a lot of these

17:00

e-vals that not just represent the

17:02

ability of a model

17:04

to solve issues in that code base but

17:06

also probably train the model better on

17:08

that specific repository. So, it

17:10

performs better on that repository in

17:12

solving issues in the long run. So,

17:13

try this out for one week. Like, pick a

17:15

random open-source project. And And

17:16

trust me, it's not going to be easy.

17:17

There's like a lot of nuance in writing

17:20

these emails

17:21

and writing the environments that

17:22

actually test a model using them. You

17:24

have one nice project to pick. As I

17:26

said, we have five videos. Let us know

17:28

which one would you like to see on the

17:29

channel next. I have seven projects. I'm

17:31

already building Superset, but if

17:32

there's a different one you want me

17:34

building on the channel, let me know.

17:35

With that, let's end it. I'll see you

17:36

guys in the next one. Bye-bye.

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