So I just had an AMA and I had a lot of replies over 300 or so

replies and I went through and. I do duped all of them and I just

created kind of like the top questions, like the themes,

the common things that people ask me. And I thought I would

just like organize it and just answer it here so that you don't

have to go through and look at all of the AMA to just like get the

nuggets of the entire thing. So yeah, I had a blast doing this.

I'll do it again. So let's do it again. So let's get right into it. Right.

Future suite jobs. This was like a common theme. I got multiple

questions about it, but Kartoa asks, Hey John, appreciate you

doing this. Where do you see the future of SWE?

For context, I'm currently a SWE 2 at Fang Adjacent Company.

I also got an offer for dental school and I'm strongly thinking of

pivoting due to the uncertainty of

our future job market.

Now, Kartoa,

you're putting me in a very difficult position

because honestly, no one knows the future. The future.

You know, obviously there's definitely like safety in doing a job

like dentistry and, you know, anything physical, at least for

a while. But maybe because I'm an optimist, I actually kind of

believe in like Jevons paradox,

where as all these coding agents and coding tools, as they go

down in price and the usage of it goes up, I think even more

and more code is going to be landed. And then that just means

that even more engineers are going to be are gonna be needed

to review the code, land the code, maintain the code. Now I do

know that like junior engineers are having a hard time entering

the field. So that problem needs to be solved. But I think

like fundamentally,

I think we're gonna be all right for a while. At least from

kind of what I've seen, like the latest models outputs,

they're good. But at the same time,

it just like kind of sucks at the same time sometimes. It's just

like LLMs are like geniuses and they just kind of also fall flat on

their face

for even the smallest things. But yeah, so I still think this is a

great job.

I honestly though, if you're just starting out, it's hard for me to

recommend 100%

because the entry to get into it is pretty hard. But if you're

already in it, I think you should just stick around

and make sure you're like learning and pushing and like

adopting these tools. But yeah, thanks for your question.

All right, the next question is around agentic coding workflows.

Essentially like what my

coding workflow is.

So missionary,

given the rate of change in AI, what is the right percent of time

spent on workflow,

harness, tooling, DevX optimizations? How has it changed

since 2022? What is the best way to test new workflows?

By just using doing real work

or is there a certain flavor

of tasks slash projects that's best? Okay, that was a very hard

sentence to read. So the right amount of time, I think it depends

on people,

But there is a balancing act, I think, between just like

always jumping on new things versus

just sticking to what works. So what I like to do is I like to get

like a really nice workflow going. Right now, I use Codex and

especially Codex app for a lot of my work, but I still use Cloud

Code mainly for Fable access if I need something like very

like in-depth to do. In depth to do but at a high level i'm kind of

sticking with what i know and i really know how to like navigate

and i'm essentially a professional at but every time there when

there's a new model that releases and some major events like

grokbot is an example it just came out and it's like kind of

changing some of the paradigms you know it's like open cloud

but it's supposed to be more polished and fully managed so

things like that i will give it a shot and i'll like test it out.

Because it's a paradigm changing like kind of event.

But besides that, I think it really depends on you

and your like

desire to be on the latest and greatest.

But yeah. And then the second one, I think doing just work

is okay. It will get you so far. But the thing is, if you just rely

on doing

work to learn these things, there is a limitation of what you

will learn. What I like to do is I just like read the docs. I'll go and

just like read all of open AI or clock codes, like entire docs.

And then I'll just like try

these things that they're saying out, like try all the hotkeys,

try all these things. So whenever something new comes out,

I have phases

of exploration. Like, oh, I think there's a big Delta. Like I'm

going to give cursor a try again, since I'm going to try Grokbot,

I'm going to play around with it. And. I already look at it and it

looks very similar to Codex app. I actually actively spend time

learning the tools deeply because I think that's the best way

to take advantage of it. And you'll be surprised how much of

your work will change based on some hidden feature that you

may never run into if you were just working. All right.

Question number three, reviewing AI code. Can you explain in

detail your diff

view workflows? How do you review code after it's generated?

Do you open an IDE or just ask another agent? And this was

from Pablo. Oh, I actually got a ton of this questions. And, uh,

if you want kind of like a written down version, I actually made a

post about it recently, but to reiterate that a little bit,

essentially code review is the one topic that I keep changing

my mind on. And it's

like a gradient.

Um,

it's like a sliding window on what diffs or PRS that I review.

Now I kind of look at code review as like, I imagine like a tree,

like I think of like

a tree when I'm looking at the code. So I'll look at the code and

I'll quickly see what files are being changed and like what kind

of code is being changed. And then because at least I have a

pretty good understanding of the code bases I work on,

I immediately know whether that code is touching everything.

If it's an infrared code, if it's touching like hotspots in the code

base that I know,

then I consider that as like a trunk change, something that is

highly interdependent. And if it breaks, it's going to cause a

big issue.

Then I'll increase my own effort level and also my agent's effort

levels on reviewing that diff.

But even when I'm reviewing that diff, I really think about the

gating and the safety

aspects first. And then I'll look at the quality and the cleanness

and those kinds of things. But it's really around safety. Like I'm

highly anchoring on safety when I'm doing code reviews.

And then there's the other like

types of code, which is like leaf node codes.

So something maybe you created a new compose component

or react component. You haven't integrated it like landing that

code only changes the bite size of the code. I guess in

JavaScript land, it wouldn't even get sent down.

But anyways, that I don't look too carefully because the safety

of that code is like very safe. It's like not being used anywhere.

Right. So. I'll definitely just kind of let it go. I'll, I'll review it,

but I'm not like reviewing it like before. So that's like kind of this

high level, like philosophical thing. It's really around safety.

And then like, how can I get the diffs to be safely reviewed?

So all of the old practices still remain of making atomic diffs,

like very small diffs, single purpose. So that is easily revertable

and also like easily testable. And then the other big thing is

getting the agents to really give proof and confidence levels on

their validation. So whether that's screenshots, videos,

or runtime data, maybe they do A-B testing on some like,

I don't know, some like runtime logs, and then they tell you, oh,

the runtime state is like this. So it's actually good.

Unit tests, all of these things that's around safety,

I overemphasize in my code writing. And I'm optimizing for like

landing code at big tech. So it's really about

trying to land it safely as possible. Because at the end of

the day, any code that you

write with AI, you still need to back it up. You can't go to like a

sub review and say, hey, the AI wrote it, it's not my fault.

You know, at the end of the day, you have to stand by

what you put out.

So I anchor really heavily on safety and then really tune my

effort based on how safe I think that piece of code is. I hope

that's helpful. I wrote a little bit more detail on the post. So feel

free to read. All right.

So the next question is around architecture

without experience. This type of question I actually got a few

times about people asking about, how to learn things when

they don't have experience. I think it's actually really cool

because it just shows that there's a lot of vibe coders or people

who are just interested in this kind of things who never

would have

really been interested in learning how to code. I think not,

I don't know, Joshua, but I think there's this whole set of

subcategory of people that just are

getting into coding because AI is making it easier and

more accessible. And I think that's really

cool. So hi, John, thanks for doing this AMA.

Coding becoming more common, how do you make sure that

AI is making good architectural and coding decisions?

Especially if I don't have

much experience. It's one event, so I'm missing some of

your posts, I think. But

the high level gist is that

honestly, it's very hard if you don't know something you

don't know. You could essentially use the AI to do

adversarial reviews.

So there's a lot of skills around like grilling your architecture

and all this kind of stuff. But at a high level, you should always

anchor on trying to get as much understanding. If the AI is

doing something and telling you some plan, if you don't

understand it,

I think if the goal is for you to get better, I think you should

slow down. You should pause and try to understand

what the AI is doing. Like every word, every term. Like recently,

I started learning about like model scaling, and I'm not a

machine learning engineer. Or like an AI researcher. But I saw

Ryan Peterman's podcast on some, I forgot his name, but he

was from like Google, DeepMind. And he recommended this

series on Jack's ML. And essentially, it's like how to think

about model scaling. And I've been kind of consuming

that material. But it's been very slow. It's like, I'll read

one paragraph, and I don't understand it. And.

I essentially have,

you know, Codex or CloudCo, both are fine, like the

Chrome extension. And

it's looking at my browser. So I'll just

say, hey, what is this? What is that? And it literally takes me

hours to get through like one page.

But after I do it, I have a much deeper understanding than if I

just had breezed through it, right? So in my opinion, I think if

you don't know good architecture, or good coding patterns,

or things like that, you know, study the fundamentals,

like question and ask and use AI to teach you how to be a

better engineer. But yeah, that's what I

think. It will slow you down. Like if the goal is to build products

and like just really stuff and like have impact in the world.

This may not be like

the best advice. But if the goal is to learn and get better,

then yeah, use AI to learn and get better. All right,

question number five is update on the coding walkthrough.

So let me read Alan's post. Can you make a V2 for this video?

And he linked one of my. YouTube videos.

With the latest CloudCo video

and

all the things you've learned since the last video, Nothing beats

a screen share and actual code, which is totally true. So I had

this video, it went kind of viral on YouTube. I think it was like

sitting at almost 500, like half a mil views. And it was my first

like major breakout video. And this was right around the time,

like, I don't know, when Opus was going viral too,

like November or like early

December or something like that.

But it's a really thorough video. And I still think a lot of the

things I covered in that video still stands. I think the way I

covered it, and it's not just about like showing things

about Clock. Code, but I cover some bits of agentic

engineering and things like that. But I should make another

video. The thing is, I did make another version, but I happened

to just like use that material to teach with ByteByteGo. But I

should probably make another version because even

since then, since I've been teaching, I've created a whole new

stack using Codex app more.

And then goals is a thing. Graph engineering is a thing.

And there's like graph

bot is like another thing that I'm like, got to look into. So there's

like new things. But I think agentic engineering and the various

pieces of it, like context engineering, compound engineering,

harnesses, agentic tooling, agentic validation, all of these kind

of core

fundamentals of agentic engineering, I think is here to stay.

They're evolving within its own thing, like their own lanes,

but they're all

like core, I think, to agentic engineering. So long answer short,

it's in the works. But these kind of videos literally take months

of research of just being in my workflow for

day in and day out. And so it takes a long time. Yeah, so I'll get

to it as soon as I can. All right, the next question is around

manual coding.

How much code do you write by hand? And

the simple answer is none. I don't manually write

any code by hand. And, you know, I think this is

pretty common. I think the only people who are actually still

writing code by hand

is people who want to because they want to feel something

different. There's a lot of there's a group of people who

genuinely just love to code like manually. And I was that I

actually don't have a CS degree. I have a non

traditional background. And I found coding

as a hobby.

And it took me like five years of just doing it on the side building

random apps with my friends until I decided to become an

engineer and then like did a boot camp and all that kind of stuff.

So when I stopped coding by hand, there was a period of time

where I really mourned it. And I was like, man, this feels kind

of bad. And, and during that time, there was a lot of babysitting

auto mode wasn't even out yet.

Right. So

it was a lot of just like constantly juggling so many

different threads, and they would stop very quickly. Like these

days is a lot different, right? You're like, writing long goals that

do bigger tasks. So you have more

kind of

ease of mind, at least in the brain, like context switching.

But yeah, so manual coding, I think is going away as sad as

that is. But yeah, it's just kind of the cost of this whole,

like change and revolution and anyone telling you that, hey,

manual coding is not going away is, you know, it's more like

they want it not to go away, I think.

And maybe for their very specific code is like very specialized,

and they don't feel like AI can do it. And I think there's some

merit to that, right? Like AI is not going to write perfect code the

way you want it exactly.

But, you know,

it will get the job done with the right context and the

right validation, in my opinion. And this is kind of where,

where the whole like coding is solved meme comes from,

like people go, oh, you can't even like update the codex app

without having you to like install an update or something.

And then they'll put like coding is solved. I think that's just

like disingenuous, honestly, because like with enough context

and with enough validation, I think,

like, the coding part is solved. The AI will do the coding for you.

But that doesn't mean there's all this like other true engineering

work that is just

like done. But yeah,

that's kind of my high level tech on the manual coding. All right.

So the next question theme was around

competitive advantage. So my hash, Vivi says,

if AI makes everyone capable of building, what becomes the

real competitive advantage? I think there's a lot to this here.

The first thing is around domain knowledge and like

domain specialty. I think like social media is like a really hard

thing to compete on is because

the true power of social media is really the graph,

the graph relationships. That's a mode that's very hard to beat,

even if you have AI. And also for all of the major big

tech companies, I think data is a big one. Like Tesla,

for example, it's hard to beat them on self-driving

because they have so many miles. And so the only way like

other companies can beat them is through different ways,

like adding LIDAR, for example. That's how like Google is

getting around it, right? With Waymo. But Waymo also has a lot

of miles, right? So I think there's still a ton of ways for you

to compete. Now, if this question is really like on individuality,

like how do I make like a successful app or whatever?

How do I like start a startup? I think this is a question that

always comes up. But if you just look at the data of all

these companies, there's been more newly minted

unicorns in the past like few years because of AI. And I actually

think there's a delta in the skill

gap of like AI usage. And I noticed that because I teach

AI coding, I just like see the knowledge gap between

various people. And like most of the students that I've taught

like the stuff to like,

like their mind is blown, right? And the thing is,

when I talk to like average non like Silicon.

Valley, like tech bros, people who are not in this little bubble of

like always talking about rock and image generation and LLMs

and open source and, you know, like model costs and all like

there's normal people are not like that. Most people don't give a

shit about any of that.

And when I talk to them, most of them are just still on ChatGPT

or Gemini. And they don't even know that like

Codex exists and Codex can do all these things. And some of

these people are like young. They're not like my mom. My mom

uses like Gemini and ChatGPT and she's like, oh, it's so great.

But like average people don't need these other things, right?

So,

but yeah, so in my opinion,

there's multiple ways to answer this question, but in terms of

like competitive advantage, I think you'd be surprised how

few people

know these AI coding tools deeply. So if you're learned that,

if you're on board with that, then you're going to have an

advantage here. All right, the next question is staying sharp

with AI. All right, this guy has a bad username, but he had a

good question. So how do you stop writing code manually and

not lose your coding sharpness? So this is very related to not

coding manually question. Personally, for me,

I'm constantly learning. Whenever I don't know

a piece of architecture within the code base, or if I run into

some kind of new patterns that I haven't seen in the code base,

while I'm reviewing the diff, I'm always asking questions and

digging into it. And for me, I actually spend a ton of time

on my own digging into other new technologies. Like I think I

mentioned that I'm going through the Jaxx ML like scaling

model scaling

course on my own because. I'm just interested. The big

advantage for me is I create content. I create content around

this topic because it's actually my hobby for some reason.

But yeah, I really love learning about all this stuff,

making content about this stuff. So for me, it's kind of like a

flywheel because, oh, I want to make some content about

some AI thing

that I'm interested in. If I want to teach, I need to understand

this stuff, right? If you want to test your knowledge, you teach,

because once you start trying to teach it and like verbalize what

you think, you know, then you'll see the gaps in

your knowledge. So

I like to try to teach. That's kind of the reason why I actually

started YouTube eight years ago, because I just wanted to like

talk about this stuff. And doing YouTube is like a really good

reason to just really get into it. So I always try to teach what

I learn.

And

because of that, I post and I get excited about it because we

were talking about it and it's fun. And then it actually changes

the way I work sometimes.

I just see a lot of benefit of continually continually learning and

trying to teach what I

learned and that's how I stay sharp so I recommend it for you if

you're interested in teaching but if you don't then you just have

to be good with having deep focus time and learning just

constantly learning all right next one is around becoming

job ready

so

Vashnavi

Patil hi John I'm not in the job market yet all right

someone young

as someone getting into software slash cs what skills or ai

related things would you recommend focusing on now to be

truly job ready

keeping it at a high level what matters the most versus what's

just noise

so at a high level i think learning

one of the like agentic coding tools whether that's like codex

clock code open code i guess cursor whatever it is learn it

really well like understand how the memory works understand

what's a skill what's an mcp

what's agents agent teams what's the difference between sub

agents versus agent teams in claude's case what's ultra code

in clock case when should you use that how good are the

different models

like do you have a good enough sense of using all these

different models that if i gave you the output of like something

from chat gpt models versus like clod do you think you could

like see the difference just by the feel of the model can you feel

the model because you kind of can if you use these models a

lot you just know that's like from opus or that's from soul you

know you kind of can get a sense of it so yeah just get really

good at all of these agentic coding tools and learn agentic

engineering there's so much to it so much to learn i i think

maybe not yet but all of these things that I'm talking about,

I think the industry is going to move towards trying to filter out

candidates that does

know these stuff or does not know it. Like it's, I think the

interview process will change at some point where these are

going to be tested. So I think that

is probably one of the most important. And then the other

big thing, in my opinion, is still learning foundational,

like fundamentals of like data structures and algorithms,

design patterns, you know,

clean code, Bob,

Uncle Bob, you know, watch his stuff, right? So

all of these things, I think the core foundation is really

important because agent decoding is essentially

an abstraction. It's a layer above, right? And abstractions are

inherently leaky. What that means is every now and then some

abstraction layer will

like mess up, like something will not happen, even though it

should theoretically happen. And that

happens all the time with AI, where you're. You're like, oh,

I think it should be working, but the agent is saying this. I don't

know if it's actually right or not. And there's all these things.

But if you don't have a foundational layer, when things get hard

and when things get stuck and the agent has a hard time,

you won't be able to steer it.

So having a really good foundational skills in engineering will

help you peel the abstraction layers when you need so you

could dig in and really debug and have confidence in your

product and your coding. So,

yeah, those are the two main things I would say. But yeah,

I hope you do well when you finish your classes. All right.

Choosing coding agents. Question number 10 through 12.

So welcome. How do you feel about Codex versus

Cloud Code?

Are you using both in any way or just Cloud Code?

So I use both

Cloud Code and Codex app.

My class is actually primarily based on clock code, but I'm,

you know, working on a version for Codex,

especially Codex app. So this question is actually more than

just clock code versus Codex, in my opinion, because

the apps itself,

they have, there is a Delta between the apps and there's a CLI

version in itself.

And recently I've been converted to essentially not use the CLI.

I've almost completely went to the app. If you asked me like two

months ago,

I thought it was like, there's no way I'm giving up the CLI.

It's just so powerful. There's so many things that you can do

with it.

But there's just like a few features on the app side

that has convinced me that the long-term priority

and all the feature sets are going to go to the app,

in my opinion. Because number one, I think all of the major labs

are targeting mass market,

right? That is their strategy. And that's kind of the ultimate goal.

I think Anthropic is like targeting enterprise, but at the end of

the day, they want to get like the entire pie of enterprise, right?

But for OpenAI, they're definitely looking at mass audience and

the enterprise.

So they're gonna keep pushing the app

because it has rich rendering, it's not a 2E, so you can actually

do a lot of very interesting things, like a full-fledged app, right?

You're not limited by the 2E's limitation. Yeah, so rich rendering

is a very

good example, but there's so many little things, right? You can

control the entire

Codex app with your voice, using the live ChatGPT voice,

you can queue prompts. I know you could queue prompts in

other places,

but you could just like drag things around, like queue 10

prompts and then drag them, edit them. There's just like a lot of

little things. And over time, I think it's just gonna get better

and better.

And yeah, I think it's gonna be hard

for just a 2E experience to

win out in the end. So that's why I switched.

That's one of the main reasons. In terms of like model quality,

I think price is becoming a factor for me, even for me,

because Fable's usage and rate limits run up so fast.

It's because this is a very expensive model.

So they limit it.

The thing is like GPT-SOL has been great for me. And because

of the price and like rate limits, I actually switched a little bit to

using Codex

and Codex app and stuff like that. And then I found like the

output very comparable to Fable.

And I actually think it yaps less.

It has like a dog in it. Like it just keeps going. Like really goes

until it gets the answer. That's how my friend Peter puts it.

It has a dog in it. When I just tested both of them out, I just

prefer the OpenAI models right now.

This may change next model release, who knows? But that's

like kind of the thing with these models. But I will say it'd be

very hard for me to give up the Codex app because I also use it

with my phone and I remote control all of my threads

and my laptop at home or whatever is just like open. It's just on

all the time.

And I just control it. I control it wherever I am. So those kinds

of things, all the schedules and things that I've created on it.

And like, yeah, it's just hard to give that harness up, I think.

But yeah, so that's kind of where I am right now, but it

may change. All right,

agent permissions. And so Daniel asks,

curious how you think about permissions in agent decoding.

At work, I stick to coding agents in the terminal. Personally,

I run multi-agent Hermes

and tight roles.

So one failure doesn't spread. So I think permissions wise,

the whole agent decoding is going towards longer runs using

goals and less permissions and then having like

agentic safeguards

or like an agentic garden that protects your code base

and all the work that you're doing. I think this is a long-term

right approach in my opinion. And then the models should just

get better and better

at doing things that most people don't want.

Like dropping a DB without, you know, approval or messing up

your OS and things like that. So, but yeah, I think less

permissions is fine. As long as the output is what you want at

the quality that you want and you have enough

agentic safeguards. All right, the next question is

fundamentals and AI. For students building

AI systems,

how would you balance learning fundamentals

with agent decoding and rapid

prototyping? For students, I think you should just focus

on learning,

learning fundamentals. Like why do you need the prototypes?

Like if you're not going to do anything with the prototypes,

what's the purpose of prototyping? The main reason why we

prototype for like at work is because we want to get a feel for a

product that we want to build so we can show leads or like play

with it ourself,

right? So there's this like reason for the prototyping. But when

I'm prototyping, I'm rarely looking at the code. I'm just like

getting it. I'm mainly focusing on how, like what the field is.

You know, that's for like client code, like product code. But if it's

like infra code, my prototype for infrastructure code is really on

exploring different paths.

So I'll have multiple agents prototyping different ideas I have

on how I think the architecture should work.

And then I'll get it to a certain point and I will play around with it

and feel the API

and see kind of the niceties and like things that I might've

missed if I, you know, miss an abstraction layer or something,

or if I didn't have certain APIs here, or if I expose

this in the wrong place, like those things, it's hard to imagine

just on like a UML diagram, but it's easy to see when you're like

trying to add new things to some existing infrastructure,

you know? So like there's,

depending on what you're doing, prototypes are really good,

but they're not really for learning. It's really to answer a

specific question

about UX or, you know, some

feel. You're trying to get a feel for something.

But if you're in school,

then you should just focus on fundamentals,

in my opinion. All right. The next question is around

AI engineering

specialties. Saravana Kumar asks, AI landscape seems to be

very broad these days. From serving models efficiently,

training models,

RL training, AI infra, security,

custom engines, and memory,

where should one focus today? I mean, this is a very

personal question, I think. Like, I think any one of these are

great. You could get a really good job doing any one of these,

I think. So

it depends, right? It depends on where you are in your journey.

Of like employment.

Like for me to like switch into

any one of these things would require at least like two years of

really dedicated effort to be like back to kind of the novice,

like kind of like mid-level of probably one of like

training models.

I'm just learning about it because I'm curious. But if I

wanted to really be like an AI researcher that trains models,

or even maybe I do

engineering where I, you know, optimize kernels for

model runs, man, I have to spend like two, like a year to two just

to catch

up on the latest like frontier, like what the frontier

knows already.

And then from there to

contribute my own

thing, my new ideas and write papers, it's going to be another

few years, in my opinion. So like it'll be at least like four years

before I'm like baseline par. And that doesn't mean that I'm

going to be even the world class, right? I feel world class at

what I do right now but it's going to take a long time for me to

get there

to do this in some other domain

right but to answer your question about one of these things

honestly i think you should pick what you think is interesting try

a little bit of it talk to a bunch of people who is doing that and

then you know just try

like the very basic things like the training models example like if

you can get through Jack's ML like model training and you're

able to train your own like little model

using a GPU or even a MacBook

with like a high RAM, like you could

see how you feel, see how you like that, see how that feels.

I think that's probably the best way

because chasing a job just for money has never really worked.

In my opinion, you always quit because it's too hard to

get through. Any one of these things require a lot of effort.

All right. The next question is next AI leap. So what do I think

the next AI leap is? So

Paul Sufka,

MD,

doctor, doctor. Hello. Hi, what are you looking forward to or

hoping for in AI

in the next six to 12 months? What will be the next leap?

Honestly, there's hints of it, but I think I really want to do

everything mobile. So the voice control is really good with like

ChatGPT's voice, but I just need better like visuals,

like feedback mobily. And I think glasses may be a good

form factor. And, you know, meta displays, really great. So I

might get pick up a pair of those. Shout out to meta. But yeah,

I just want like really good interface just mobily so I could like

code on the go or like juggle my agents on the go. That to me is

like the most interesting because one thing about this

profession is that. I have to sit and be in front of a computer for

a really long time. And if I could

figure out a way to not do that, I think it would really change

my life. It would really change like the quality of my life. Oh,

and another thing that I would be really interested in

is

video. Input as like a raw input. So right now,

most video inputs, if you add video, what happens behind the

scenes is they take the video, they use like FFM peg

or whatever, and then they'll get all the frame data and try to like

piece together what the video is through frame data. Now, I use

this workflow when I'm debugging animations or trying to build

some animations that my designer made or whatever.

And it's OK, but it's really slow and it doesn't work all the time.

Yeah, it's just not. Great. So it would be great if we get video

multimodal like support like natively. That would be

really interesting. But yeah, it's a video multimodal input. I think

those two are probably very interesting for me personally.

All right. So the next one is around portable agents. Makan

asks,

as

agent coding evolves, do you think the agent will eventually

become portable

across models and coding harnesses with things like identity

payment permissions

and memory

following the agent rather than being? Owned by Claude Code,

Codex, Cursor and et cetera. So this is a really

interesting question. And I think this is kind of hinting at some

of the

world that will eventually live in, which is like agents that

operate for people will have their own entity. I actually do think

that's going to happen. And I think

Stripe is seeing that. That's why they purchased like

open router. I think I think they're they're just trying to get ahead

of that and they'll probably create something like that,

in my opinion. And it's going to happen very quickly where

agent traffic is going to be more than human traffic in, in,

in the internet, obviously. So the concept of agent as an entity

that operates on your behalf, that has its own credit

card information,

Gmail credit score, maybe,

I don't know. It's like all these kinds of things. I think it's going

to be a thing. And I think as personal agents like

gain popularity,

I don't know if Grogbot or whatever is going to be a thing,

but yeah, as all these things gain popularity, the, the necessity

for the agents to have their own entity,

I think is going to be

a thing. I don't know if like the whole memory and all this kind of

stuff is going to get ported around with this, but I'm talking

more about what it required to be an entity. I think that concept

will be created and ported around. I actually think Stripe is

going to be the one that creates it, but we'll see. All right.

So the next data is confidential data.

Welcome John. We'd love to hear your take on uploading

confidential documents.

To AI even when data control is off? You know, I get this

question actually quite often, especially people who are in

these fields like HIPAA compliant

or EU, like people who do a lot of work with the. EU, they just

have this like

higher sensitivity towards data. And

my answer to this has always been figure out a

way to like host your own like servers and host your

own models, whether that is doing a small build out. And the

thing is like local models are so good with open rates. If your

company has a budget to buy a few like really good like H100s

or whatever,

and then have that set up and just have one person that's like

dedicated to just managing that system. Yeah, just use open

source to like use these models. You know, maybe you just use

open code and you could get so much done, I think, just using

these like open tools and open weight models. So that's kind of

my answer to that. If not, then you're always going to be just

kind of

like have to use

some enterprise like box version of Anthropix systems or

open AI. I think they have an enterprise solution, but I

don't know. All right. Open models and security.

Question number 17. Pragnya and Ramath. Thoughts on open

source models being used for cybersecurity warfare when they

are as good as closed source ones in a few years. So I think it's

already happening. Like this, this, this is already happening.

I think Hugging Face had to use an open source model

because the

closed source models were like rejecting the

like cybersecurity

like prompts because it detected cybersecurity. So they had to

use like an open source to get around

like that

post training, that post training kind of like

safety concerns. So I think it's already happening

the other way. Right. This is saying like, what do you think

about cybersecurity warfare happening? Honestly, that's a

scary question. And I think it's a legitimate thing to be

scared about. And a lot of people joke about, you know,

Anthropix or open AI or any of them AI labs like saying that,

we want to slow down. Like we're very concerned. Like people

joke about it because they only see, I don't know. I think like a

lot of people just see that and they think, oh, they're high.

They're just like trying to hype their models up or things

like that. But I don't know. I think it is pretty scary.

Like cybersecurity itself is pretty scary because these models

are pretty powerful and there's just a lot of like zero day

security holes. So yeah, I don't know. Like I try to not to be

so pessimistic

and that like these companies have so many ulterior motives,

but I think they're just calling a spade a spade sometimes that,

you know, there is serious cybersecurity

like concerns with models and open source models and all this

kind of stuff, because, you know, let's be real. These, the open

source companies, like especially the ones in China are not

going to care

that much

about like security. You know, I'm not, I'm not trying to get

political here, but yeah, that's just the reality. All right. The next

question is getting into meta or I guess any big tech. Hi,

what advice would you give your past self to make it easier to

look for a job at a company like meta? So I actually made a

bunch of videos on this topic. So feel free to look at that on my

YouTube channel. But the main advice that I have is there's like

a few things like I wish I had good grades and went to a

good school, then it would have been a lot easier.

But the other path is first, you actually do need to just know

the fundamentals, like data structures and algorithms to even

pass these interviews.

But even to get an interview, you need to have been working for

a few years. If you don't have the pedigree of like a really Ivy

league school, then you need to figure out a way to get

into like, like I went to capital one, like I worked at a bank for a

few years, like four years or something. And around year three,

that's when

meta and Netflix and all these other companies started

knocking on my doors and I got interviews. Like before then I

would send a bunch of resumes. I never got the interview.

But after around year three, that's when everyone wanted

to talk. So I think you just need to have those things. And then

once you get the opportunity, you just have to nail

the interviews. I think it's actually not as hard to get into a

company like meta or one of the big techs, in my opinion,

because it's like interviews in a sense is a solved problem,

but it's hard to commit to

studying for it enough

that you can pass it. You know, I think neat code is really great if

you want to like really dive in and learn like the code, but yeah,

neat code is great. If you just know like 80% of neat code and

then you could get into Fang

or, or mango, whatever you call it. In my opinion, if like, you got

to just get so good at those kinds of things where you see,

you read a problem. And as you're reading, you already have

a sense

and feel what kind of data structure algorithm problem is it?

Is it a sliding? Window problem is it a tree problem is it a link list

problem is it a graph problem I don't know there is a subset of

these things and you can start pattern recognizing

so for me personally. I tried twice the first time I got my ass

kicked I failed every single interviews I did I did basically

everywhere and I got I just like failed that taught me what I was

missing that taught me the gaps and then the next time I

studied literally for a month and a half four hours a day I was on

during that time and then I got two offers like a meta and Netflix

and I took the meta offer that's a whole nother story because

it's like anyways

and then the last thing I think I mentioned this. I think this is a

new thing but learn AI coding

I think agentic engineering is

the future and a lot of these companies are going to test for it

and then another like side tip that's like not necessary is to

build in public

like try to build and put yourself out there make connections

you know doing social media like doing like content creation.

I meet so many cool people and it's, it's one of the

best ways to network. It just opens doors that you'd never

thought was possible.

And then social media is one of the places where it's an even

playing field. Like you could be 10 years younger than me

or

10 years older than me. Doesn't

matter if you just make good content and you put a lot of value,

people will notice, and then they'll reach out to you and then do

stuff with you. Maybe give you a job. I don't know. All right.

Next question, getting some career questions now, which is

growing to staff.

So hi, premium goblin. It's weird saying this because this is like

a random name I came up with

for my experiments that I was running on X and threads and

other social media platforms. But what is necessary to become

a senior staff software engineer? And could you give tips on

becoming an excellent suite? So like, this is a, I don't know,

there's

so much, I actually have a lot of insights into this topic,

mainly because to

do really well at a company like meta, you not only need to be

text technically excellent, because that's just the bar.

Like everyone's just very

smart. And then they have a very high technical bar. That's just

like the minimum,

but you also have to have pretty high EQ. Like people will need

to like you,

people need to like working with you. Right. And you need to

be pretty easy to work with, but not also not a pushover at the

same time. So there's like all these like various factors.

But from going from E3 or L3, like junior level to senior, you just

have to do

your job really well. You know, E3 is just coding

really well, just landing a lot of stuff. E4 is like taking a project,

like by yourself all the way to the end with minimal,

like handholding. E5 that's the baseline but you're now like

mentoring other mid-level engineers and you're slowly creating

some scope for yourself

and you're taking on even more complicated projects but from

senior to staff it's essentially a different job like from a technical

perspective a senior engineer and a staff i think there is not a

huge amount of delta to be honest and even for like senior staff

but obviously there is a delta but like the the true delta between

like senior roles and like staff plus roles is around behavior.

Actually, there's a lot of behavior things that make you really

stand out. And it's like, it's a totally different job. Like my bar for

staff is you have so much ideas and so many things that you

want to do, and you're good at executing on all of those ideas,

but you have so much that you have to and want to do that you

just need people to help you. And that's inherently

scope creation, right? So that's why a lot of tech leads is the

archetype for sixes because they help create scope. And then

they essentially have impact that scales beyond

just themselves.

And seven plus is that, but at org level scale, not just

your team. So you're like multiple teams and cross org. And it's

really around influence.

So it's really hard to get into in just this like small like video,

but it's essentially, you got to know that staff plus

is like a different job.

It's a different role. All right, so

we're at question number two. 20, and this is agents outside

of coding. What's been your favorite way to apply agents

outside of coding?

So I actually use agents for all of my stuff, like whether it's

helping out with taxes, you know, the typical ChatGPT

use cases,

but I also run my entire content pipeline through Codex. Yeah,

I think at one point for my personal stuff, I was using Codex

more for content pipelines

than like coding itself because I didn't have like a side,

like coding project to do.

But yeah, that has been my favorite. All right,

product decisions

and AI. Mahesh, provide insights on product design,

no responsibilities and meta and paradigm shifts towards

human AI interactions and collaborations.

What work are they doing now? And what skills do their

roles require?

So not just product design,

but product management,

data science, data engineer, like all of these were, a lot of these

roles are kind of blending a little bit. Obviously,

people have specialties, but like our designer is amazing.

They're basically a designer and an engineer.

She essentially codes and makes a bunch of like UI changes.

And like, she doesn't ask us to, you know, fix this pixel or

that pixel. She just like does it.

And it's been really nice. It's really great. And they're closer to

the code and they're like, it's like they get more empathy to

what we do. I have always like gave my own opinions about

designs and stuff, but I'm not like doing the designs yet. Right.

But in terms of like data science, I'm doing my own queries now

I'm doing my own analysis. I mean, I've always done that,

but it's a lot easier with. AI to run a queries and get like impact

sizing and things like that. So roles are definitely blending.

Yeah. It's, it's, it's very interesting, not just us, but I don't think

the entire industry has figured it out yet. That's why everyone's

being just labeled as member of technical staff, like the labs.

But yeah, it's a very interesting like state we're in right now.

All right. Lessons

from meta products. Nice to see another new L7

suite on here. What have you found as the most interesting

projects of the ones you worked on and why?

Interested to hear both from a technical perspective and

organizational one. So obviously working on reels was like a

dream come true at the time. I was like a E4, right. When I got to

join reels and essentially I got to like senior during that time,

my tenure at

reels, but I was there when we launched it. So I got to join the

team that created it. I always get to say that even though I

joined at the very tail end of it, but yeah, that product was

really interesting. It was, I like, it was very technical and hard

for me because I switched from full stack web to mobile during

that time because I really wanted to work on reels and it's a

really interesting product.

And a lot of people don't know that reels is actually

a music based app and everything's indexed around music.

So that's a really interesting project. I also got to work on

meta AI, like the original, like connect launch of meta AI

on messenger. I know like meta AI is not like the most widely

used right now, but I, this was during. Lama three days where it

was very popular and like a lot of people were using it,

but yeah, it was, I think that was really interesting because

none of us, the entire industry hasn't really figured out what AI,

like interacting with AI look like. So all of these new UX patterns

and like client

had to be like, just kind of imagined. And I really enjoy that

like experience. All right. So the last question is growing

content reach. Draco asks, Hey, yeah. And welcome. What are

some of your best tips for content creations in the agentic era?

I've recently just started sharing content and so far limited to

my blog and LinkedIn. How do you approach the reach of

your content? So

a few things I have content pipelines with my agents. And the

other big thing is I use. LinkedIn, which is my friend,

Kevin's a startup. And what I do is essentially post to one place

and it fairies all of my content. And it has been really helpful

helping me go viral in like places that I'd even know I was

going viral. But as a content creator, like the best advice I can

give to you is don't rely on just one platform and just make sure

you post, but make sure you reuse a lot of the content that you

have repurposed them because people hang out in

different places.

They don't always cross paths and some people only stay on X

some people only stay on threads. Some people only consume

on YouTube. So just be everywhere every creator is doing this,

you know. But they just have a team do it. I'm just by myself.

So I leverage ferryman so shoutouts to my

friend Kevin for making that but yeah, that's basically it. Alright,

so that's all the questions on the AMA that I thought I kind of

like dedupe as much as I can. I'm sure I missed a few but feel

free to add more questions here and I'll try. To get back to you

as quickly as possible. But yeah, I really love doing AMAs.

I feel like it's just like a fun way to interact with the community

of like people that have the same hobbies that I do. Like same

interests and I could talk about AI and like software and career

stuff all day. I don't know why I'm weird, but this is what I enjoy

talking about. And I don't get to do that all the time because my

wife doesn't want to talk about that kind of stuff. She just wants

to talk about Korean mom dramas.

But yeah, thanks for. Watching this little series of videos or this

little video. I hope to make more of these kind of videos and

thanks for supporting me in this journey. All right