YT YouTube2Text

YouTube2Text → Transcripts → How AI Makes Him Crores: Second Brain, Automations & Systems | Vaibhav Sisinty | FO557 Raj Shamani

How AI Makes Him Crores: Second Brain, Automations & Systems | Vaibhav Sisinty | FO557 Raj Shamani — Transcript

by Raj Shamani · 18,856 words · 2,920 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00What's your biggest fear today actually
0:01with AI with jobs with people watching
0:04this?
0:04>> My biggest fear is people not being able
0:06to understand the balance between what
0:09we should get in AI to do and what we
0:11should do because AI can do everything
0:12today and you should not get AI to do
0:14everything today.
0:16>> How do I know a line between how much
0:18should I let my team or myself should I
0:20use AI versus how much should we know?
0:22>> Do the work that was low value for you
0:25all throughout which was operational and
0:26get that done by an AI. The last layer
0:28of work is something that you have to
0:30focus on. Eventually
0:32options
0:35what is that I want to take and why and
0:37build on top of that. If you don't try
0:39to touch it, that is what happens. You
0:40get average outputs.
0:43[music]
0:43If my job is to get an AI to work at the
0:46level that I operate or get it close to
0:48me, it will not happen overnight. It
0:50will probably a 12 to 18 week journey
0:52for me to fine-tune it. How will I do
0:54it? I will build a second brain. Very
0:56simple. What are the things that I
0:57consume on a every single day level?
0:59Let's say YouTube content that I'm
1:00watching every day is very very
1:02valuable. How do I feed the same context
1:03to AI? So I built a daily rapper which
1:06opens my YouTube history, goes through
1:08all the videos, ranks all the pieces of
1:10content, everything that I've seen of
1:12AI, it pulls that data, converts the
1:14transcripts and looks at what are the
1:15key pointers, saves it in form of JSON
1:18cards. JSON cards, what is the value
1:20pack of each one of the podcast is what
1:22it saves. That is one. Two is a lot of
1:24times when I watch content on AI, I
1:26share it with my team. Where do I share
1:28it? I share it on my Slack. So, I have a
1:29triage running for my Slack which looks
1:31at what are the conversations I'm
1:33having. This is number two. Three, I
1:35realized one of the highest value
1:36conversations that I have today is
1:38inside of standups with my team. Every
1:40single meeting is transcribed, [music]
1:42put in a GitHub repository and inside of
1:44the repository, my team can access that
1:47information at any point of time.
1:48Finally, oh my god, how can I forget
1:50this? This is a game changer which is
1:54>> what is the easiest way to make an agent
1:56who do all of this?
1:57>> There is something [music]
2:02>> I have a small favor to ask you. I need
2:04you to subscribe to our channel. The
2:07more subscribers we have, the better and
2:09bigger guests we can bring and provide
2:11you more value through these
2:13conversations. And the full audio
2:15experience of this show is also
2:17available on Spotify where you can
2:19follow us and listen to the new episodes
2:22as well. Now let's get into the episode.
2:24If you want to know how top creators and
2:26companies are getting hundreds of
2:29millions of views a month and making
2:31money out of it and scaling their
2:33revenue to 10x. This episode breaks down
2:37exactly how they do it with AI. The
2:39difference is the system behind the
2:41tool. They've built a second brain that
2:44feeds AI. Everything they read and
2:46think, AI does 90% of the work for them.
2:49In this episode, WebV and I will show
2:52you how to build that system step by
2:54step, how to set up your second brain,
2:57how to use AI agents that find your best
2:59topics, how to create scripts that go
3:02viral, and how to upgrade your own
3:04thinking. We've also built a free AI
3:07focused community where we teach AI in
3:10depth. If you want to learn how to make
3:12AI work for you, join the community. The
3:14link is in the description below.
3:20I want to understand last time when you
3:22were here, you told me that
3:26you have scaled your company
3:27significantly faster than anyone.
3:30>> Yes.
3:30>> From what point before AI, if your
3:33revenue was 100 rupees, how much is it
3:36today?
3:38>> 1,000 rupees.
3:4010x.
3:41>> So you have 10xed your company in last
3:43two years.
3:43>> Two and a half years.
3:44>> Two and a half years with AI.
3:46>> Yeah.
3:48>> Most of the efficiency has come.
3:49>> What did you do? What what was the
3:51efficiency? How did you make how many
3:53people are there now?
3:55>> 400.
3:56>> And this work was possible with 400
3:59people before?
4:00>> No. That is what we will talk today. So
4:03there are some things that we spoke
4:04before.
4:06>> We spoke about jobs. first podcast.
4:11>> A lot of things that we spoke then are
4:13not true anymore.
4:15>> Things that I have said are not true
4:17anymore because I've seen the other side
4:19right now.
4:20>> Okay.
4:20>> Which we should capture.
4:22>> The audience needs to know both the
4:24sides of the story is what I feel.
4:25>> Tell me [clears throat] what is not
4:26true. What's your biggest fear today
4:29actually then you come to tell me and
4:30what's
4:31>> my fear today
4:32>> with AI with what's happening in the
4:34world with jobs with people watching
4:37this
4:38>> my biggest fear is people not being able
4:41to understand the balance between what
4:44we should get an AI to do and what we
4:46should do
4:48>> because AI can do everything today
4:50>> and you should not get AI to do
4:52everything today
4:54>> what do you mean by that what what are
4:56the things you
4:57not let AI do.
4:59>> Dude, tell me something. What do you do
5:01on everyday level at uh at work?
5:06Are you the one sitting and working on a
5:08computer?
5:08>> No.
5:09>> Are you the one being all by yourself
5:12drawing on?
5:13>> No.
5:14>> You take decisions.
5:15>> Yeah.
5:16>> Right.
5:16>> Then I'm judging what's right, what's
5:17wrong.
5:17>> Correct. If you think about it,
5:20leaders always have done this. It's
5:23nothing new.
5:25What we understood on top of that is you
5:28you learned how to build a team. You
5:30learned how to orchestrate the whole
5:32thing. So you get to a point right now
5:34where you are not sitting and doing
5:36anything any of those things.
5:38>> You're sitting and taking decisions.
5:40>> True.
5:41>> You're using your judgment to take the
5:42right calls.
5:44>> You have a flare of understanding.
5:47>> Agree.
5:48>> That is taste. Knowing what to do, what
5:50not to do is the decision-m that you
5:52build on top of.
5:54Now the problem with this is you can do
5:57it very very well. But someone who's
6:00just getting started who's using AI
6:03imagine you would have started this
6:04company by hiring someone to write
6:07YouTube script to write titles to think
6:09of what the copy is and everything and
6:10you just be the actor. Do you think your
6:12uh company would have been this big?
6:14>> That is exactly the mistake people are
6:16making today with AI just because not
6:19that they could not have done they could
6:21have done it. You could have hired very
6:23good people from everywhere and they
6:24could have done it and a lot of people
6:25do it.
6:26>> That is the difference between a brand
6:28and that's the difference between a
6:30creator.
6:31>> Creator mindset you are always in a
6:33founder mode. So the problem with this
6:36is someone is getting started who's
6:39getting the seeing the flare of AI you
6:42just letting it do everything end to end
6:43without even knowing why it is doing
6:45what it is doing gets to a point where
6:48AI is taking decisions for you. So
6:50here's my problem as an entrepreneur
6:52which is we you picked up spot on
6:54because of all the content that we make
6:56you come here you teach me AI and I get
6:58blown away and I get into like
7:00>> this geek mode of learning everything
7:02what you said whatever is relevant for
7:04me for next 2 3 4 days and then I go
7:06deep dive and then I ask and probably
7:08force my team to start using air
7:10>> perfect [clears throat] so now since
7:12last one and a half year I've made them
7:14do it
7:15>> correct
7:15>> and now I've realized some of them have
7:18become dumber
7:19They were much smarter before I gave
7:21them AI.
7:22>> So, and I'm like, I don't want this chat
7:26GPT clot stuff. I want your original
7:29thinking and there's no original
7:30thinking anymore. And I'm sad about it
7:32because my [clears throat] team members
7:34are becoming dumber and dumber. I
7:37started becoming dumber in middle and I
7:39shut down. And I've stopped now using by
7:41the way the way I would use
7:46all of these AI stuff for my questioning
7:49for my research and stuff like that.
7:52I've reduced it like to 14th
7:55>> of what I used to do because it was
7:57making me dumber. It was giving me just
7:58same kind of things which anybody could
8:00have done and that's what's happening in
8:02the team. Everybody no matter what task
8:04I'm giving
8:06they are giving me dumb stuff boring bad
8:10stuff. So how do I know a line between
8:13what how much should I let my team or
8:16myself should I use AI versus how much
8:18should we not?
8:18>> That's why I'm saying no
8:20>> do the work that [clears throat] was low
8:23value for you all throughout which was
8:24operational which was if this and that
8:27>> and get that done by an AI. The last
8:30layer of work is something that you have
8:31to focus on, right? Eventually options.
8:36Yeah.
8:39What is that I want to take and why and
8:41build on top of that? It's a 20% layer
8:44that is left right now. If you don't try
8:47to touch it, that is what happens. You
8:49get average outputs. I'll tell you
8:50there's a study by anthropic. Uh I
8:52forgot what it is called. uh the u it's
8:56it's a study on agents working with each
8:59other
9:00>> okay there are multiple AI agents
9:02>> when they work with each other agent
9:04orchestration
9:08I was reading it and it had something
9:10very evidential in this when a bunch of
9:13AI agents from same anthropic were given
9:16the same task
9:18>> of writing a book or something okay
9:20>> four out of 10 agents came up with the
9:22same book name
9:24some socialistic something I don't
9:25remember the exact your team can pull
9:27and put a screen on it okay put it on
9:28the screen
9:29>> what does that say
9:32all models are thinking alike of course
9:34there
9:35>> there are different way of solutioning
9:36that you can bring in to get
9:38perspectives I use different tools like
9:39multi and all to bring different AI
9:42models to
9:43>> discuss differently because everybody
9:45has their own biases so I want to know
9:47everybody's biases that's how you
9:48operate technically to get the best
9:50answers but it's happening
9:52So when the output is similar for
9:55example we working on a project
9:58let's think about names everybody had
10:01come up with 10 names three to four
10:03names were same in everybody's paper
10:05>> because everybody used the same bloody
10:06AI models
10:08and gave the same prompt
10:09>> that's what was happening
10:10>> yeah so how are you tuning that
10:13conversation by adding more context
10:16so here's what do you mean by adding
10:19more context because here's what people
10:21are At
10:22least in my org which I've seen
10:25>> they ask AI to do deep research work
10:28>> ask them to come up with like 50
10:30questions then I have given my own
10:32process of how I finalize topics how I
10:34finalize research and then come up with
10:36questions and stuff like that it's a big
10:38it's a 30page process because I've
10:40written down everything right and I've
10:42written like I've genuinely done it and
10:44I've given that to them what they've
10:46done they've like now based on Raj's
10:50process choose the questions. [laughter]
10:54So the last judgment layer earlier they
10:58were
10:59watching hundreds of things coming up
11:01with 100 questions but they were the one
11:03taking decision what are the 10
11:04questions which are which should reach
11:06me just to example what is the right out
11:10of 100 scripts what are the two scripts
11:12which should reach me
11:13>> now they are dumping those 100 scripts
11:16and 100 things on AI and asking cloud
11:18GPD do you choose and that is choosing
11:20something
11:21>> and I can read and tell that this is Not
11:24you.
11:25>> Yeah.
11:26>> So they're letting AI only decide
11:28>> take decisions.
11:30That is I you didn't you didn't lead me
11:33to this. I I led you to this
11:35>> where I literally told you this is what
11:36scares me the most.
11:38>> I keep saying this to my team
11:39>> that people will keep getting dumber.
11:41>> They I don't know how to put it like I
11:44don't think net net their number.
11:46>> No. So [clears throat] people are doing
11:47average work. Let's just say that people
11:49are just keeping
11:50>> there is a new average right now. H
11:52>> the new average is AI slop.
11:54>> Ah there
11:56>> that's that's a new average. It's better
11:58than the last average but it's just an
11:59average again.
12:01So at this point of time like when
12:04internet came in everybody has the
12:06information same information that you do
12:08right because everybody can search.
12:10>> When AI came in everybody has a smartest
12:12agent right next to you. If a I don't
12:15know if a Sachin Tandulkar while playing
12:18cricket would have gone to a terrible
12:19coach do you think he would have become
12:21a sachinder? No. Right. So that is what
12:23is happening right now. You everybody
12:25has a sachinda. Now you have to be a
12:26very good coach.
12:28If you can't be a good coach that person
12:31will never make it into cricket. I mean
12:32I will not name a couple of other
12:34cricketers who had the potential could
12:35not make it but for whatever reason they
12:38were not guided the right way. So, how
12:41should I tell like what you said it's
12:43about context? How should I give my AI
12:45more context in a better way to get
12:47better results? Not average work.
12:51>> Yeah, it is not
12:52>> AI is giving me average That's I'm
12:54just going to go out and tell you that's
12:56that's become the problem. Maybe I'm not
12:58using it right away.
12:59>> I think your baseline has shifted. Your
13:01expectation from AI
13:03>> has gone up significantly over the
13:06course of time because you're seeing the
13:08potential of it.
13:09But I've seen it in data. I'm talking
13:12about data. Like I've tried question A
13:15with X guest which is written by me.
13:18Question B with B guess which is written
13:21by AI. Question A has a better spike
13:24than version B
13:25>> and multiple times. So with probably
13:29similar guest, similar options, bunch of
13:31other places and we experiment with
13:33hundreds of pages.
13:34>> What level of context does AI have?
13:37>> A lot. How why do you think it has a
13:40lot? Does it have data of every piece of
13:44content that you have consumed?
13:47every piece of content that I've
13:48consumed. Why do you think you come up
13:50with ideas the way you do smart
13:54content?
13:57>> None of us are smarter than an AI. That
13:59is very clear based on data.
14:01>> Agreed.
14:01>> Right. Your context is different and
14:04based on your context, you're a
14:05specialist in something because of which
14:09you're able to come up with feelings and
14:11gut and taste. That is what AI is not
14:14able to pick up. and you're like you're
14:15comparing that to this for I'll give you
14:17one simple example right when you're
14:20let's say when you knew that
14:24you must have passively been consuming
14:26something
14:31it passively happens right like
14:36you're doing that passively you're doing
14:37your mind is already working in those
14:39lines if you're meeting a president of a
14:42country you're actively reading about
14:44your mind is automatically working on it
14:46right but AI is on silos
14:49>> the moment you ask question it wasn't
14:51zero it becomes one
14:53>> it's trying to become one so it is
14:55trying to compete to with you who has
14:58insane amount of context and that is not
15:00the only context what is context how do
15:02we take decisions let's take two steps
15:03back one is a close proximity layer
15:11that is one two is what have I been
15:14doing over the course of last 6 months
15:15to one year to bola I'm able to come up
15:18with better questions of course you'll
15:19be able to come with better questions
15:21because you're the guy sitting and
15:23asking the questions before even the
15:26data goes out before even the podcast
15:28goes out you have the taste of knowing
15:31podcast
15:34you have the taste I know after this
15:36podcast you go and say yeah
15:39because I've been with you after this
15:40podcast you have it running on your head
15:43because You have that knack.
15:46>> AI doesn't have it because AI has not
15:48sat next to you to do all these
15:50podcasts. But can it? Yes, it can.
15:54>> Can it get close to it? Yes, it can.
15:56>> How?
15:57>> That is by giving it context. For
15:58example, try to note out. You remember
16:01we had built a skill for research back
16:03in the day. That was only one part of
16:05it. What is a human AI employee?
16:08>> I want a AI employee. For an AI
16:10employee, what all do we have access to?
16:12You have access to a memory which is
16:14yours. [snorts]
16:15>> You have access to tools which is your
16:17computer this that and all. You have
16:19access to skills and SOPs. Some are
16:22built out mentally. Some are built out
16:24on paper.
16:25>> M
16:25>> right. And four is you basically take I
16:28mean you have a brain which thinks
16:30through all of these vectors and a few
16:32more to take decisions. Today AI has all
16:35of them.
16:37What we are doing is we are not using
16:38the tool well enough to get the actually
16:41it is smarter than us
16:43>> and I have instances to prove it also in
16:45our cases but it's also dumb in a lot of
16:48places we can talk about that also it's
16:49not there
16:50>> where we want to but if I have to if my
16:54job is to get an AI to work at the level
16:56that I operate or get it close to me it
16:59will not happen overnight it'll probably
17:01a 12 18 month a 12 to 18 week journey
17:04for me to fine-tune it but I will do
17:07everything possible for AI to get
17:09exposed to what I'm exposed today. How
17:12will I do it?
17:14>> One, I will build a second brain.
17:17>> What do you mean by that?
17:18>> What is a second brain? Very simple.
17:20What are the things that I consume on
17:21every single day level? I consume
17:23YouTube.
17:24>> Mhm.
17:24>> There are cons there are things that I
17:26consume that I don't want AI to see. I
17:28don't want it to see that I was watching
17:29some Netflix show.
17:30>> It has no relevance to the work that I
17:32do. That's entertainment. So, I'll
17:33probably keep it aside. I'm sure there
17:35are correlations there also.
17:36>> Absolutely.
17:37>> But I will keep that aside for now.
17:39>> Yeah. The movies I see I have I learned
17:41so much from you.
17:42>> Yeah. For you definitely. Yes. I can
17:43imagine. Right. Uh let's say YouTube
17:46content that I'm watching
17:47>> every day is very very valuable. At
17:49least I can say for myself I consume a
17:51lot
17:52>> podcasts and you know we spoke about lex
17:55fitments of the world and all I consume
17:57a lot. I build lot of perspectives from
17:58that.
18:00>> How do I feed the same context to AI? So
18:02I built a basically a daily rapper which
18:06opens my YouTube history
18:09goes through all the videos ranks
18:11[snorts] all the pieces of content. If
18:12I'm watching some I don't know like Mr.
18:14beast video or let's say if I'm watching
18:16some health video of some podcast of
18:18yours or whatever it'll ignore all of
18:21them because I'm focused on AI
18:23>> right everything that I've seen of AI it
18:25pulls that data
18:28>> converts the transcripts and looks at
18:30what are the key pointers saves it in
18:32form of JSON cards JSON cards what is
18:35the value pack of each one of the
18:36podcast is what it saves that is one
18:39[clears throat] two is that is not it
18:41lot of times when I watch content on AI.
18:45I share it with my team.
18:47>> Where do I share it? I share it on my
18:48Slack. To the content team, I might say,
18:50"Guys, this is very good perspective. I
18:52might have dropped a voice note to my
18:55programs team because we teach AI a
18:56lot." I must have shared someone saying
18:59that I like this SOP. We should teach it
19:01to our learners to our implementation
19:03team who's probably implementing
19:04something at Motorola right now. Let's
19:06say I found something technical there
19:07and I think in the project that we
19:09working with Motorola, this could be
19:11useful. So I'll send it to them and all
19:13of these things are happening on Slack
19:14for me.
19:15>> So I have a triage running for my Slack
19:19which looks at what are the
19:20conversations I'm having
19:21>> actively. This is number two. Three, I
19:24realize one of the highest value
19:25conversations that I have today is
19:28inside of standups with my team. Every
19:31single meeting is transcribed put in a
19:34GitHub repository.
19:37GitHub is where people push code, right?
19:40is on a GitHub repository and inside of
19:43that repository my team can access that
19:46information at any point of time because
19:50this this by the way started very
19:51recently. I said when we were creating
19:53content why why are our perspectives so
19:56stronger when I'm with Raj
19:58>> but when we are shooting content you ask
20:00me a question
20:01>> because
20:05how do we bring the flow that I get with
20:07Raj we started daily standups
20:10>> in daily standups we talk about AI
20:12topics I give my perspectives there
20:14because
20:19we are not able to speak a lot of things
20:21So those things are transcribed. Those
20:24are easy. You can use a whisper flow,
20:25granola, fireflies, whatever
20:27>> those notes are there.
20:29>> Okay.
20:29>> Right. So all the finally, oh my god,
20:32how can I forget this? This is a game
20:35changer which is it is a little risky
20:38also. Every conversation that I have
20:41with Chip, Claude, Gemini, Grock, cursor
20:47sometimes everything every day is
20:50exported and fed to AI
20:54because my raw thoughts [snorts] are not
20:56having with or are not the conversation
20:58that I'm having with you are not the
21:00conversation that I'm having with my
21:01team are the conversations that I'm
21:03having with my AI.
21:04>> My perspectives are there. I'm a huge
21:07voice mode user. Okay.
21:10When I'm in the gym and all just
21:13brainstorming, I I treat AI like a
21:15brainstorming partner.
21:17All those perspectives are fed into a
21:20single memory layer called as Cognney.
21:23>> Okay.
21:23>> Cogni is a memory layer.
21:31JSON bits save.
21:33>> Okay. As a result, next time I want to
21:37do anything,
21:39I can say tomorrow I'm having a
21:41conversation with Raj. These are the
21:43podcasts that we have done. What are the
21:45strong perspectives that we should put
21:46in the podcast that was spoken the last
21:483 months? I have all the data ready.
21:51>> Nice.
21:52>> So, and on top of this, let's say
21:54tomorrow I want to write a important
21:56email. Forget about all that email
21:59whatever email everything all that data
22:01is basic.
22:02Tomorrow if I want to take an important
22:04decision today, tomorrow I want to meet
22:07someone
22:08>> and I want to know what questions can I
22:10ask them.
22:11>> I might not remember exactly what I
22:13could have asked them which was a
22:14question of mine. But I could have asked
22:16AI that.
22:17>> It would be like oh you're meeting Alex
22:18Wong right next week. You should ask
22:21these three questions because we debated
22:22about these three questions. Our
22:24perspectives were different. He could
22:26give us a very different perspective.
22:28when when I was uh by the way I did a
22:32podcast I hosted a podcast where a
22:34couple of people commented saying that
22:36web has become Raj Hammani right now
22:38with you know right the conversations I
22:40do with open AI and all those right so I
22:43was speaking with Wolfie before the
22:45conversation happened my prep work was
22:48when I sat in the car
22:50>> with this data asking this is Wolfie
22:53Bane he heads startups for open AI I'm
22:55having a conversation with him uh coding
22:58New codeex is launching because we got
23:01to know what was launching whatever
23:02right so that was a conversation
23:03>> so what are the questions that I had
23:06apprehensions that I had things that are
23:09people are asking me you pull all that
23:13data and tell me what is it and it had a
23:16report ready no team can beat this
23:19>> no AI can beat this because this is you
23:22[snorts]
23:23this is context now this is where you
23:25draw a line do you give this access to
23:27your team or do you keep this to
23:29yourself? I'm not given access to
23:31everything to my team. Anything which is
23:32public is public. For example, meetings
23:35that is happening with the team that is
23:36available for the team.
23:38>> But personal conversations,
23:39>> personal AI conversation channel is very
23:42much limited to me.
23:43>> Okay,
23:44>> that is personal context,
23:46>> right? And this is what I mean by an
23:49overall context. This is your decision m
23:52every day.
23:53>> It is one time setup. Now you set it up
23:55and leave it. Now sadly this is the
23:57problem right that is running on a
23:59computer at my home and that computer is
24:02dedicated for my AI agent
24:05>> AI agents not AI agent AI agents right
24:08>> but then don't they get context rot
24:11>> very good question context rot that is
24:14why I don't use a direct LLM memory
24:19I use something called as cogni
24:21>> I didn't say I save all of this inside
24:24of Uh chat GPT or context
24:28context window
24:30so
24:32okay so the yeah go ahead I got
24:35something but go ahead
24:36>> no go on go on
24:37>> no no no so it's like you're telling me
24:38that the cognney is like the master
24:42memory it's not a per chat memory and
24:44per in per chat context gets the context
24:47rot right
24:48>> no no no no no so basically what is
24:50happening is we have llms [snorts]
24:53right this is your uh GPT claude etc.
24:59>> Okay.
24:59>> Uh and these are all LLM layers. What we
25:02have done is an agent
25:05is like a human.
25:06>> M
25:07>> in the simplest way put an LLM is the
25:10let's say mind.
25:12>> Okay.
25:13>> The thinking part.
25:15>> Okay.
25:15>> That is what for an for a this is like
25:18your brain dude.
25:19>> Okay. This thinking thinking
25:21>> then it needs tools. H tools like let's
25:28say web search uh MCP MCP is basically
25:32using your slack
25:33>> got it
25:34>> etc. Whatever tools it uses this is tool
25:37use.
25:38>> Okay.
25:38>> Right. And then there is something
25:40called as memory.
25:42>> What is memory? To remember everything.
25:46>> So whenever you ask a question, I mean
25:48there are more layers to this. Whenever
25:50you ask a question, you're using memory
25:52to understand what is in there. If
25:54there's something that you can retrieve
25:56based on habits, you're using tools to
25:58do a job. And LMS are your thinking
26:00part. H now in these LLMs there's also a
26:05layer of memory which is basically
26:07called as in a simplest way put context
26:10window h
26:12what is context window so humans I don't
26:15know if you know this we can only
26:17remember like seven digit things the
26:20best way possible in a way passively
26:25seven digits is like our uh context
26:28window of some sorts
26:30>> okay what do you mean by Seven digit
26:31>> anything seven digit so all the numbers
26:34and all that you see right
26:35>> what now let's say how do I put it
26:37that's the amount of things that we can
26:39remember very very well
26:40>> okay
26:41>> right but an LLM can actually remember
26:45context of up to 1 million right now LLM
26:47is like 1 million 1 million is 1 million
26:49is like I think uh let's say five large
26:54novels it can remember at any point of
26:57time digit number okay 765 whatever I
27:02told you
27:08and you what you will try you'll try to
27:10make sure you're not remembering
27:12anything else
27:13>> that will stay in your memory that is
27:15imagine that to be a context window
27:16>> got it
27:17>> till I ask you remember the moment I
27:19tell you remember this you'll probably
27:21forget the last one
27:22>> that is the human brain
27:25>> but for a LLM today it can remember 1
27:28million
27:29It can pretty much have five novels open
27:33and you can ask
27:34>> 18th page novel one fifth word it can
27:37tell you
27:38>> that is the level of brain.
27:40>> Okay.
27:40>> Now most of the context used to revolve
27:42around this
27:45was regular context window. This was
27:47regular memory and context.
27:50>> Okay.
27:50>> But
27:53you're adding a secondary memory layer.
27:56Imagine this to be like a notepad. M
28:07starting index numbers to remember from
28:10web numbers to remember of YouTube. You
28:13made a index of it and that is your
28:14memory layer. So I'm not pushing all
28:18this context here. I'm not giving it all
28:20to Chad GP saying
28:23because
28:26it will feel like a lot but it's not a
28:28lot.
28:29>> Imagine if you watch 10 YouTube videos
28:33which is a podcast like you and me do
28:36that's 10 hours of content.
28:39It will go over context window easily.
28:41So memory is beautiful because what a
28:43tool like Cogni does is it uses graphs.
28:46[snorts]
28:47Okay, it uses graph memory and it
28:50retrieves the same set of memory which I
28:52need to do this task right now
28:56>> and it only requires like it'll
28:59only pick up a book that it needs at
29:01that time.
29:01>> Yes. So it has built a beautiful index.
29:05>> M
29:06>> okay. Every time when I give a task
29:09saying, "Hey, write an email for me or
29:11whatever,
29:12>> it'll first go to an LLM lm context or
29:15agents.md
29:18which is basically a markdown file or
29:19instruction file which will have a
29:21direction saying look
29:31Okay.
29:39And this is basically how an agent
29:40operates in the simplest possible way.
29:42I've simplified it. And for you to use
29:45agents like this,
29:46>> you need something called as a agent
29:48harness. You're harnessing an agent.
29:52And these tools today are the ones that
29:55we were using who have evolved from
29:57being an assistant to an agent harness.
29:59A chat GPT has become a chat GPT work
30:03plus codecs which are basically agent
30:05harness. A claude has become claude
30:07co-work and claude code. These are agent
30:10harness tools which has access to memory
30:13which has access to tools which has
30:15access to brain which has access to a
30:18few other things and it orchestrates
30:20everything together to execute your job.
30:24>> Now let's say this doesn't exist at all.
30:28>> Yeah. and you say webarch
30:32it is as good as a fresh intern who is
30:36bloody smart
30:40context how can he compete with you or
30:43he and she compete with you isn't that a
30:45wrong comparison to make to start with
30:47>> true
30:48true
30:50so the memory is a game
30:52>> context is everything it's everything
30:54and on top of that look in this also
30:56there's one more layer right when you
30:58build your agents, you build your
31:00skills.
31:02>> We you remember we built the skills back
31:04in the day. Skills have become much much
31:06better right now. Our whole company
31:08operates out of skills right now. That's
31:09the proprietary thing that we have in
31:11the company at this point of time. How
31:13do I translate my brain into skills? How
31:16can everyone translate their brains into
31:19skills and make those skills better
31:21every single day with new things that we
31:23are learning where it gets to a point
31:25where it can truly do 90% of the things
31:28much better than you can do.
31:32That's the game you play.
31:34>> And skills are a bunch of markdown files
31:36>> which is all that we made last text
31:38document. It's like a text document. It
31:40looks like a instruction document. Step
31:42one, step two, step three, step four. I
31:44have skills for everything. No,
31:46>> that's the whole point. That is the
31:48whole point. But there are a few layers
31:50that are missing right now. There's a
31:51memory layer that is missing. Tool call
31:53layer that is missing.
31:59Slack.
32:00I've just attached my appy
32:03>> as to use whatever unlimited go
32:07>> made bunch of skills and then doing it.
32:10context layer is missing and that's why
32:12it's giving me dumber and dumber
32:13>> because till now
32:15>> I'm doing everything
32:16>> till now we always used to think this
32:19layer
32:20>> was actually very
32:23>> not important is like
32:24>> no no it was always important it was not
32:26solved
32:27>> okay
32:27>> memory management in a way was not
32:29solved it's getting much much better and
32:32on top of that context windows are also
32:34very very short it is getting much much
32:36better right now it's going to a million
32:39right No by default it has become a
32:41million. So by default it has become
32:42five books.
32:43>> Nice
32:44>> right? So you have to think about
32:46>> this is how you build a second brain and
32:48then
32:49>> this is your second brain. This only
32:51this bit is your second brain.
32:52>> And when you do that with your process
32:54[clears throat] that is how I get AI to
32:58start thinking probably much better than
33:01me in the context that I am thinking.
33:05That's
33:05>> it gets to think like you. it it you
33:08make it your own [clears throat]
33:11>> more information as well right like at
33:12the same time I can't remember that much
33:13information so it
33:14>> that's fine that's fine you anyways
33:18can't remember because your brain is
33:19this seven digits
33:20>> m
33:20>> you anyways couldn't
33:23>> so this and then you build your second
33:25brain and that's how your all your
33:26content all your everything comes better
33:28>> all context yes it gets better every
33:30single day like I said you right you
33:32will see this graph okay of quality of
33:34output that you'll see if this is time M
33:37>> this is quality
33:38>> m
33:40>> this is how you will see most people
33:42will give up here
33:45saying because this is going to take you
33:47like 3 weeks this journey will take you
33:49seven more weeks but this is where
33:51you'll see the exponential results sorry
33:52I'm making it
33:54>> this is where you'll see exponential
33:56results so most people I believe give up
33:58before you get here
33:59>> yeah because they're like oh I'm trying
34:01so much and still not working
34:03>> it's not working that's the point right
34:05but that is Third differentiation when
34:08everybody has the same bloody tool.
34:10>> But tell me point blank question is
34:15when you get a script
34:17>> to write a real for example
34:21>> is your judgment better or AI
34:24today
34:25>> even [clears throat] after doing all of
34:27this is my judgment better or AI?
34:29>> Yeah like you I can give you 10 scripts
34:31I can give
34:32>> I don't let AI make a judgment only.
34:34I'll tell you how I use AI.
34:35>> But it will give you one or two things
34:37that you have to choose.
34:37>> I'll tell you I'll tell you how I use
34:38AI. That's not the question that I ask
34:40to start with. I don't expect it to do.
34:42That's my job.
34:44>> If I'm creating content,
34:46>> my job is to put out something that is
34:50valuable for people. So I decide I'm the
34:52decision maker. AI is my
34:54>> helper
34:54>> team [snorts]
34:56who's going to help me get there better
34:58to serve my audience better. So AI's job
35:02is to so I go if you're talking about
35:04content right one of the reasons that we
35:06win I believe or we've been winning in
35:08short- term uh short form content and
35:11also long- form content is taking very
35:14datadriven decisions. Okay, let's make
35:16your datadriven AI decision maker or
35:21whatever you're doing right now your
35:24AI content system I want to understand
35:26how many views you're getting per month
35:28>> maybe 70 80 million views across
35:32everything 10 million will be Twitter 15
35:34million will be Twitter
35:36>> Twitter is rampant 10 million is crazy
35:38>> 15 millionishabi
35:40>> you're putting res there also
35:42>> no only Twitter is this conversation
35:46>> brain dumps
35:47>> people either like it or don't like it.
35:49The same since the back in the day it's
35:52the same.
35:52>> Okay. So let's let's make your system
35:55right here.
35:56>> 100 million view system.
35:58>> 100 million views system. Okay. I don't
36:00think we're hitting 100 million every
36:01month yet, but I think uh we could be
36:04averaging here on a yearly level because
36:06there could be appreation months.
36:09>> Dick the way it kind of works for us.
36:12>> This is 100 million view content system.
36:14you are teaching us.
36:15>> Yes.
36:15>> Through AI what you are doing because
36:18you tell me you don't create content
36:20much.
36:22>> So I don't create content at all. In
36:23fact there is a joke uh in our audience
36:26webhub only talks in Raj's podcast.
36:29everywhere [laughter] else it's an AI
36:31and it's not and and I'm not saying you
36:33you should read the comments, [laughter]
36:35>> right? like web
36:39you should come you [laughter] should
36:40come because on my channel it's AI it's
36:43my AI clone but look
36:45>> here's the problem most people think we
36:48are [clears throat] doing 100 million
36:49views because I use a AI clone
36:54>> this is what people think
36:56>> I'll tell you the reality if I was not
36:58using AI clone this would have been 200
37:00million that is something that you need
37:02to understand
37:03>> like if you were not using
37:04>> if I was not if I was shooting the
37:06content I would have made 200 million
37:07views.
37:08>> Oh,
37:08>> but
37:11for the life that I live I don't have
37:14the time to shoot enough content for
37:16even half of it.
37:17>> So you're telling me real face, real
37:20person will always like till now
37:21>> there is flare. There is flare.
37:23>> Okay.
37:24>> Of course there is flare.
37:24>> So a real person will drive way more
37:26views than an AI person.
37:28>> Yes.
37:29>> Right now it's the thing or is it going
37:31to stay like that for I think platform
37:33is going to incentivize real people. We
37:35will see what happens. Uh I think
37:37platform incentivizes only one thing and
37:39that is retention and engagement. It
37:40doesn't care. So far it doesn't care
37:42till people re retaliate and they're
37:44seeing that in consumer behavior
37:47>> because platforms does what is needed
37:48for the platform not for you not for
37:50users.
37:50>> Agreed.
37:50>> Right. But anyways that's a
37:52philosophical conversation as well that
37:53we can debate on too. But what I'm
37:55trying to say is most people think I'm
37:58winning
37:59because of AI.
38:01>> And the [clears throat] answer to that
38:02is what they believe is AI clone.
38:05>> I would want to start by saying this.
38:07This is not the answer.
38:08>> Okay. What's the answer?
38:09>> The answer for winning is a process of
38:12how we think with AI
38:14>> where
38:17every time I try to set up a process
38:20across company. Here we are going to
38:22talk about content example right now
38:24across the company. The biggest mistake
38:26people do is they will see what is a
38:29human doing right now where all I can
38:31plug AI so that it can be done better.
38:34That's a fundamentally wrong process.
38:36>> Okay,
38:36>> you have to reimagine that whole process
38:39with capabilities of AI in the mind.
38:42That is when you build a great process.
38:44For example, for content, what is the
38:46first step for someone like us who
38:49create education content? One is
38:53topic selection.
38:54>> M
38:55>> okay.
38:56>> This plays a very important role.
38:58>> This is game
38:59>> game. No, topic selection is one. I'll
39:01tell you what game is. Game is this
39:05packaging.
39:08>> What we select as a topic here? The
39:10topic is we'll talk about AI. Okay. I
39:13mean you know packaging really well but
39:15you get what I'm saying.
39:16>> Yeah.
39:17>> Second is
39:17>> for me topic is packaging.
39:19>> Okay. For me topic is topic. What am I
39:21choosing?
39:22>> The theme. You me you meant theme.
39:23>> Okay. You can call it theme
39:25>> because topic for me is the title. It's
39:28the absolute bang on packaging title. So
39:30probably I'm thinking in a different
39:31>> correct. So you you do you tell your
39:33stuff. Okay.
39:363.8 3.8. That's a theme.
39:39>> No. How is it a theme?
39:40>> Like for me that's a theme. No, I I have
39:43options.
39:44>> I'm saying that in my head I I think of
39:47it like a theme.
39:48>> I think like that as a thing for me the
39:51>> So what what is theme for you is a
39:54category for me. So let's say your theme
39:56would be today we'll talk about AI. For
39:57me that's
39:58>> theme for everyday is AI.
39:59>> Huh. So
40:00>> that's why [laughter] one level down
40:01>> exactly. So I'm like that's
40:03>> I'm curious to know what people think is
40:05the right thing. Topic selection or
40:07theme? Let's see. I want people to tell
40:09us in the comments.
40:09>> No, everybody can have their own names.
40:11I'm just saying my name. My brain works
40:12like that. There's a category, there's
40:14theme, and then there's
40:14>> that's your lingo.
40:16>> That's my language.
40:17>> Topic selection.
40:17>> You let's do yours. Okay.
40:18>> Topic selection, packaging on how I'm
40:20packaging it.
40:22>> Then comes script
40:25>> and then comes uh posting.
40:30>> Okay.
40:30>> Okay. I'll get to each one of them on a
40:32>> and then data and then all of that.
40:33>> So all will come in this that becomes a
40:36loop. Now I feel most of the game is one
40:39here.
40:41Topic selection and packaging. That's
40:4380% of the game.
40:44>> Fair. All right.
40:45>> Fair.
40:46>> Agreed.
40:47>> Yes, you will agree. I know
40:49>> topics. So, how do I do an incredible
40:51work
40:52>> at finding the right topic
40:55>> normally?
40:56>> Okay. So if I have to create content
40:59>> on AI
41:00>> or health or anything for that matter,
41:03the first thing that we'll think about
41:04doing or you should think about doing is
41:07hey what is that that the consumer
41:10outside who's consuming content every
41:12single day giving me as a signal.
41:15>> What are they consuming more? What are
41:16they liking more? What are they engaging
41:19with more? What is driving me better
41:21data? On a human level, what will we do?
41:24We will have daspandra profiles. our
41:27feed our sc our feeds will be optimized
41:29for that.
41:30favorite stuff and we'll use that data
41:33and when you put a layer on it, what
41:35will you say? You scroll the feed for
41:37me.
41:37>> Yeah.
41:39>> What will you say? You will say go and
41:40read this people and tell me. So you'll
41:43limit yourself. But with AI coming into
41:46the picture, I have unlimited employees
41:50>> with unlimited human labor that I have.
41:52How will I think about topic selection
41:53now?
41:54>> Now I will go bonkers. Okay,
41:57>> I will look about every single source of
42:01information that is related to AI is a
42:03signal to me.
42:05>> Every single.
42:06>> So it could be now when it comes to
42:08topic selection if I break this down
42:09right just this for me on AI it could be
42:13something like product hunt.
42:15>> It could be something like X [snorts]
42:17viral tweets.
42:18>> It could be something like real other
42:20reels
42:21>> or Tik Toks
42:22>> in other countries. It could be YouTube
42:24videos.
42:26It could be podcast
42:28in fact
42:30it could be research papers
42:35>> news there are etc. It depends on the
42:38topic that you're doing.
42:39>> Yeah.
42:40>> Everywhere this piece of conversation is
42:43being debated about
42:44>> I need to know.
42:46>> Yeah. Articles
42:48nothing. I will leave nothing.
42:51>> I want to know what's happening.
42:53>> A human cannot do. Yeah,
42:55>> Reddit massive source
42:57>> heavy
42:58>> kora massive sources what's happening
43:00inside of small slack circle school
43:04communities
43:06very important I don't have
43:08>> I would not even imagine of going into
43:10all of this because I don't have 100
43:12employees to do topic selection
43:13>> true
43:14>> but I have AI agents so now I have
43:17nailed down on the sources that I want
43:19to tap into
43:20>> okay
43:21>> I will deploy AI agents for everything
43:24whose job will be wake up every single
43:26day or do it every 3 days. For example,
43:30product hunt has become one source. Xia
43:33X feed is one source which is my feed.
43:36So an AI wakes up four times a day for
43:38me, opens my Twitter and scrolls on my
43:40behalf and pulls everything. But that is
43:42only one side of Twitter.
43:44But then there are 100 other people who
43:46create content on Twitter.
43:47>> Okay,
43:48>> on AI AI is getting all that information
43:50as well. topic based cluster everything
43:54related to AI some 100 keywords that
43:56information just inside of Twitter reals
43:59same logic
44:01uh YouTube same logic podcast same logic
44:04what real today in last 24 hours or 48
44:08hours if a real on AI has blown up in
44:11any language I should know
44:14>> and how do you what's the easiest way to
44:17make an agent who would do all of this
44:19>> to build something like this I told you
44:20there is a You have to build an agent
44:22for it.
44:22>> No, but then we want to do every this is
44:25only topic selection.
44:27>> Topic selection.
44:29>> Okay.
44:29>> In topic selection, we've built
44:31something called as a triage system to
44:34get all the ideas in one place.
44:36>> That's all we have done.
44:38>> We have not even gone to the packaging
44:40part. I'll get to let's go to the next
44:41step.
44:42>> Okay. But how will we get all of this?
44:44>> How to get like how do I make so that I
44:47get all of these things?
44:48>> Okay. So now how do we build this system
44:51where it will pull all the ideas where
44:53it will pull all the ideas back in
44:55>> well uh I don't know if you can see my
44:58screen there is something you have to
45:00pick up an agent system right now for
45:02this
45:03>> okay [snorts] and agent system is
45:06something like Hermes that we already
45:09dabbled with sometime
45:10>> last podcast yes
45:12>> or something like openclaw
45:14>> which are again agentic systems
45:16>> that you can use here and these are
45:18trade solutions if you want to do it for
45:20free.
45:20>> Okay.
45:21>> But if you have $200 to spend per month
45:24and you don't want to get technical and
45:26you want it to be way more reliable,
45:28then recently there is something called
45:30as Grogbot which is by Elon Musk has
45:32come out which is actually quite good.
45:34>> Okay.
45:35>> Okay. These are like team of agents
45:37designed for nontechnical people to be
45:39able to do a lot more.
45:40>> Okay.
45:41>> Right. You can use any of these systems
45:43to be able to build this. For example, I
45:46want to go with a free solution. Of
45:47course, you can do Grogbot and all. Uh,
45:49I want to go with a free solution. So,
45:51I'll go with Hermes.
45:53>> Okay.
45:53>> Right now,
45:54>> Grogbot is what? Easiest.
45:55>> Easiest. It's the easiest, but it'll
45:57it'll cost you $200 a month.
45:59>> But what is it? Like, they're just all
46:01agents and I can give different agent
46:02different job to do.
46:03>> Yes. So, I'll show you a simple example.
46:07I personally don't use Grogbot. Why?
46:09>> Right?
46:10>> Because I have a much more sophisticated
46:12system than this. I [laughter] don't
46:13need Grogbot. Uh but again this is
46:16Grogbot the bunch of bots you can see I
46:18have a partnership inbox trading radar
46:21content OS my WhatsApp injection engine
46:24>> uh for every it's like your team
46:26>> okay
46:26>> it's like your team you can it's like a
46:28slack this is Grogbot
46:29>> and every person has a different job to
46:30do
46:31>> correct for example this is Watsi what's
46:33job is to go check my WhatsApp
46:35>> Mhm. every three hours and tell me
46:37what's in there. Look at this. Right.
46:38And the other advantage of Grogbot is
46:41that this is subscription tracker. It
46:43basically tracks all my subscriptions on
46:46my email every single day. And
46:49so it tell me, bro,
46:53so any problem that I have, I've given
46:54it to an employee.
46:56>> Nice.
46:56>> Right. But this is me testing. All
46:58right. So this I I I was trying because
47:01>> And you can connect everything to this
47:02like your WhatsApp, email. connect
47:04anything to everything. [snorts]
47:06>> You can connect anything and everything
47:08all over the place. Right? So this is
47:10Grogbot. Now I'll just tell you because
47:11we've opened Grogbot. I also want to
47:13talk about the advantages of Grogbot.
47:15Right?
47:15>> One advantage of Grogbot is that the
47:18advantage of Grogbot is that
47:19>> every bot inside of this has its own
47:24computer. You can see this on your left
47:25side, right side. Now I've opened it
47:27right now. Every Grock bot has its own
47:30computer.
47:32because it has its own this is the
47:33biggest advantage that I see so far
47:35>> because this h it has its own computer I
47:39can let's say I have a subscription
47:40tracker it only has access to my email
47:45and has access to nothing else
47:48>> I have basically context because it's
47:51not connected to a different memory
47:52layer at this point of time this
47:54subscription tracker which is this
47:55employee is very good with my emails and
47:59it gets better every single day as I
48:00talk to it but only for tracking
48:02subscriptions.
48:04>> So over a course of time, it'll it
48:06should be able to also learn what
48:07subscriptions I usually cancel, what
48:09subscriptions I don't cancel, what are
48:11the ones that I'm using every day and
48:14all of that, right? Same goes with
48:15everything. And the other big advantage
48:17of beyond having a computer is that it
48:20can talk to each other. That is I can
48:23tell my
48:24to say that hey can you talk to
48:26subscription tracker and tell me have
48:28you paid for this tool because this tool
48:30is not working. Someone the team has
48:32messaged me, right? So it can go talk to
48:34subscription tracker on my behalf, get
48:36that information and give it back to
48:38this.
48:39>> Nice.
48:39>> So it is agent.
48:41>> How do you build? Let's say you built a
48:42subscription tracker. You just go open a
48:45chat ad.
48:45>> Let's try let's try building a simple
48:47one.
48:47>> Now let's say topic like all news in the
48:50world. I want to know what are the top
48:51newses in the world. Exactly.
48:52>> All news in the world. Okay.
48:53>> Like exactly.
48:54>> I'll click on new chat.
48:55>> Like I want to be really smart person in
48:57the world of geopolitics. I want to know
48:59every news in the world related to
49:04first we'll create a new bot.
49:07>> Hey, uh I want you to be my research
49:10agent. I want you to be able to use my
49:13Twitter. Scroll through my Twitter every
49:152 hours. I'm very keen on US
49:19geopolitical news. So look up for all
49:21the keywords on US geopolitical news
49:25that's happening. Look for tags inside
49:26of Twitter and also scroll my feed if
49:29you're able to find anything. Triage all
49:31this information and give me in a
49:33condensed format right here. And I want
49:35you to keep updating this every 3 hours.
49:38Set it as a schedule as well.
49:43Now I can call it whatever. I didn't
49:44really name it. I can call it the
49:46researchy or whatever you want to go
49:48with. Now what it will do is because
49:51Twitter
49:53advantage
49:55>> so it has access to best access to
49:56Twitter.
49:57>> So it will now be able to access
49:59Twitter. It'll ask me for my credentials
50:01and all that to log in.
50:02>> I have to log into Grogbots's computer.
50:06>> But I think it already has access to it
50:07because it's connected to my Twitter
50:08account.
50:09>> Okay,
50:09>> because I use Grock, right? I've already
50:11set this up. So it's already doing it
50:13like on it. I'll set up a US
50:15geopolitical digest from X and drop the
50:17first one here once I've scanned
50:20>> because this is already there.
50:22>> This is already connected.
50:24>> So now it's just the agent is set.
50:26>> This is this became too simple. The
50:28agent is set
50:28>> like that's it.
50:29>> That's it. It's done.
50:31>> Like Max either it would have asked you
50:33for
50:33>> by the way it renamed itself to
50:35Geioatch. [laughter]
50:37>> Max it would have asked you to just
50:39login and connect as login.
50:42>> Wow. Now if you if I want the same thing
50:45to happen on I don't know YouTube
50:48YouTube
50:48>> this is like inshots of the world are
50:50dead after this I can have my own
50:53>> pretty much yeah you can have
50:54>> you don't need news in shorts you don't
50:55need all
50:56>> triage but they're not dead because not
50:58everybody will build it
50:59>> ah
51:01but at some point the way it is growing
51:03it's done then because everybody will
51:05want to have their own customized feed
51:07like I want to be doesn't mean that
51:10everybody wants to have Raj that's the
51:11whole point But I want to be updated
51:13about the news of the world. Everybody
51:14wants to be updated of the world. No,
51:17>> very few people. You don't want to
51:18actually I know this.
51:19>> Yeah. [laughter]
51:20>> But I'm still see not that I'm not
51:21interested in geopolitics.
51:24>> It is I'm I want to know what's
51:26happening but it's not that not of your
51:29concern like not
51:30>> Yeah. It it doesn't drive me. It doesn't
51:32do anything around me
51:33>> like AI news. I don't
51:35>> you will not make it but I have tried
51:36just for that because that matters to
51:37me.
51:38>> Like I don't care what new product is
51:39launched on a product.
51:40>> Exactly. Yeah,
51:41>> exactly. All right. So, there you go.
51:43Schedule is on for 3 hours on weekends
51:45weekday and it'll do the digest. It's
51:47still watching right now. It's doing all
51:48the work. Once it is done, it'll dump
51:50it.
51:50>> Nice.
51:51>> For example, let's say I want my
51:54so all the things that that we decided
51:56for topic selection like all five six
51:59things I can make one one each for all
52:02or I can just give one person only to do
52:04all of these things.
52:05>> I would break it down into one one for
52:08all.
52:09>> Why? The more channel spec, the sharper
52:12the task, the better it is.
52:14>> Okay?
52:14>> Right? When you have unlimited
52:16employees, why do you want to worry
52:18about giving one employee every
52:20>> $200 per employee? No.
52:21>> No. Unlimited employees for $200.
52:24>> Okay?
52:25>> Okay. [laughter]
52:27>> Because it's too expensive for an
52:29>> then then India will set up new
52:31companies saying don't use AI, hire on
52:33people. [laughter]
52:34New service economy will pop in. That's
52:37not how it is. like it's it's for all
52:39the employees. In fact, that is how I do
52:42it.
52:43>> I assign an agent for as less work as
52:47possible and I spin up agent swarms.
52:51>> I don't do one agent. I don't do two
52:52agents. I spin up agent swarms
52:55>> like 50 agents at Google.
52:56>> 100 agents, 200 agents, it doesn't
52:59matter the number of
52:59>> smallest like for the smallest sharpest
53:02task, one agent.
53:03>> Yes,
53:04>> that's a better strategy. That would be
53:05better.
53:05>> That's always a better strategy. M as
53:07long as the context layer is same.
53:13So there is a boss the boss
53:20you're lost.
53:22>> So here we did Twitter for example. But
53:26can it do u
53:29can it do Instagram? Can I do YouTube?
53:31>> Yeah. So for Instagram or something like
53:33that what I have to do is
53:34>> No, but Grock only can do it.
53:35>> Yeah. Anything. Okay. But it cannot go
53:37like this. See, this is where your brain
53:40comes into the picture. Last time also I
53:41told you this.
53:44>> It can do it. But you should recommend
53:47how you want it to do it. For example,
53:49Grock, it was able to do out of the box
53:51because it was Twitter's.
53:53>> Yeah.
53:53>> Instagram. So what do you have to do?
53:56You have to give it a source of data for
53:58that. Either you can ask for example in
53:59the simplest way say I want uh similar
54:03data from Instagram res also how do we
54:08go about it in most cases it should be
54:10able to recommend they're smart enough
54:12to tell you hey these are three four
54:13sources should I integrate with it at
54:15max it will say
54:21same idea look at this checking if
54:22Instagram already is set up
54:26>> right now we're using grog M
54:28>> so it has access to Twitter for free
54:31>> Instagram API
54:32>> okay
54:33>> so you
54:35for example you end up using appy
54:37>> but you don't know appy because you
54:39didn't the person didn't see our podcast
54:40or whatever you have not researched how
54:42do you go about it
54:43>> ask
54:44>> look at this once you sign in my
54:45computer session stays I'll fold real
54:47into the same three-hour dig so what it
54:49will do is it's asking me sign in into
54:51your Instagram account I will scroll
54:53your Instagram accounts but that's not
54:54what I want
54:56>> I want data from hundred hundreds of
54:58Instagram accounts. So, I have to use a
54:59scraper. So, I can use like a tool like
55:03Appify,
55:05>> right? Or
55:06>> Appify at this point should start paying
55:08you.
55:08>> Yeah, man. [laughter] I know. I know.
55:11>> Every podcast you you give me.
55:14>> I'll give you one more name. That's why.
55:15Super data. [laughter]
55:18>> This I came across. It's cheaper than
55:20Appify.
55:21>> Okay.
55:22>> Cheaper than API, but it's not as
55:23comprehensive as a but use cases. This
55:27is pretty good super one of my team
55:29member recommended or super data you and
55:31now
55:31>> this is better or appi is better
55:33>> for social data this has been cheaper
55:35everything is the same there is no
55:36better but again right how do you you'll
55:38be like but webub
55:41in claude and all there was a button I
55:43used to click a button and do it now how
55:45will I do it I don't do anything right
55:46now I just go to documentation copy the
55:48URL and I'll be like no
55:52I want
55:54>> do appy do appy because that's the best
55:56And now everybody knows that's a thread.
55:58>> No, I want you to integrate uh super
56:01data or appi for me. So I'll give you
56:04both the URLs. You go through the data,
56:06you go through the documentation and
56:07tell me which is better. In fact uh uh
56:10build one layer as a primary data layer
56:12and if that fails, use the second
56:14service as a backup layer. Uh and I want
56:16this to be done for free. So try to use
56:19both the accounts in a way the free
56:22limits are not crossed so that I don't
56:23have to pay for it and still the work is
56:25done.
56:28Why not, right? And just in case it
56:31doesn't know what appy and this is, I'll
56:34just usually do this. But today you
56:36don't need to do it. You don't need to
56:37give URLs also. Uh these days I just
56:42give this data man. The agents are smart
56:43enough to figure out now
56:46>> connect they I've done more. No, now
56:48I've given more context.
56:50>> I've told the problem. I've given the
56:52solution also. It only has to execute.
56:55In most cases, I don't know the
56:56solution. I only have a problem.
56:59>> You figure out and tell me.
57:00>> I talk to AI to solve the problem. But
57:02this is this is where your brain comes
57:04into the picture, right? You're like,
57:08I want to build a bloody triad system
57:11where I'm scanning 500 Instagram
57:13accounts every single day.
57:15>> So, it was like, okay, he just wants to
57:17scroll tweets. No, scroll scroll res.
57:19No, I'll log into his account. But if I
57:21do that same login of traging or or
57:24using like 300 Instagram accounts using
57:26my Instagram account, my Instagram
57:28account will get back.
57:29>> True.
57:30>> It didn't have that context. So this is
57:32where your judgment layer comes into the
57:33picture. This is where your experience
57:34comes into the picture. Boss, I'm
57:36building a triage system. I can't just
57:38use my Instagram. I don't want to get
57:40banned.
57:40>> It'll push the limits. So I have to find
57:42a different way.
57:44>> So what is the way? Okay, I'll go with
57:46finding a service which can do this for
57:47me.
57:48>> True. We did that research a couple of
57:50podcasts back. People have not seen can
57:51go and see it, right? And then we use
57:54that data saying ampify and super data
57:56use.
57:57>> Nice.
57:58>> Or in this case, you could have done a
57:59push back also. You like use some other
58:01data scraping service to make that
58:02happen. It would have figured it out
58:04quite frankly.
58:05>> But this is how this is how agents are
58:07done.
58:08>> But effectively, right, look at this.
58:09Appify actually can discover res but
58:12it'll cost you 2.60 per thousand res.
58:15Uh
58:18>> super data cannot search Instagram.
58:21>> Oh
58:24where is it?
58:26Oh [snorts] if the real URL is present
58:28for the free account for the free
58:30account.
58:31>> But anyways it'll figure out you can see
58:33this add app. If I connected I just say
58:35add and it'll go on to do its stuff.
58:37This is how we would go about
58:40>> finding the things. But today we used in
58:43this case we use Grogbot but a lot of
58:46people
58:49in that case I usually recommend
58:51something open source
58:52>> slightly more work to do but we can use
58:54Hermes agent.
58:55>> Okay.
58:56>> Now Hermes agent interesting and that is
59:00before this if you remember when I
59:01showed you Hermes agent I spent 25
59:04minutes just setting it up.
59:05>> Yeah I remember that it was technical
59:08>> very technical and difficult. Well, I
59:10didn't set up. That's why.
59:11>> Now, [laughter]
59:12you will set up
59:13>> because it's become one click. Once you
59:15do this, you'll have a Hermes agent like
59:18this. This is the Hermes agent right
59:20now. Beautiful desktop app. Now, just
59:22like we had that on uh
59:24>> Grogot. Grogbot. No, it also has
59:26something called as bots right here.
59:29>> If you go to here, you can see I have a
59:30caller. I have a inbox.
59:32>> Oh, it's same like having chats the way
59:34you did there. It's not. That's why I
59:36was asking you not to do arms because
59:38last time it
59:39>> look how beautiful this is right now.
59:40>> This is just like WhatsApp chat.
59:42>> For example, look at this. I have an
59:43inbox here. I mean, I built this to show
59:46you. I don't even use this. I use
59:48everything on my Slack like you know.
59:50But uh for example, my Proton mail is
59:53connected. I just run a query saying
59:54that hey like what is the
59:57video? What are the integrations that
59:58have come across to my email? Here are
1:00:00all the integrations that companies have
1:00:02reached out to me to work on with.
1:00:04Right. Uh my 7-day social media report
1:00:06is right here. How is my content
1:00:08performing everything? It is connected.
1:00:09My finances are connected here. Uh I
1:00:12just made you a call. That call was not
1:00:14me. It was my AI who called you some
1:00:15time back. That was also done by this
1:00:18Hermes agent.
1:00:19>> But wait, you did social what did you
1:00:21set up there? Data and analytics.
1:00:23>> Yeah. So all my social data gets
1:00:26harvested onto single platform to see
1:00:28what's working, what's not working. And
1:00:29how do you like you've just given access
1:00:31of
1:00:32>> so the way I given access to this to
1:00:34understand my social data is using this
1:00:37tool called as metricool everybody
1:00:39should pay us no all these guys should
1:00:41pay us what the hell anyways this is
1:00:43metricool if anyone from metricool is
1:00:45looking pay us a lot of money [laughter]
1:00:48>> but metricool is basically a social
1:00:51media scheduler
1:00:52>> okay
1:00:52>> but it's API so I have integrated
1:00:55metricool
1:00:57again if you say integrated
1:01:00technical. Okay. I logged in. I went
1:01:03into I've connected all my Instagram
1:01:05accounts and whatnot.
1:01:06>> Okay. LinkedIn, Instagram, Tik Tok,
1:01:08YouTube, all of it.
1:01:09>> It harvests all the information. For
1:01:11example, let's say Facebook did 20
1:01:13million views in the last 30 days,
1:01:15>> right? And all that you can see and it
1:01:17has all socials, right? Instagram,
1:01:18LinkedIn, every can you connect multiple
1:01:21YouTube accounts?
1:01:21>> Yes, you can.
1:01:23>> Or can you connect to 100 accounts
1:01:26>> per social handle?
1:01:27>> Nice. Is it free or
1:01:28>> No, no, nowhere close.
1:01:30>> It's not. How much is it for?
1:01:32>> Some $100 per month something. It's
1:01:34expensive. It's for uh It's not for
1:01:36regular people. It is for people who run
1:01:38multiple social accounts.
1:01:40>> So, [snorts]
1:01:41every time you come here, you increase
1:01:43my company's
1:01:45>> I increase your efficiency also. No,
1:01:46>> but you increase subscription money for
1:01:48my company. Cost of my company just goes
1:01:50up after every podcast or whatever. It
1:01:51should be otherwise I should make more
1:01:53money after it. You don't make money by
1:01:55spending less money. You make more money
1:01:58by making more money. And you can only
1:02:00make more money when you spend more
1:02:01money and improve efficiency. I heard
1:02:04something like this from this guy called
1:02:05Raj. [laughter]
1:02:07Wow. Wow. Wow.
1:02:10>> Okay. Uh look uh this is Metricool. All
1:02:14right. Uh by the way, just to make it
1:02:17very clear,
1:02:18>> you could have got all of these things
1:02:20done for free also.
1:02:21>> Okay.
1:02:22>> Right. I could have used five different
1:02:23services. For example, YouTube has this
1:02:25API. You can integrate the API. Meta has
1:02:28its API. You can integrate Meta API.
1:02:30Twitter has its API. You can
1:02:32>> Instagram doesn't have an API which
1:02:33gives you
1:02:33>> meta API. Okay.
1:02:35>> There is Meta API. You have to create a
1:02:37app, personal app in the developers.
1:02:40And do it.
1:02:41>> I was just lazy to do all of that,
1:02:43[clears throat]
1:02:44>> right? I was okay to spend this $50
1:02:46whatever per month then figuring out all
1:02:47of that. So, I just chose to do this.
1:02:50when you also do API, you have to build
1:02:52an infra layer to save all the data.
1:02:54>> Okay.
1:02:56>> You don't try to be the expert at
1:02:57everything. No.
1:02:58>> Yeah. Yeah. Fair. Fair.
1:02:59>> That's why I use this. But again, I
1:03:00don't use I don't open metricool at all.
1:03:03>> I this is this is a software for my
1:03:05agent.
1:03:06>> Okay.
1:03:06>> So, I come to
1:03:07>> And you don't use it forululing?
1:03:09>> Nothing. I use it.
1:03:10>> You use it for just analytics.
1:03:12>> Yes. And I also don't watch that data.
1:03:15>> I just go to if you go to this uh
1:03:16settings, right?
1:03:18>> Sorry, not brand settings. If you go to
1:03:20account settings I guess huh account
1:03:22settings there's something called as API
1:03:24I just copy this key
1:03:26>> right and I come to something like
1:03:29Hermes create a new agent let's say uh
1:03:32let's call this the social data whatever
1:03:36okay I can call it whatever I want this
1:03:37is inside of Hermes right now I can do
1:03:39the same inside of Gro also
1:03:40>> yeah yeah
1:03:41>> and I'll be like connect [snorts]
1:03:43to my metricool
1:03:46account using this
1:03:49>> API
1:03:50>> API key
1:03:52and I will not paste it or you can
1:03:53actually blur it. this
1:03:56I just copy this
1:03:58>> paste it in few cases like I said right
1:04:01I usually go and give it the API
1:04:02documentation for example I can give
1:04:04this documentation but these models are
1:04:06smart enough to understand it I just
1:04:07click on send it will work for 20
1:04:09minutes it'll pull all the data
1:04:10automatically once the data is there I
1:04:13can say every seven every uh every day
1:04:16end of day send me a report it will send
1:04:18you the report
1:04:19>> nice
1:04:20>> that's pretty much what it is like today
1:04:24The world is not so complicated like it
1:04:26was before. It looks complicated because
1:04:28I need a data guy.
1:04:30>> The friction layer there is me being
1:04:33okay to say that I can go and copy and
1:04:36paste the API key. Not that the moment I
1:04:38heard API like oh yeah and you run away.
1:04:41>> That's all you had to do. You had to
1:04:42take that one extra step of saying I
1:04:45don't I I am not I'll not be scared.
1:04:48I'll be I'll be figuring this out.
1:04:50Once you do it, you're unlocked forever.
1:04:54I'm sure appi when you were doing for
1:04:55the first time was scary
1:04:58>> scary but once you did it you're like
1:05:00this is it
1:05:01>> yolo
1:05:02>> yes this is it. So that's the whole
1:05:04point right. So this is uh Hermes agent
1:05:07that people can build on top of but the
1:05:09ones that we have done is slightly
1:05:10different. Okay. The ones that we have
1:05:13done is I'll show you maybe towards the
1:05:15end how it looks like as the output.
1:05:17>> Okay.
1:05:17>> Okay. But what where were we? We were at
1:05:21>> you said no before you finished the
1:05:23thread
1:05:24>> where you [snorts] told me
1:05:26that you don't uh
1:05:30>> I I said data guy and you I asked I told
1:05:34you that you don't need a data guy then
1:05:36>> do you
1:05:37>> bro
1:05:39we do have a couple of data people in
1:05:42the company but they exist to verify
1:05:45>> when we building a new data system if
1:05:47it's correct or wrong
1:05:50I don't know if that makes sense to you.
1:05:51>> Yeah, that that does but only the
1:05:54beginning. Or do they just keep random
1:05:56checking as well and as act as an admin?
1:05:58>> I don't think they do anymore. First few
1:06:01verifications happen.
1:06:03>> So data and analytics guy is gone.
1:06:05>> I think their jobs are evolving is how I
1:06:06would put it. Are you being that data
1:06:09person who's leveraging all these AI
1:06:10tools to be able to do a lot more than
1:06:13what you were able to do before? Because
1:06:15see what's the data guys job to look at
1:06:18all the data put it in a report make
1:06:21some sense out of it put pick up the key
1:06:24highlights of the key experiments and
1:06:25the key things which have worked and
1:06:26which have not worked and present that
1:06:29to a person who will actually implement
1:06:30these things and turn around all of this
1:06:33now your agent is giving you
1:06:35>> who's going to ask the questions
1:06:37>> you only know you don't know always
1:06:40that's that's actually the core job of
1:06:42the data guy to ask the right questions
1:06:43to the data
1:06:45As in
1:06:46>> if as in you are you you will say up
1:06:50views come
1:06:53the data guys job is to understand views
1:06:56come break it down into 10 15 questions
1:06:59and then look for the data
1:07:01you understood the job of a data person
1:07:05before very good data person before used
1:07:08to be take your highle problem which is
1:07:11why are our views down or
1:07:13>> yeah okay Example, why are our views
1:07:15down?
1:07:16>> This is working well and this is not.
1:07:18>> Okay. Why is this real working well? Why
1:07:20is this not?
1:07:21>> That's your question. Your data guys job
1:07:24is Raj said this re is working. This re
1:07:27has not worked. What could what are the
1:07:29questions that I can ask. Was the topic
1:07:31right? Have we created topics like this
1:07:34before? Was the uh script written in the
1:07:37right way? Did we post it in the right
1:07:38way? Was the pattern right? Did we use
1:07:41any words? how was the retention graph
1:07:43of this data versus other data. Once he
1:07:46has 10 15 questions or she they used to
1:07:49dig into data to find evidences for each
1:07:51of these hypothesis.
1:07:54>> That was the job of a data layer person.
1:07:56But the problem was we the average data
1:08:00person used to say
1:08:04that's a terrible data person.
1:08:07>> You're not driving decisions. You're
1:08:08doing what was considered as smart work
1:08:11before because you had to write SQL and
1:08:13all that was not easy either. You and I
1:08:14could not have done it. So we were still
1:08:16appreciating and respecting and whatnot.
1:08:18Today that is gone.
1:08:19>> Today what is left is are you able to
1:08:22translate my problem into 10 valuable
1:08:26questions that I'm able to ask, get the
1:08:30data and take conclusions of it. Because
1:08:33this middle layer of doing data
1:08:35research, figuring out cleaning of data,
1:08:37building data sets around it, building
1:08:39hypothesis, looking back into the
1:08:41databases, writing SQL queries.
1:08:42Sometimes SQL doesn't work so you have
1:08:44to write Python queries, whatever that
1:08:46is is all being done by AI. So as of
1:08:49today, I again this is not the actual
1:08:52slack of mine like actual computer of
1:08:54mine like right.
1:08:56>> I only have one project in this which
1:08:58I've been using which is GS data.
1:09:00>> What is GS data? GS data is a it's
1:09:03inside of codeex right now you can see
1:09:05this right
1:09:05>> h
1:09:07>> GS data is basically a project that we
1:09:09have created
1:09:10>> which has access to my meta ads like
1:09:14real [snorts] time my database
1:09:17>> my database when I say users revenue
1:09:22zoom data
1:09:23>> everything is connected whatever my uh
1:09:26whatever my uh
1:09:27>> wherever you need data and
1:09:29>> huh whatever data I have it's connected
1:09:30so
1:09:40okay,
1:09:40>> so all the finance data is here. So all
1:09:43I have to ask is a question right now
1:09:46saying something like
1:09:48hey uh you know uh I've been seeing a
1:09:52steep decline of ROI from last month to
1:09:55this month uh based on the ad spend to
1:09:57the revenue that we have had. Could you
1:09:59go through and understand what could be
1:10:01the three to four key metrics uh that
1:10:03could be the key indicators for me to
1:10:05understand what are some metrics that I
1:10:07need to optimize for to pull up the
1:10:09revenue back again.
1:10:12>> What is this codeex for?
1:10:14>> It's a agent harness
1:10:17memory tools
1:10:20skills everything is in there. Right.
1:10:22What are the tools here? The tools here
1:10:25is the data is the data that it has
1:10:26access to.
1:10:27>> Look at this. just loaded tools a bunch
1:10:29of tools. So first it read the business
1:10:32intelligence skill
1:10:33>> which you have built. So which I have
1:10:35built which is what Joe it needs to
1:10:38understand what my business is how it
1:10:39works what is what so that is a business
1:10:42intelligence skill then it read the next
1:10:44skill revenue attribution sematic layer
1:10:46skill
1:10:47>> where I'm basically we have taught AI
1:10:49how [snorts]
1:10:50to capture revenue from the source
1:10:53concept
1:10:58okay and this is available to everyone
1:10:59in the company and it'll go inside of it
1:11:02dig into data it has Aurora DB. You can
1:11:04see it has started to write Python right
1:11:06now. It has the metad pulling in the
1:11:08data.
1:11:09>> It'll pull every I didn't don't even
1:11:10need to know what it's doing. Okay.
1:11:12It'll find
1:11:13>> this is impressive. This is crazy. If it
1:11:15gives you the real data,
1:11:17>> we are running performance marketing for
1:11:18a product that we have, right? And I
1:11:21gave a case study here. Key if my
1:11:25product would have been would have been
1:11:27X price
1:11:30>> versus Y price. And this is the
1:11:32conversion that I saw. This is the
1:11:34conversion that I saw. It's beach. I
1:11:36added a new add-on which is much
1:11:38cheaper.
1:11:39>> So [snorts] build the whole simulation
1:11:40for me and tell me which is driving to a
1:11:43better revenue in all these simulations.
1:11:46That is what it is doing right now.
1:11:48>> Crazy.
1:11:49>> And this led me to understand that if I
1:11:51do this executed properly, it can give
1:11:53me a 3.5x more revenue than I'm getting
1:11:57on the same spend today.
1:11:59But you're just killing the guess.
1:12:03>> Like anything and everything which was
1:12:05an intuition and guesswork, you're
1:12:06killing it. You're like, I want
1:12:08concrete.
1:12:10>> That's wrong. I guess more
1:12:13>> but I only make those guesses based on
1:12:15data. [snorts] Everything is a guess.
1:12:18>> You experiment more.
1:12:19>> Yes, you experiment more experiments.
1:12:21>> But you're trying to minimize everything
1:12:23which was a guess work.
1:12:24>> My experiments are 10 times better.
1:12:27>> Every experiment. So the one that I was
1:12:30showing you of that simulation that we
1:12:32did and then I executed it yesterday
1:12:34night and you saw the lift also evidence
1:12:35of that. All of this happened because we
1:12:39thought
1:12:44we used to come up with ideas that we
1:12:45want to do this what will happen. We
1:12:47never used to implement it because we
1:12:49were too scared something will break and
1:12:51at the scale that we operate if anything
1:12:53breaks it's very bad. It could just
1:12:55destroy the whole quarter for us and we
1:12:57could go to losses.
1:12:59>> Right? Right now we are able to simulate
1:13:03the risk profile for me.
1:13:05>> What is the worst case? What is the best
1:13:07case? If you see some conversations, we
1:13:09we run ultra mode conversation. So on on
1:13:13this right, if I go to new chat and
1:13:16there is something called as here,
1:13:20this is called as ultra. When you turn
1:13:23on ultra mode and ask data,
1:13:26it will spin up multiple agents to
1:13:29crossverify every single bit. So that
1:13:32hallucination chance. So these ultra
1:13:36mode tasks run for 4 five hours before
1:13:38it gives you a decision
1:13:40exact. But if it's given a decision, it
1:13:43will tell you exactly. I'll tell you
1:13:45something that will blow your mind.
1:13:47Okay.
1:13:49We basically I don't know if you spoke
1:13:51about this but we have started to do
1:13:54implementation of AI for brands
1:13:57>> okay
1:13:58>> that [snorts] is for example there are
1:13:59companies out there who want to get this
1:14:01done for their company because you saw
1:14:03how valuable this is
1:14:04>> and they don't know how to do it because
1:14:05it's obvious right AI implementation is
1:14:07a big play we have spoken about it you
1:14:09have recommended me to start it
1:14:10>> so started in a very small way
1:14:13>> and we were looking at that data of what
1:14:15is working what is not working what is
1:14:17it and then we realized We're wasting a
1:14:20lot of time [snorts] talking to a few
1:14:22people who are very very problem aware.
1:14:25>> They also want a solution but they have
1:14:28a budget but that's not a budget that we
1:14:30can work on because we can only pick
1:14:31five or 10.
1:14:33>> So what are the experiments that we can
1:14:35run to make sure that more people are
1:14:39aware of this and right audience come to
1:14:41us and we ran a lot of data because our
1:14:43funnel works in a way where if people
1:14:45are interested we do a call with them.
1:14:46>> Yeah. Inside the call, we understand
1:14:48what the problem statement is and then
1:14:50we also teach them how to do it. If
1:14:52they're not able to do it, we will help
1:14:53them to do it if they're willing to pay.
1:14:56>> We we had a lot of this data of Zoom
1:14:59recordings of the sessions, transcripts,
1:15:01when they attended, when they dropped
1:15:03off, how they reacted, when Zoom also
1:15:06has reactions. You can drop a heart,
1:15:08>> you can drop a thank you, you can drop
1:15:10messages. We took all that data of
1:15:13multiple rounds that we had done and I
1:15:15gave I mean all the data is already
1:15:17available. I just asked my AI to look at
1:15:20all that Reddus database which has all
1:15:22this data of mine and tell me what are
1:15:26some things that I had told or my sales
1:15:30team had told or my advisers have told
1:15:33which led to a positive reaction signal
1:15:36either via message or reaction which led
1:15:39to more people
1:15:42coming to us on a higher ticket price
1:15:44rather than a smaller ticket price.
1:15:47It was able to give me three experiments
1:15:49to run.
1:15:52Okay, saying change. It knows what we
1:15:54are talking. It knows how the
1:15:55conversations are going.
1:15:57>> It knows what it was able to understand
1:15:59what these user inhibitions are
1:16:02>> and how we are not able to solve their
1:16:04problems because we're getting too
1:16:05technical in a few cases. It asked us to
1:16:07make these three changes.
1:16:11>> You won't believe conversions went up by
1:16:1344%.
1:16:1744 45%.
1:16:19Eventually ROI went up by 44 45 not
1:16:22conversions 45% ROI went up.
1:16:25>> Revenue from the same spend went up by
1:16:2745%.
1:16:29>> That's insane.
1:16:30>> But again the reason why we don't talk
1:16:32about all of this is these are we are
1:16:34able to get to this level of data purely
1:16:36because we are able to ask the right
1:16:37questions.
1:16:38>> People need to learn how to ask right
1:16:40questions. Give it right context. If you
1:16:42don't give right context, it give you
1:16:43right answer wrong answers. So this is
1:16:45where when something big moves like this
1:16:47are being made, we got hypothesis from
1:16:50AI this could work.
1:16:52>> I don't want to rely on AI here because
1:16:54if AI was wrong anywhere I'll get
1:16:56screwed. So I have a human layer
1:16:58verifying everything.
1:17:00>> Got it?
1:17:01>> High qualations but once it was verified
1:17:03once I know now I don't have to ask 10
1:17:05times verify verify verify verify it
1:17:07becomes easier for me.
1:17:08>> Got it. Okay. So we were we just
1:17:11selected the topic. [laughter]
1:17:16>> Okay. For topic selection
1:17:19we had data sources.
1:17:21>> Mhm.
1:17:23>> Right. We built this data sources. So we
1:17:26got let's say 100 topics here.
1:17:28>> Okay.
1:17:28>> But every day I can make one topic.
1:17:30>> What will I do with 100 topics? Now
1:17:32>> now this is like picking needle from the
1:17:34haststack.
1:17:35>> Okay.
1:17:36>> I got the whole hay stack. I have to
1:17:38pick the needle. Yeah.
1:17:39>> How will I pick the one topic that I'll
1:17:42create today?
1:17:43>> How?
1:17:44>> That is where again data comes into the
1:17:45picture.
1:17:46>> What do I do?
1:17:48>> Two things. One is that out of these 100
1:17:51topics that I have,
1:17:54have I created any similar piece of
1:17:56content that has worked for me in the
1:17:59past?
1:18:00>> Okay.
1:18:01>> Two, has anyone else created any similar
1:18:05piece of content that has worked for
1:18:06them? M
1:18:07>> two evidential layers I want.
1:18:09>> M
1:18:10>> and everything will have a weighted
1:18:12score.
1:18:14>> If you have taken ideas from an
1:18:15Instagram reel that has gone viral
1:18:18>> that comes with a score by default. If
1:18:20you have taken just news that also comes
1:18:23with a score. All these 100 ideas are
1:18:25basically sent to our data bank.
1:18:30What this does is this is a DNA
1:18:33playbook.
1:18:34>> What does this DNA playbook have? It's a
1:18:37big document which captures pretty much
1:18:41what works for me as content.
1:18:44>> Okay,
1:18:45>> what has worked in the past. It has
1:18:47data. It has uh the topic. It has the
1:18:50transcript. If it is a written post, it
1:18:52has the post. It has likes, comments,
1:18:54views, shares. All the metric that I
1:18:56possibly have across social medias is
1:18:58all in a Google sheet. Simplest way
1:19:01Google sheet. And also we have a DNA
1:19:04road map. These are the angles that have
1:19:06worked. We also have human layer data
1:19:08where my team for all the pieces that
1:19:11have worked have written by themselves
1:19:16because of this the human touch I have
1:19:19tagging of all this data. Okay.
1:19:21>> Now I take all these 100 pieces of data
1:19:24>> spin up 100 AI agents parallelly.
1:19:29Each agent is attached with one idea and
1:19:31they go into the data bank and the DNA
1:19:33playbook to rate each one of the topic
1:19:35from a scale of 1 to 100 on the
1:19:37potential of it to go viral based on my
1:19:40past data based on secondary data and
1:19:42everything around it. But then this is
1:19:45aren't you limiting yourself [snorts]
1:19:49to just keep creating content on the
1:19:51basis of what has worked because after
1:19:54some time after like as a creator right
1:19:57or as a marketer as someone who's trying
1:19:59to create a viral content
1:20:02>> you need something fresh which probably
1:20:05is not reflected in your past data bank.
1:20:08>> So how do you do that? Because this will
1:20:10only rank topics virality based on what
1:20:14has worked for you and the underlying
1:20:16structure behind it. But maybe let's say
1:20:18a new thing works for you.
1:20:19>> No, could work for me.
1:20:20>> Then this you will not get exactly the
1:20:22same topic, right? Any which ways what
1:20:24you're looking for is signals.
1:20:28Something like this is audience is
1:20:30interested. Something like this audience
1:20:31is not interested. Also one other thing
1:20:33which is a fair question for you to ask.
1:20:36If you think every single topic happens
1:20:38through this, that's not the case.
1:20:39There's always a 25 30% experimentation
1:20:41layer which we have never done.
1:20:45>> You'll go into a silo.
1:20:46>> Exactly. And then at after some point
1:20:48it'll stop working.
1:20:48>> Yeah.
1:20:50>> So you know Boris Churnney who is the
1:20:52creator of cloud code
1:20:54>> basically says every every time there's
1:20:56a new model that comes in you should
1:20:58delete all your skills, delete all your
1:21:00skills uh delete all your markdown
1:21:02files. Delete all your memory and let it
1:21:04play again. I take that concept very
1:21:07very strongly. saying I'll tell you what
1:21:09happens in this process also it works as
1:21:12long as the audience is accepting after
1:21:14a point of time it stops working so your
1:21:17other 20% layer place the job now where
1:21:20you build a new playbook
1:21:23so you're parallely building that
1:21:24playbook you always at any point of time
1:21:27if you want to grow your growth is
1:21:30directly proportional to the number of
1:21:31experiments you're running this is your
1:21:33bread and butter
1:21:34>> so you run your bread and butter like
1:21:36this
1:21:36>> got it
1:21:37>> this cannot go wrong. This is the most
1:21:39scientific way of going about it. But on
1:21:41top of bread and butter, you should also
1:21:42go play cricket.
1:21:43>> It's an 80/20 rule. Yeah.
1:21:44>> 7030 in our case. Sometimes 50/50 also.
1:21:47>> Got it.
1:21:48>> But this allows the process to run
1:21:51without me breaking my head on.
1:21:56Fair
1:21:57>> because a lot of times when we talking
1:21:58we'll get an idea and we'll implement
1:22:00it.
1:22:00>> Fair.
1:22:01>> And we don't do everything for just
1:22:02views.
1:22:03>> All right. But this is a framework of
1:22:05thinking.
1:22:05>> Okay. So spins up 100 agents looks up
1:22:08data and gives me the top 10 topics
1:22:13>> out of 100. So from 100 we come to 10
1:22:15topics. Now from 10 I come to one and
1:22:18the way I do from 10 to one is slightly
1:22:20different. What I do on 10 to 1 is I
1:22:24just have the topic right now today uh
1:22:27GPD 5.6 Six soul dropped
1:22:30>> ultra mode dropped which is very
1:22:32powerful is one of the topic topic
1:22:34direction but topic present are angles.
1:22:41So from here these 10 topics are pushed
1:22:44for 10
1:22:46>> angles each
1:22:48>> and now once I have 10 angles which is
1:22:5010 topics into 10 angles
1:22:52>> 100
1:22:53>> it goes back to the bank again to see
1:22:55have these angles worked out and these
1:22:58are all built off skills. So it is
1:23:00always learning.
1:23:01>> So with this tuning process we are
1:23:03eventually able to come up with
1:23:07>> it go. Yeah, sorry. With this tuning
1:23:09process, we are eventually able to come
1:23:11up with like eventual five
1:23:15topics into two angles or three angles.
1:23:20Here is where
1:23:23human comes into the action.
1:23:26>> Judgment is up
1:23:30last call, last decision, last judgment
1:23:32has to be ours.
1:23:33>> Has to be ours. So the judgment comes in
1:23:35here where I'll be like okay you know
1:23:37what topic
1:23:42topic feeling should be ours that's the
1:23:45judgment and on top of that I don't go
1:23:47basis of this when I see these I get
1:23:50directions maybe we should try this
1:23:52maybe we should try that
1:23:54effect but the hard work of
1:23:58datadrivenness
1:24:00ability of making it work everything is
1:24:02already run
1:24:04you open our Instagram okay or YouTube
1:24:08for that matter we are at this point I
1:24:11say we because it's majorly the team
1:24:13that drives everything for me right now
1:24:15we are at least 3x
1:24:19with respect to every single metric when
1:24:22you compare to anyone with AI across the
1:24:24world
1:24:26>> nice
1:24:27>> and I will tell you everybody else 99%
1:24:30of them are shooting content if I could
1:24:32just record this my delta will be 5x
1:24:36>> and it's because of one reason and one
1:24:38reason obsession
1:24:40of building a process which is super
1:24:42datadriven but judgment left to us
1:24:48>> but this is only topic selection.
1:24:50>> Yes, this is topic selection and in a
1:24:53way uh we have come to packaging as
1:24:55well.
1:24:55>> So how do you now package it?
1:24:57>> Packaging also follows a very similar
1:25:00thought process. For packaging we run a
1:25:02very similar agent again what if it's
1:25:05Instagram or let's say in this case
1:25:06let's say take YouTube in YouTube
1:25:08packaging is very complicated you have
1:25:10thumbnails
1:25:12you have titles and the same topic could
1:25:14be positioned in 10 different ways
1:25:18>> and then that goes and checks your data
1:25:20bank
1:25:20>> my data bank
1:25:21>> sees what has worked IQ MCP to pull the
1:25:25data from there to understand what are
1:25:27other packagings that have worked what
1:25:28are the videos that
1:25:30blowing up because I know that there's a
1:25:32direct correlation to CTR of the video.
1:25:35What are the packagings that are working
1:25:36that could fit in here?
1:25:39>> And then we take 10 of those packagings
1:25:43and run ads.
1:25:44>> Nice. Which agent this is this? Which
1:25:48platform is best? Grock, Hermes, Codex.
1:25:51>> All built on uh for us it's all built on
1:25:54Hermes and Codex.
1:25:57>> All built on Hermes. But all can do same
1:25:59you think.
1:25:59>> Yeah. Yeah. Pretty much it's all depends
1:26:01on how you train them. How good is your
1:26:03skill? Models are there man. Models are
1:26:05there. I don't think models is a problem
1:26:07>> and you trust all of them. Doesn't
1:26:08matter.
1:26:10>> There is evidence. No.
1:26:13I have a topic saying 5.6 soul
1:26:22or AI is giving me three topics which is
1:26:24giving me a saying this is right.
1:26:27Oh man, I why didn't I think of it? Oh,
1:26:29that's right. I had done this some time
1:26:31back. And the other beauty of this is it
1:26:34doesn't only has data of what has
1:26:36already gone out. It also has data of
1:26:38what it had come up with, but we didn't
1:26:40select.
1:26:42>> And by the way, all of this, you
1:26:45remember the second brain I was talking
1:26:46about of what content I consume?
1:26:50>> Yeah. Yeah. Yeah.
1:26:51>> It has access to all that. Oh, FYI, the
1:26:53second brain also includes every single
1:26:55podcast that we have done. Nice.
1:26:58So
1:26:59>> do you actually also run all the things
1:27:02that you're not choosing and if someone
1:27:04else chooses what is the result? Do you
1:27:06ask your agent to go check that as well?
1:27:09>> Come again
1:27:10>> like let's say out of 10 topics you
1:27:13decided to make real on two
1:27:15>> the eight are left
1:27:16>> but someone else in the AI world would
1:27:18be creating real on one of those eight.
1:27:20Do you track that as well because to see
1:27:23the judgment where something that you
1:27:25didn't choose and if someone else chose
1:27:27that how did it perform?
1:27:29>> No, we have not tried doing that. That's
1:27:31a good idea but we could track actually.
1:27:36>> So then if agents are only doing it
1:27:38>> I never thought that's a good idea.
1:27:39That's that's a very good idea I feel
1:27:41because this is essentially like
1:27:44reinforcement learning for me. This is
1:27:46making your judgment better.
1:27:48>> AI's judgment better as well. Uh I mean
1:27:50effectively AI's judgment better. That's
1:27:52a good idea. That's a good idea. And my
1:27:53judgment also better. Like it can change
1:27:55the weighted scores as well for me.
1:27:56>> Like I do that for me. That's why I'm
1:27:58>> That's very smart. Yeah, we should do
1:27:59that. Sir, next podcast.
1:28:02[laughter]
1:28:04>> No, but I do that for my podcast. Every
1:28:05podcast that I purposely intentionally
1:28:08choose not to do, every guest, every
1:28:10angle or the angle with a specific
1:28:12guest, I look for the radar in the
1:28:14world. who is the person who's doing it
1:28:15who has touched this topic in even in a
1:28:17clip
1:28:18>> and then I need to learn that so that
1:28:20next time I can improve my judgment
1:28:22thinking okay what is working
1:28:24>> but now if I ask AI agent to do it it
1:28:26can do it 10x better than me
1:28:27>> yeah yeah but that's the translation is
1:28:30what is important no
1:28:31>> you being able to translate your thought
1:28:34process into a process that works 24/7
1:28:36without you
1:28:38>> is what changes the game for you and
1:28:41this is just idea right then you get
1:28:43into script the moment you get into
1:28:44script you break it down your hook your
1:28:46body your
1:28:47>> how do you do script now from this let's
1:28:50say you've decided the topic
1:28:51>> yeah from here script is actually a mix
1:28:54of we basically don't write the scripts
1:28:58end to end a lot of work is done by AI
1:29:01lot of work is still done by human as
1:29:03well today
1:29:05>> but my perspectives come I told you
1:29:07about the standups that I do right
1:29:09>> my team says these are the three topics
1:29:11that we're considering talking about and
1:29:13all the work is done before I get on the
1:29:15call.
1:29:15>> But break me a script for you what your
1:29:17agent knows so that agent writes.
1:29:19>> So agent basically uh once the topic is
1:29:22selected actually agent when it gives
1:29:25right it gives the scripts also by
1:29:27default rough scripts
1:29:29>> but how does it break the script because
1:29:31you must have given some structure to
1:29:32give the script.
1:29:33>> Oh that is there is a skill that is
1:29:35created there's a skill that is created
1:29:37>> which is built based on topics. For
1:29:40example I have five buckets of topics.
1:29:43Mhm.
1:29:44>> Let's say
1:29:45this is let's say tools.
1:29:47>> Mhm.
1:29:48>> This is let's say uh models.
1:29:50>> M
1:29:51>> this is let's say future tech
1:29:53>> where I talk about this is let's say
1:29:54robotics
1:29:56>> and this is let's say business overall
1:29:58India and business. These are the five
1:29:59buckets. For each of the buckets I have
1:30:01data of what has worked in the past for
1:30:04me and for others. This has 70%
1:30:08weightage. This has 30% weightage for
1:30:10everything.
1:30:11>> Nice. for everything,
1:30:14right? And every time a new script kind
1:30:16of blows up, it gets added to the
1:30:18script.
1:30:18>> Okay? And then it automatically breaks
1:30:21down
1:30:22>> and it automatically breaks down the
1:30:23structure of a script.
1:30:24>> Based on the topic, it goes says which
1:30:26bucket is it falling into? If this is
1:30:27the bucket, what has worked for this
1:30:29uses that skill to come up with an
1:30:31output and skill
1:30:34topic maybe we do something called as
1:30:36loop.
1:30:36>> Okay, what is that? It's called loop
1:30:40engineering. What we do here is we get
1:30:42AI to test. So when we give a topic like
1:30:45this before we build the skill right the
1:30:47way we improve the skill of the AI to a
1:30:49very high level is
1:30:51[clears throat and cough] let's say the
1:30:52topic is uh
1:30:55example
1:30:57uh let's say hermace agent
1:31:00>> okay I uh for me to see if AI is able to
1:31:04come up about her agent very very well
1:31:06as a script what I will do is I have an
1:31:09old script I have which I've written
1:31:12already in the past which has worked for
1:31:13me Okay.
1:31:14>> And the topic was Hermes agent.
1:31:16>> Right. I will train AI with all these
1:31:20topics and I say come up with a skill
1:31:22which will replicate the style of a
1:31:24winning script for me.
1:31:25>> Yeah.
1:31:25>> Once you say I've done what I basically
1:31:28do is I will say okay now Hermes agent
1:31:31is your topic that I want you to
1:31:32generate script on.
1:31:34>> You generate the script. It'll generate
1:31:36the script. Then I will be like okay now
1:31:38that you have generated the script
1:31:39compare it to the old script that I have
1:31:42written.
1:31:43Don't read the script that I've written
1:31:45in the past. Compare it to the old
1:31:46script that I've written. And now tell
1:31:49me how much would you rate the script
1:31:50that you come up with from a scale of 1
1:31:52to 10. It will basically give you five,
1:31:54six, something like that. That's where
1:31:56the skill stands by default.
1:31:58>> No way you can hit a bigger number than
1:31:59that. When you ask it to compare because
1:32:01it's very very specific,
1:32:02>> brutal, huh?
1:32:04>> Once it does, then I'll say all right.
1:32:06You know it's a 5.5. Now your job is to
1:32:10selfimprove the skill. So I want you to
1:32:13run a loop right now for me with a goal
1:32:16that this skill has to generate every
1:32:20script which is a 9.5 out of 10 no
1:32:22matter what. And the way I want you to
1:32:25improve is pick up a topic that I've
1:32:28already written a skill already written
1:32:29a script on that has gone viral. Give an
1:32:32AI model and give a skill. Isolate it.
1:32:35Don't give it the data. It should not
1:32:37know what the scripts that I've written.
1:32:39Just give it the skill. Just give it the
1:32:41topic and the research for it and ask it
1:32:43to write the script. Once it writes a
1:32:45script, you compare it with the original
1:32:48script written and give it a rating.
1:32:50>> If the rating is less than 9.5 out of
1:32:5210, do a comparison like an examiner and
1:32:54tell what are the things it can improve.
1:32:56Once the things are improved, take those
1:32:58things that can be improved and edit the
1:33:00skills so that it can get added and then
1:33:02do the next topic again and then look
1:33:04what the score is and continue this
1:33:07process till you get to 9.5. Don't stop
1:33:09till then.
1:33:10It will run for all night.
1:33:12>> Nice
1:33:13>> to recursively self-improve to get to a
1:33:15score of 9.4.
1:33:18I tried 10 initially.
1:33:22It never finished. It would come to 9.6
1:33:259.5. I made the exam even harder. I said
1:33:27you have to get 10 out of 10 five times
1:33:30back to back. Not once.
1:33:359.5.
1:33:369.5 is the middle ground that I found.
1:33:38So it runs there. But something crazy
1:33:40happened, dude. When I was running this
1:33:41loop for the first time, you'll be
1:33:44blown. When I was running this loop for
1:33:46the first time, this I was doing this
1:33:48exam experimentation for my LinkedIn
1:33:50scripts. Okay,
1:33:51>> it was on claude.
1:33:53>> I literally said this. Here's a topic
1:33:56that I've written in the past. Here's a
1:33:57LinkedIn post. I use this skill to
1:33:59generate a LinkedIn post on this topic.
1:34:01Compare both of them. Run till you get
1:34:03to 10. This is the first time I'm
1:34:05running. I thought it will run all
1:34:07night.
1:34:09slept, woke up next day and I saw the
1:34:11processor completed in 30 minutes. I'm
1:34:14like before I left it was a five. How
1:34:16can it go to a 10? So 9.5 so quickly.
1:34:19>> Then I said can you tell me a last five
1:34:21examples of the LinkedIn post that you
1:34:23came up with and what is the original? I
1:34:24want to see both of them because I
1:34:25thought it was not measuring right and
1:34:27all the five LinkedIn posts it came up
1:34:29with was gibberish.
1:34:31It was not even written on the topic. It
1:34:33was random like [snorts] I it has no
1:34:36meaning to it. The topic is how Hermes
1:34:38agent is awesome or how GPD 5.6 six is
1:34:41awesome or whatever that topic is and
1:34:43it's written Loram ipsum cool Travis
1:34:46>> full full
1:34:47>> full gibberish
1:34:48>> I was like what the hell is this they're
1:34:50not even the same and it says oh I
1:34:54apologize I cheated
1:34:57and I was what do you mean by cheated I
1:35:00was like no so what happened is we it
1:35:02ran a few rounds after that I was not
1:35:05able to improve the score then I
1:35:07realized the way the exam was designed
1:35:09was I was the one who was reviewing my
1:35:11scores. I was the one who was giving
1:35:13myself a score and then the examiner was
1:35:16actually just measuring and saying
1:35:17giving me a new topic. So because I was
1:35:19not passing, I just gave myself score
1:35:21full scores everywhere because the
1:35:23examiner never saw the output. So I
1:35:25passed. [laughter]
1:35:28I'm like what the hell? And these are
1:35:30what these are very solid five class
1:35:32models. These are very very smart
1:35:34models. And that is when I realized it
1:35:38was not trying to cheat me.
1:35:41It was just trying to pass the exam
1:35:44>> and it found it to be hard and it figure
1:35:46out a way to win.
1:35:50>> Isn't that insane dude? Like when
1:35:51[clears throat] humans do
1:35:53>> when that happened I was I lost my
1:35:55I was like how can this even happen? Now
1:35:57when I write loops I know how it can so
1:35:59I orchestrated. There's something called
1:36:01as graph engineering for the same thing
1:36:03where you say agent one will do this
1:36:05agent two you're building a system so
1:36:08that it doesn't cannot cheat
1:36:11>> and every agent you're actually giving
1:36:13them a specific task to do
1:36:15>> yes and they're isolated there are
1:36:17multiple verification layers you don't
1:36:19rate yourselves there's always someone
1:36:21else rating and their job is to make
1:36:23sure that you're not winning
1:36:25>> so that there is no bias because there's
1:36:27a lot of agent bias that comes into the
1:36:28picture as well and sometimes times the
1:36:30best thing that you can do for things
1:36:32like this is to get multiple agents from
1:36:35multiple AI models rather than the same
1:36:37AI model.
1:36:39>> That's called as an agent council,
1:36:41right? Or a council where
1:36:52different
1:36:54and then they're doing the task.
1:36:55Everybody wants to win. Everybody wants
1:36:57to do their job with
1:36:58>> because then agents do one of the these
1:37:00three things as well, right? They fight.
1:37:02>> Yeah.
1:37:02>> Agents can fight as well. So then
1:37:03they're always fighting and not coming
1:37:05up with a good answer.
1:37:06>> No, but they can sabotage each other.
1:37:08But it's easy to sabotage if they know
1:37:09you. If they don't know you, how will
1:37:11they sabotage?
1:37:12>> You are just increasing the variables of
1:37:14making it hard for them to cheat.
1:37:16>> And it's easy to fool each other as
1:37:18well, right? Now you see look human
1:37:21generation has gone through decades to
1:37:24understand how to live with each other.
1:37:26As agents are just born they're
1:37:28extremely good when they're working all
1:37:30by themselves. The moment
1:37:32>> you give them a team
1:37:34>> and that two of the same it becomes a
1:37:37problem. There is you know you will you
1:37:40will see lot of fight lot of sabotages
1:37:43happening. There are a lot of studies
1:37:45that were published on this by anthropic
1:37:47as well, right? Where agents literally
1:37:49where they were given a task of
1:37:51refactoring a codebase
1:37:54code
1:37:58just to make sure that their language is
1:38:00picked.
1:38:02They fought with each other. Eventually,
1:38:04one agent managed to block the other two
1:38:06from even doing the work.
1:38:10This is multi-agent orchestration
1:38:12issues. They do cheat. This is beyond me
1:38:15right now. [laughter] This is just like
1:38:17like all of these things. I can't even
1:38:19imagine what agents are doing and what's
1:38:21going on. I was just every time you come
1:38:24up here and then you tell me 50 things
1:38:26which is like wild.
1:38:28>> Yeah. I'm also learning. No, like I
1:38:30think uh the world is just evolving way
1:38:32too fast, right? And we are all trying
1:38:35to keep up and trying to build systems
1:38:36that will help us to do some fun.
1:38:39>> This what you've built is insane. Okay.
1:38:41So then the script.
1:38:42>> Yeah. Yeah. And it happens automatically
1:38:44and then the posting
1:38:45>> script also in the script there's a lot
1:38:47of context that comes into the picture.
1:38:49Where is my personal context coming in
1:38:51in the meetings that is also automated
1:38:53where meeting transcripts like I told
1:38:55you before.
1:38:55>> So everything starts from the second
1:38:56brain.
1:38:57>> Yeah. It all gets pulled from second
1:38:59brain and goes to the second.
1:39:00>> So there's a second brain. Then there's
1:39:01a your 100 million views content system.
1:39:04>> Yes.
1:39:04>> Which is broken into topic selection,
1:39:07packaging, scripting and posting. And
1:39:10then it's just like on an autopilot. It
1:39:12just keeps going on and on and on. It's
1:39:14improving your business. It's improving
1:39:16your content. It's improving everything.
1:39:17And everything is done by agents and not
1:39:19you.
1:39:19>> Yes.
1:39:20>> And you are just the master who's asking
1:39:22questions and deep questions so that
1:39:23they can come up with
1:39:24>> they're giving a lot of context and just
1:39:26trying to get great answers from an AI.
1:39:31>> This is how do I do this?
1:39:34>> By practicing and implementing small
1:39:36small things.
1:39:37>> You said that you implement it for
1:39:38brands. Yes.
1:39:40>> So why why aren't you not doing for me?
1:39:42>> You're not big enough.
1:39:43>> Oh my god. [laughter]
1:39:45What's what's the price?
1:39:47>> Correct.
1:39:49Mazak.
1:39:50>> But how big companies are doing? What
1:39:51ticket size you're doing?
1:39:53>> The starting is 100K
1:39:55for implementation and we're picking
1:39:57very very specific use cases right now
1:40:00>> which is not bad. 100k is not
1:40:01>> yeah 7day project 100k is a starting
1:40:03point. Uh
1:40:05>> but it's not a lot. A lot of people will
1:40:06pay. I know I know I don't have the
1:40:08capacity to take.
1:40:09>> Ah that's your capacity 100k is not a
1:40:12problem because
1:40:13>> I know back of my head
1:40:15>> that there are at least 25 people who
1:40:17will do like on fingers like I can text
1:40:19them today and they'll do 100k with you
1:40:22>> but we will get there. We are also
1:40:23trying to build playbooks or else it
1:40:25will get very expensive for us to run
1:40:26this because if you can you can just
1:40:28imagine people who are able to do this
1:40:30are also people who are very expensive.
1:40:32H now people who know this game they
1:40:35will come and they will be like
1:40:39no no no
1:40:43you don't have the time
1:40:45>> plus is it because also their client
1:40:47attracting capacity will also be low
1:40:48>> compared to individuals who are coming
1:40:50and do it yeah that is also there we do
1:40:52a lot of training no so that
1:40:53>> like you can command 100k in the market
1:40:55very easily someone else who can even do
1:40:57it can't command 10k
1:40:58>> yes
1:40:59>> as long as they have like a big brand to
1:41:00do it
1:41:00>> yes but the problem also is that one
1:41:02person cannot do it. Some implementation
1:41:05is only one part of it but being able to
1:41:08think about how do I solve this event
1:41:10pro solve this problem event. You're
1:41:11like a consultant, service provider and
1:41:14an outcome driven like provider figuring
1:41:17out. Okay. We're trying to figure out
1:41:19what is the right model here.
1:41:21>> We want to do it. We want to help.
1:41:23>> It's in the autopilot mode. You're doing
1:41:24that game.
1:41:26>> What do you it cannot be an autopilot.
1:41:28This cannot be an autopilot. Uh we have
1:41:30to understand it's a very serious uh
1:41:32>> No, the service that you are providing
1:41:33isn't it will be an autopilot for a lot
1:41:35of companies.
1:41:36>> Correct. Correct. So
1:41:37>> you're giving outcomedriven service. We
1:41:39are building marketplace.
1:41:41>> A marketplace where you can find great
1:41:43talent who can come and implement it for
1:41:45you with our playbooks that we have
1:41:48built so that it can go right not go
1:41:50wrong.
1:41:52>> But uh next time we'll talk about it.
1:41:54>> Interesting.
1:41:55>> Right now it's an early testing phase.
1:41:58>> Nice. [clears throat] So we and then
1:41:59early testing phase we'll talk about all
1:42:01the case studies. Oh
1:42:02>> we can do that. Happy to do it. Yeah.
1:42:04>> And all the things that Oh my god.
1:42:10>> [laughter]
1:42:14>> Are you sure? Are you sure I'm not
1:42:15giving anything? [laughter] You're
1:42:17getting a lot of value, man.
1:42:18>> That I agree. That I agree. Four hours,
1:42:215 hours.
1:42:26[laughter]
1:42:27But thank you so much.
1:42:30Okay. Thank you for watching this
1:42:32episode till the end. We would love to
1:42:34know what you liked or disliked about
1:42:36this episode and which guests you would
1:42:38like to see on the show. Let us know in
1:42:40the comments. Your feedback help us
1:42:42improve and make every episode a little
1:42:45better. I'll see you next time. Until
1:42:47then, keep figuring out
1:42:50[music]
1:42:58[music]
1:43:05>> [music]

About this transcript

This page contains the full transcript of How AI Makes Him Crores: Second Brain, Automations & Systems | Vaibhav Sisinty | FO557 Raj Shamani by Raj Shamani, generated from the public captions YouTube serves with the video. The transcript has 18,856 words across 2,920 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.