WEBVTT

NOTE Sentence-level transcript of https://www.youtube.com/watch?v=Yphdry8ttAQ

NOTE One cue per sentence. Cue ids are the line anchors on /transcripts/Yphdry8ttAQ.html. A cue ends where the next begins, or 2 s after its last word.

s1
00:00:01.309 --> 00:00:03.309
[music]

s2
00:00:12.600 --> 00:00:14.800
Hello everyone.

s3
00:00:15.480 --> 00:00:20.560
So, my name is Ayush Bhardwaj and I did applied AI for a hedge fund.

s4
00:00:20.560 --> 00:00:26.520
And now I do everything tech plus applied AI for a pharma tech startup cuz you know the way startups are.

s5
00:00:26.520 --> 00:00:28.880
You have to do everything, wear multiple hats.

s6
00:00:28.880 --> 00:00:32.480
So, before I start the session, I would like to do a small survey.

s7
00:00:32.480 --> 00:00:36.400
Can I get a raise of hands for all the engineers in the room?

s8
00:00:36.400 --> 00:00:38.680
Okay, that's a tough room.

s9
00:00:38.680 --> 00:00:41.880
Now, can I get a raise of hands for managers?

s10
00:00:41.880 --> 00:00:45.080
Okay, just to be clear, managing AI agent does not count.

s11
00:00:45.080 --> 00:00:46.280
You have to manage people.

s12
00:00:46.280 --> 00:00:48.600
Okay, we have few managers as well.

s13
00:00:48.600 --> 00:00:49.000
Interesting.

s14
00:00:49.000 --> 00:00:51.640
So, they will help me like fine-tune my talk a bit.

s15
00:00:51.640 --> 00:00:58.040
So, today my aim is to take you through the journey of how do you actually build and iterate in applied vertical AI?

s16
00:00:58.040 --> 00:01:02.240
And my experience is from the hedge fund and the pharma tech company.

s17
00:01:02.240 --> 00:01:11.240
So, before delving deep into the recipe, I'll just like take you through what do I even mean by applied vertical AI cuz I don't know if it sounds like a very weird term.

s18
00:01:11.240 --> 00:01:13.640
It's like the vertical word is kind of forced.

s19
00:01:13.640 --> 00:01:14.800
I won't lie, it is.

s20
00:01:14.800 --> 00:01:16.440
I coined this term probably.

s21
00:01:16.440 --> 00:01:24.240
So, applied AI is like built for So, what applied vertical AI is essentially applied AI but built for one very specific industry.

s22
00:01:24.240 --> 00:01:29.760
It's It's aim is to simulate a job of a person in that particular industry in a sense.

s23
00:01:29.760 --> 00:01:35.160
So, an example of applied AI is Google Translate which is like general purpose, helps you translate.

s24
00:01:35.160 --> 00:01:44.800
It could be used in education tech and it can have like tons and various sorts of uses is whereas Elos, which is my employer, the pharma tech company, we specifically build drugs with AI.

s25
00:01:44.800 --> 00:01:46.960
So, that's a very specific use case.

s26
00:01:46.960 --> 00:01:51.920
Another examples of applied vertical AI field could be the legal tech firms that are now coming up with.

s27
00:01:51.920 --> 00:01:53.640
You must I'm sure you must have heard about them.

s28
00:01:53.640 --> 00:01:58.000
So, those are like another the examples of applied vertical AI.

s29
00:01:58.000 --> 00:02:01.920
So, when I left the hedge fund, right?

s30
00:02:01.920 --> 00:02:07.600
So, I was expecting that the world would change for me cuz you know, hedge funds are like really fast and really pressure sensitive.

s31
00:02:07.600 --> 00:02:12.360
Whereas, pharma is like, "Okay, we're going to take 15 years, but we're going to do it right."

s32
00:02:12.360 --> 00:02:16.360
Hedge fund was all about like, "You need to do it fast and mostly right.

s33
00:02:16.360 --> 00:02:20.160
It does not matter if we lose at one paradigm as long as we are overall winning."

s34
00:02:20.160 --> 00:02:23.720
Whereas, a pharma firm is like, "We have to be absolutely right.

s35
00:02:23.720 --> 00:02:25.240
You can take a week more."

s36
00:02:25.240 --> 00:02:26.320
And it was true.

s37
00:02:26.320 --> 00:02:28.280
It's it's a completely different world.

s38
00:02:28.280 --> 00:02:33.880
But, to your surprise and to mine as well, nothing changed, actually.

s39
00:02:33.880 --> 00:02:46.080
My job increased, but the core part of my job, applied AI, remained the exact same and I cannot uh express how surprised I was cuz I thought that it'll be a complete different thing, but apparently it was not.

s40
00:02:46.080 --> 00:02:55.080
So, uh So, then I spoke to other people as well across legal AI and the people coming up with the prop tech firms, which is essentially the real estate tech firms.

s41
00:02:55.080 --> 00:03:00.040
And I realized that everyone is kind of building the applied vertical AI in a very similar way.

s42
00:03:00.040 --> 00:03:03.720
I could see some steps that could be essentially abstracted out.

s43
00:03:03.720 --> 00:03:06.000
And that's what we'll do today.

s44
00:03:06.000 --> 00:03:15.440
So, uh before again delving the deep into that, I received a few reach outs saying, "Are people actually putting agents into production?"

s45
00:03:15.440 --> 00:03:17.720
And I was like, this is such a wrong question to ask.

s46
00:03:17.720 --> 00:03:23.160
Everyone is putting agents into production, even like 15-year-old 15-year-old kids these days.

s47
00:03:23.160 --> 00:03:34.720
The question to ask is whether they actually work, whether they actually make or save money, whether they justify their ROI, whether uh they're making way more than the amount we are investing into it like end-to-end.

s48
00:03:34.720 --> 00:03:37.840
And I can say from my anecdotal experience, yes.

s49
00:03:37.840 --> 00:03:43.480
At the both places I worked, the agent either saved the money or made more money.

s50
00:03:43.480 --> 00:03:47.040
So, with that, let's get started.

s51
00:03:47.040 --> 00:03:54.960
So, the recipe I'll take you through a series of seven steps, roughly, and try to like make this process as simple as possible.

s52
00:03:54.960 --> 00:03:57.880
So, the first step is formulate the problem.

s53
00:03:57.880 --> 00:04:03.080
So, this is sounds like very trivial, but a lot of people, specifically startups, get this wrong.

s54
00:04:03.080 --> 00:04:05.320
They just try to do too much at once.

s55
00:04:05.320 --> 00:04:10.720
Whereas, from what I have learned and what I think a lot of colleagues would agree, you need to pick a very narrow task.

s56
00:04:10.720 --> 00:04:12.840
You just cannot ask it to do everything.

s57
00:04:12.840 --> 00:04:21.560
A good example for this could be, let's say if you build something in finance, you won't ask it to like, "Hey, can you fetch me top three market opportunities that I could invest in?"

s58
00:04:21.560 --> 00:04:22.600
No, that won't work.

s59
00:04:22.600 --> 00:04:24.840
You have to be like very specific.

s60
00:04:24.840 --> 00:04:28.320
Like you pick a market, you say, "Let's take the US equities."

s61
00:04:28.320 --> 00:04:31.160
Then you pick an industry, let's take IT.

s62
00:04:31.160 --> 00:04:40.240
And then you ask it to like rank stocks based on some parameters like capital expenditure or let's say the AI um uh investments.

s63
00:04:40.240 --> 00:04:46.800
So, you pick like very specific things, and then you uh sort of formulate a very narrow job for the AI agent to do.

s64
00:04:46.800 --> 00:04:49.160
And you can build like n number of AI agent.

s65
00:04:49.160 --> 00:04:51.640
Last I checked, there was no tax on building more AI agents.

s66
00:04:51.640 --> 00:04:54.480
So, why do you want your single agent to do everything?

s67
00:04:54.480 --> 00:04:58.600
So, this is important, and this is in the same uh in the pharma context is the exact same.

s68
00:04:58.600 --> 00:05:03.800
We just break down the process into steps, and then ask really pointed questions with the agent.

s69
00:05:03.800 --> 00:05:06.400
We model our agent for a task.

s70
00:05:06.400 --> 00:05:13.120
So, once we have our problem right off the way, we know what we're trying to solve, the next step is identify the data.

s71
00:05:13.120 --> 00:05:14.600
And I cannot stress this enough.

s72
00:05:14.600 --> 00:05:19.960
This is a really, really, really important step, cuz everyone has news data.

s73
00:05:19.960 --> 00:05:24.160
Everyone has like seller side reports from JP Morgan, Morgan Stanley.

s74
00:05:24.160 --> 00:05:29.200
Uh everyone has the arXiv preprint server or PubChem or your research papers, right?

s75
00:05:29.200 --> 00:05:33.919
But what actually makes your application better than let's say ChatGPT or Claude?

s76
00:05:33.919 --> 00:05:35.960
It is your proprietary data.

s77
00:05:35.960 --> 00:05:40.960
So, the thing with proprietary data is it's really expensive to buy, and most people won't sell it to you.

s78
00:05:40.960 --> 00:05:43.000
So, you need to curate it by yourself.

s79
00:05:43.000 --> 00:05:45.520
Imagine your organization has been working for 3 years, right?

s80
00:05:45.520 --> 00:05:47.080
They already have a lot of data.

s81
00:05:47.080 --> 00:05:48.640
It's just unstructured.

s82
00:05:48.640 --> 00:05:55.040
And in the age of LLMs, I think this is a very fairly easy task to make unstructured data into structured data.

s83
00:05:55.040 --> 00:05:57.920
Like a LLM workflow could do it overnight.

s84
00:05:57.920 --> 00:06:10.240
So, to give you a great example of the proprietary data that finance industry has, it's the trade thesis, which is like what trade work and why it worked.

s85
00:06:10.240 --> 00:06:13.400
And in pharma, it is the data for failed experiments.

s86
00:06:13.400 --> 00:06:19.320
For successful experiments data, yes, you can get it, but failed experiments, that's relatively hard to get.

s87
00:06:19.320 --> 00:06:21.960
So, now we have the problem, we have the data.

s88
00:06:21.960 --> 00:06:23.919
What's the third step?

s89
00:06:23.919 --> 00:06:27.200
That is to model the problem, like write the prompt.

s90
00:06:27.200 --> 00:06:35.800
So, while writing prompt, you like what we should aim is to model it after the person who you are trying to replace.

s91
00:06:35.800 --> 00:06:42.680
I mean, that's the hypothesis, but yeah, no offense, we're not trying to replace anyone with AI, but that's the ideology behind writing prompts.

s92
00:06:42.680 --> 00:06:46.400
Encode how a person would solve this job into multiple steps.

s93
00:06:46.400 --> 00:06:48.560
So, it's just like a like a mental model.

s94
00:06:48.560 --> 00:06:51.120
So, this is again fairly simple.

s95
00:06:51.120 --> 00:06:58.840
Next thing, observability, I'm sure you have been in this conference at 3 years and this word, I think I don't know, you'll be hearing about like a thousandth time.

s96
00:06:58.840 --> 00:07:01.160
There are tons of observability provider.

s97
00:07:01.160 --> 00:07:03.520
If you can't see it, you can fix it.

s98
00:07:03.520 --> 00:07:11.680
So, you need observability to see the traces, understand what your uh AI application is doing, and debug it.

s99
00:07:11.680 --> 00:07:15.800
So, sorry, but all of this was the easy part, to be honest.

s100
00:07:15.800 --> 00:07:17.560
All of this fits one screen.

s101
00:07:17.560 --> 00:07:21.760
The mythical 10x engineers can do this stuff in minutes.

s102
00:07:21.760 --> 00:07:25.600
Like literally, this is the code you precisely need to build an AI agent.

s103
00:07:25.600 --> 00:07:27.480
So, that's why it's not the moat.

s104
00:07:27.480 --> 00:07:29.760
Uh of course, except your proprietary data.

s105
00:07:29.760 --> 00:07:31.520
So, what do you do now?

s106
00:07:31.520 --> 00:07:42.160
What do you do after doing the first four steps, which is observability, and prompts, and like uh getting the data right, and everything?

s107
00:07:42.160 --> 00:07:44.040
UI trade.

s108
00:07:44.040 --> 00:07:49.240
Now, the thing with iteration is like when I joined the hedge fund, I thought how hard it can be.

s109
00:07:49.240 --> 00:07:50.680
I mean, everyone can iterate.

s110
00:07:50.680 --> 00:07:54.200
I mean, we have been iterating our whole life for each of the task.

s111
00:07:54.200 --> 00:08:03.520
But, to be honest, I could build it, but I just could not tell if it worked cuz I'm not a trader.

s112
00:08:03.520 --> 00:08:06.880
I'm not someone who has a PhD in biology or chemistry.

s113
00:08:06.880 --> 00:08:12.400
I just don't understand what the model is saying, what is the output of my AI agent is.

s114
00:08:12.400 --> 00:08:15.160
And since most of you are engineers, you would relate.

s115
00:08:15.160 --> 00:08:20.520
You can instantly tell that Sonnet 5 sucks because you have your own training.

s116
00:08:20.520 --> 00:08:23.000
You understand, okay, this code is not great code.

s117
00:08:23.000 --> 00:08:30.040
Whereas, some X model, let's say Fable 5, you see, okay, this is great but not as great as the high base cuz you've been trained for this for life.

s118
00:08:30.040 --> 00:08:33.880
You have a mental model to judge these things.

s119
00:08:33.880 --> 00:08:44.480
But, you just do not have the same kind of mental model when it comes to like predicting trade thesis is or doing like really specific task that vertically our industry does.

s120
00:08:44.480 --> 00:08:55.360
And this is also the place where like a lot of vertical AI projects quietly die because on the surface it looks like you have made it, you have built it, let's put this into production and start selling it.

s121
00:08:55.360 --> 00:09:00.160
But, no one would buy it the same way you won't use an inferior coding model.

s122
00:09:00.160 --> 00:09:04.040
So, as an engineer when I ran into this, I just couldn't accept honestly.

s123
00:09:04.040 --> 00:09:06.160
I thought, no, there's certainly more that I can do.

s124
00:09:06.160 --> 00:09:07.880
We don't need other people.

s125
00:09:07.880 --> 00:09:13.760
So, I thought I could LLM as a judge my way out of it.

s126
00:09:14.136 --> 00:09:14.800
[sighs and laughter]

s127
00:09:14.800 --> 00:09:21.680
And this was a really, really stupid mistake to be honest cuz what LLM is essentially doing, it's it's predicting the next probable word.

s128
00:09:21.680 --> 00:09:24.480
So, if you see, it's just like jargoning its way out.

s129
00:09:24.480 --> 00:09:26.560
It does not understand what alpha means.

s130
00:09:26.560 --> 00:09:31.800
It does not understand how to actually create value unless you have like taught it some way.

s131
00:09:31.800 --> 00:09:36.080
And whereas a human can just tell it instantly what's and what's not.

s132
00:09:36.080 --> 00:09:42.480
So, I'll just try to dwell a bit more deeper on why you can just iterate.

s133
00:09:42.480 --> 00:09:58.880
So, first thing is that model cannot verify itself, specifically in these fields, because reinforcement learning via verifiable rewards is really good at math and code because you have like answer keys, you can verify your code is uh compiling or not, and there are tons of stuff you can just model

s134
00:09:58.880 --> 00:10:00.680
uh the complete thing around this.

s135
00:10:00.680 --> 00:10:04.160
But, when in these fields, there is just no way to model it.

s136
00:10:04.160 --> 00:10:07.880
And And let's say if any error gets in, it's just compounds with every stuff.

s137
00:10:07.880 --> 00:10:11.080
And that's what LeCun seems to think as well.

s138
00:10:11.080 --> 00:10:17.240
And now, the more important part that we touched upon previously, the data.

s139
00:10:17.240 --> 00:10:23.040
So, the interesting thing with pharma and finance is the data was never there.

s140
00:10:23.040 --> 00:10:26.440
And I'll explain to you why.

s141
00:10:27.120 --> 00:10:38.560
So, any institutional manager holding over $100 million in qualifying US equities are forced to publicly file their holdings, long position holdings, every quarter.

s142
00:10:38.560 --> 00:10:48.800
And once a hedge fund does this, this is the percentage decrease in their returns because everyone just sees those reverse engineers and takes away their moat.

s143
00:10:48.800 --> 00:10:51.200
And when it comes to pharma, right?

s144
00:10:51.200 --> 00:10:57.839
So, this is the number of uh so, by law, you are like required to disclose every clinical trial pass or failure you have done.

s145
00:10:57.839 --> 00:11:01.800
But, 30% of the funds, which is like nearly 1/3 of firms, never do.

s146
00:11:01.800 --> 00:11:14.160
And in like 2026, FDA had to like publicly remind over, I don't know, about 2,000 sponsors that they are, I mean, doing injustice by not uh releasing unfavorable results

s147
00:11:14.160 --> 00:11:22.440
because this is the exact data which helps the model thing, which helps your LLM actually reason through these complex and niche industries.

s148
00:11:22.440 --> 00:11:26.520
And they hide it because for them, it's like a chicken laying golden eggs.

s149
00:11:26.520 --> 00:11:28.080
Why would they sell their chicken?

s150
00:11:28.080 --> 00:11:33.160
So, naturally, neither OpenAI nor Anthropic has that has this data because it's like gatekeeper.

s151
00:11:33.160 --> 00:11:41.000
You just cannot hire a trader for $100 an hour and have them annotate that stuff because there's like lots of NDAs and they definitely earn more.

s152
00:11:41.000 --> 00:11:45.520
So, okay, now I have told you about tens of problems, right?

s153
00:11:45.520 --> 00:11:48.960
Now, you would naturally think, okay, yeah, right, then what do we do?

s154
00:11:48.960 --> 00:11:53.080
How do we build a startup in like a vertical space space?

s155
00:11:53.080 --> 00:12:03.120
So, very self-explanatory, you hire the person who you want to sell it to cuz there is, to be honest, no other way around.

s156
00:12:03.120 --> 00:12:04.720
I have tried a lot of stuff.

s157
00:12:04.720 --> 00:12:06.920
You just need to hire the user.

s158
00:12:06.920 --> 00:12:12.240
In finance in a hedge fund, this was very easy because the user was kind of like my boss, the trader.

s159
00:12:12.240 --> 00:12:16.240
We worked together, but in the PharmaTech startup, it was very weird.

s160
00:12:16.240 --> 00:12:22.400
We were like a bunch of young engineers and we were like, oh, we need a 20-year-old scientist in our company to tell us what to do?

s161
00:12:22.400 --> 00:12:24.320
Yeah, I guess we do.

s162
00:12:24.320 --> 00:12:25.760
And then we hired someone, right?

s163
00:12:25.760 --> 00:12:28.920
And that someone actually changed the trajectory of our tools.

s164
00:12:28.920 --> 00:12:30.480
Our tools started making sense.

s165
00:12:30.480 --> 00:12:41.000
When we pitched to the other pharma companies, the big ones, the big pharma, they started liking our tools because it's kind of spoke their language versus the normal jargonish LLM language.

s166
00:12:41.000 --> 00:12:47.640
So, once you have hired the user, let's say, then what would you make that user do?

s167
00:12:47.640 --> 00:12:50.400
You try to build a learning loop out of it.

s168
00:12:50.400 --> 00:12:56.560
The domain expert can start at the like a very, very low level, the ground level, where they just think about prompts.

s169
00:12:56.560 --> 00:13:00.760
Okay, yeah, I mean, let's not ask LLM to do this.

s170
00:13:00.760 --> 00:13:03.080
Let's ask a very specific query again.

s171
00:13:03.080 --> 00:13:04.640
They'll help you curate data.

s172
00:13:04.640 --> 00:13:15.760
Just like engineers know which conferences are which are not, which research paper sites are great, which are not, which are like top leaders in engineering, which is which are just like influencers.

s173
00:13:15.760 --> 00:13:20.720
Similarly, a pharma expert or let's say a trader knows which sources are more reliable than the other.

s174
00:13:20.720 --> 00:13:22.120
So, they help you create their data.

s175
00:13:22.120 --> 00:13:29.240
They help you like refine your prompts better, and they try to create like thinking models of how they would think about a problem.

s176
00:13:29.240 --> 00:13:33.960
Cuz I mean, let's say if you if you follow five steps to solve a problem, right?

s177
00:13:33.960 --> 00:13:36.440
You just cannot do it in in any random order.

s178
00:13:36.440 --> 00:13:37.760
There has to be a logical flow.

s179
00:13:37.760 --> 00:13:38.920
There has to be a natural flow.

s180
00:13:38.920 --> 00:13:46.320
That's So, that's what they uh try to curate like decompose a problem, gradually refine, and then finally judge.

s181
00:13:46.320 --> 00:13:56.640
So, the person who sort of has lived through the complete of the industry that they're trying to revolutionize, their judgment is now like turning into agents.

s182
00:13:56.640 --> 00:13:58.880
So, that's what's happening behind the loop.

s183
00:13:58.880 --> 00:14:02.040
So, uh to do this there are like again multiple ways.

s184
00:14:02.040 --> 00:14:05.720
I mean, each of these could have been a hour-long session on its own.

s185
00:14:05.720 --> 00:14:10.560
And I wish I could take, but these are like few ways that I identified.

s186
00:14:10.560 --> 00:14:14.920
Uh I'll just like take uh you through them like really quickly in the interest of time.

s187
00:14:14.920 --> 00:14:20.760
So, supervised fine-tuning I think most of you would know where like model mimics human nest demonstrations.

s188
00:14:20.760 --> 00:14:28.680
Uh reinforcement learning from human feedback is like a kind of uh a very efficient way where human preferences train a reward model.

s189
00:14:28.680 --> 00:14:34.920
Then rubrics as a reward is I I like to call it reinforcement learning from AI feedback.

s190
00:14:34.920 --> 00:14:44.720
This is because that you can human can just create a rubric, and then AI will just like grade itself based on that rubric, and that improve its own processes.

s191
00:14:44.720 --> 00:14:49.760
But again, there is a slight chance that you might run into an echo chamber with rubrics as rewards.

s192
00:14:49.760 --> 00:14:54.480
And the cheapest of all, and I think the highest ROI is the error analysis.

s193
00:14:54.480 --> 00:14:59.320
Whereas the observability part that you set up earlier, you just analyze the logs plain and simple.

s194
00:14:59.320 --> 00:15:03.040
You understand where model is going wrong, and then you just try to correct it.

s195
00:15:03.040 --> 00:15:12.160
So, this is where you have like don't have to touch any weights, and the most highest ROI way to get the impact from a like start on.

s196
00:15:12.160 --> 00:15:23.520
And once you understand like uh what more you could do, or if error analysis is solving or not, you can just gradually climb up the ladder, and probably uh later on go to the ultimate reinforcement

s197
00:15:23.520 --> 00:15:32.960
learning from human feedback cuz that's I think in our industry kind of the golden standard these days that you need to do RLHF to actually get some edge.

s198
00:15:32.960 --> 00:15:35.880
But uh certainly there are some pitfalls of it.

s199
00:15:35.880 --> 00:15:39.600
Like for example, now there's GLM 5.2, right?

s200
00:15:39.600 --> 00:15:42.080
You fine-tuned it, right?

s201
00:15:42.080 --> 00:15:46.600
Uh Alibaba Cloud or let's say Deep Seek will release a newer model, then you have to fine-tune that too as well.

s202
00:15:46.600 --> 00:15:47.560
So there is a cost.

s203
00:15:47.560 --> 00:15:49.480
It's not cheap.

s204
00:15:49.480 --> 00:15:54.120
So once you have done all this, you just create a loop and you just like go on to that loop.

s205
00:15:54.120 --> 00:16:00.880
You hired one user, you hire more users, they ask more queries, the scoping increases, the data increases.

s206
00:16:00.880 --> 00:16:03.320
At this point you're kind of generating your own data.

s207
00:16:03.320 --> 00:16:05.760
The exercise you have been doing in loop, right?

s208
00:16:05.760 --> 00:16:11.720
That exercise itself is generating a very I would say a crazy data set of what works and what does not work.

s209
00:16:11.720 --> 00:16:13.400
And this loop never stops.

s210
00:16:13.400 --> 00:16:21.760
Once you feel confident enough in your application, you just ship it, provide it to the external paying users, and then you see the magic of it that it actually works.

s211
00:16:21.760 --> 00:16:30.360
So I just pulled the stat from Stanford AI Index report cuz it's a really nice report that gives you an idea of what the state of AI is.

s212
00:16:30.360 --> 00:16:35.120
And this says like 80% 89% of enterprise AI agents never reach production.

s213
00:16:35.120 --> 00:16:36.920
Again, I disagree.

s214
00:16:36.920 --> 00:16:42.040
Every AI reaches production, but it just fails to work or like justify its own cost.

s215
00:16:42.040 --> 00:16:43.520
So that's the real thing.

s216
00:16:43.520 --> 00:16:48.240
You can just build and ship AI agents whenever you want, but you need to justify ROI.

s217
00:16:48.240 --> 00:16:54.240
And finance and pharma are two such industries where if it does not make money, it's shown the door.

s218
00:16:54.240 --> 00:16:54.880
Simple.

s219
00:16:54.880 --> 00:16:57.680
They won't like wait and say, "Okay, maybe it'll work in 2 years.

s220
00:16:57.680 --> 00:17:00.160
Maybe the cost will be lower in by the third year."

s221
00:17:00.160 --> 00:17:02.200
No. It has to instantly make money.

s222
00:17:02.200 --> 00:17:03.600
It has to like hit the ground running.

s223
00:17:03.600 --> 00:17:06.280
And if it does not, shown the door instantly.

s224
00:17:06.280 --> 00:17:15.800
So just to summarize the seven steps that I feel are like good enough to give you an abstraction of how the vertical AI industry moves.

s225
00:17:15.800 --> 00:17:27.240
You formulate the problem statement, you source your data sources, you prompt it well, you define those prompts, you observe how your tool is performing, you don't iterate yet, you hire the user.

s226
00:17:27.240 --> 00:17:31.080
And this user or users now play with the tool as much as possible.

s227
00:17:31.080 --> 00:17:43.680
They like kind of form a learning loop, an endless learning loop that goes on and at a point when you feel yeah, it's it's really delivering that alpha over let's say Claude and ChatGPT, you just ship it, you start earning money.

s228
00:17:43.680 --> 00:17:46.480
So, one more interesting thing.

s229
00:17:46.480 --> 00:17:50.960
So, HITL is like kind of a thing everyone is like yeah, let's add human in the loop.

s230
00:17:50.960 --> 00:17:52.320
I would say not yet.

s231
00:17:52.320 --> 00:18:03.400
Finance and pharma are still those two industries where it's AITL, AI in the loop cuz everything is like done by the expert, but the AI assistant really helps save time.

s232
00:18:03.400 --> 00:18:18.760
Like for example, uh it may take an X amount for a trader to form different trade thesis, and AI can just give him five candidate trade thesis, but which one would actually work in the market and which won't is the discussion the discussion still lies with the trader.

s233
00:18:18.760 --> 00:18:29.240
And same for pharma when you're like picking drug candidates, which one to pick, the expert still does it, but you just like reduce the time of expert by a lot lot.

s234
00:18:29.240 --> 00:18:32.200
So, and and it will stay this way for really long.

s235
00:18:32.200 --> 00:18:39.160
So, uh for the models to actually make good decisions, they don't need to do correlation, they need to do causation.

s236
00:18:39.160 --> 00:18:45.760
And as Ya as Jan LeCun puts it, these are like text statistics, not real-world models.

s237
00:18:45.760 --> 00:18:49.520
You cannot just pattern match with past and use future to predict to it.

s238
00:18:49.520 --> 00:18:51.560
And so, we are like kind of not there yet.

s239
00:18:51.560 --> 00:18:54.040
That's what I call as the AGI line.

s240
00:18:54.040 --> 00:18:56.880
Once we are there, yeah, probably then models will just like make drugs.

s241
00:18:56.880 --> 00:18:58.680
You will have vibe coded drugs.

s242
00:18:58.680 --> 00:19:02.600
Someone would be vibe coding market, but yeah, not yet.

s243
00:19:02.600 --> 00:19:09.920
So, a final takeaway that I would call if if if there's one thing you are taking away from this talk, this is it.

s244
00:19:09.920 --> 00:19:15.840
Model infra ecosystem, everyone selling you tons of stuff at this conference is just commodity.

s245
00:19:15.840 --> 00:19:16.680
Everyone has it.

s246
00:19:16.680 --> 00:19:17.920
If you have it, everyone has it.

s247
00:19:17.920 --> 00:19:21.120
Everyone can pay X number of dollars for a subscription.

s248
00:19:21.120 --> 00:19:24.720
But, what is moat and no one will come and sell it to you.

s249
00:19:24.720 --> 00:19:28.040
You won't have to curate it on your own is the domain expertise.

s250
00:19:28.040 --> 00:19:28.920
You need your data.

s251
00:19:28.920 --> 00:19:29.960
You need other people's data.

s252
00:19:29.960 --> 00:19:32.400
That is just not out there on the internet.

s253
00:19:32.400 --> 00:19:34.800
And that that's what will form your moat.

s254
00:19:34.800 --> 00:19:36.520
So, thank you for your time.

s255
00:19:36.520 --> 00:19:40.000
I think you enjoyed the talk and yeah, let me know if you have any questions.

s256
00:19:40.000 --> 00:19:41.040
We can meet outside.

s257
00:19:41.040 --> 00:19:42.943
Thank you.

s258
00:19:42.943 --> 00:19:44.943
[applause]

s259
00:19:58.663 --> 00:20:00.663
[music]