WEBVTT

NOTE Sentence-level transcript of https://www.youtube.com/watch?v=WOszaC1WWbQ

NOTE One cue per sentence. Cue ids are the line anchors on /transcripts/WOszaC1WWbQ.html. A cue ends where the next begins, or 2 s after its last word.

s1
00:00:13.040 --> 00:00:14.548
Thank you to everyone who came.

s2
00:00:14.548 --> 00:00:18.840
I am very pleased to deliver this report.

s3
00:00:18.840 --> 00:00:22.105
So, a brief introduction about me.

s4
00:00:22.105 --> 00:00:23.947
My name is Harnit.

s5
00:00:23.947 --> 00:00:28.099
I am the founder and CEO of Axon Intelligence.

s6
00:00:28.099 --> 00:00:33.232
At Axon, we create an intelligent layer for investing in distressed loans.

s7
00:00:33.232 --> 00:00:37.557
This is also what we will talk about today.

s8
00:00:37.557 --> 00:00:43.127
The title of my talk: Can AI understand problem loans?

s9
00:00:45.110 --> 00:00:53.732
I think the first question that arises for everyone who is not familiar with loans is: what does a problem loan actually mean?

s10
00:00:55.186 --> 00:00:57.260
Credit is pretty simple.

s11
00:00:57.260 --> 00:01:01.749
It is simply a bond or loan from a corporation.

s12
00:01:01.749 --> 00:01:06.415
A Coca-Cola bond is a loan.

s13
00:01:06.415 --> 00:01:10.618
And what does it mean to be under pressure or in a difficult situation?

s14
00:01:10.618 --> 00:01:14.441
Is n't that right?

s15
00:01:14.441 --> 00:01:25.951
Pressure arises if a company has some underlying risk that could cause them to run out of money or be unable to repay their debt.

s16
00:01:25.951 --> 00:01:34.674
Every year, a certain portion of the credit market faces some level of pressure or problems.

s17
00:01:34.674 --> 00:01:38.410
This may be caused by cyclical problems.

s18
00:01:38.410 --> 00:01:42.148
Let's say the construction market is in decline, the housing market is in decline.

s19
00:01:42.148 --> 00:01:45.324
This may be caused by operational issues.

s20
00:01:45.324 --> 00:01:49.408
For example, if you are a restaurant business and there is a large-scale product recall.

s21
00:01:49.408 --> 00:01:53.082
This may also be caused by long-term structural changes .

s22
00:01:53.082 --> 00:01:58.389
If you are in software today and are facing the impact of AI.

s23
00:01:58.389 --> 00:02:01.972
Some companies overcome difficulties and recover.

s24
00:02:01.972 --> 00:02:06.032
Some companies face challenges and find themselves in difficult situations.

s25
00:02:06.032 --> 00:02:08.027
What does a difficult situation mean?

s26
00:02:08.027 --> 00:02:10.328
In reality, it's not that simple.

s27
00:02:10.328 --> 00:02:15.558
Bankruptcy is one of the consequences of a difficult situation.

s28
00:02:17.671 --> 00:02:28.710
Liability management measures are also possible—I can explain these terms in detail, but just know that this is another option, as well as an out-of-court restructuring or an agreed exchange.

s29
00:02:30.537 --> 00:02:32.716
So, next slide .

s30
00:02:32.716 --> 00:02:39.673
As you can see, it's not that simple, is it?

s31
00:02:39.673 --> 00:02:43.856
Historically, the issue of predicting problems has been quite binary.

s32
00:02:43.856 --> 00:02:51.458
Bankruptcy could be predicted : either you predict default or you don't.

s33
00:02:51.458 --> 00:02:55.800
Today it has become much more difficult.

s34
00:02:57.277 --> 00:03:10.527
Many traditional machine learning methods simply can't keep up with this, nor can the assessments that were important 10, 15, or 20 years ago.

s35
00:03:10.527 --> 00:03:22.698
Um, uh, the main problem is that Altman's Z-score has become very static in a situation that changes, you know, sometimes monthly, quarterly, or annually, and in a much more opaque

s36
00:03:22.698 --> 00:03:24.340
market.

s37
00:03:24.340 --> 00:03:33.688
So, if you've been following the credit market, the growth of private lending and some other sectors has led to, you know, a lack of pricing.

s38
00:03:33.960 --> 00:03:49.519
On the other hand, there is a very high-quality method of understanding and analyzing problem loans, when you, as an investor in problem debt, can analyze financial statements, understand what is happening on the part of the lender, and read loan documentation.

s39
00:03:49.993 --> 00:03:57.879
Basically, you know, all the things a typical hedge fund or private equity investor would do to understand credit.

s40
00:03:57.879 --> 00:04:01.878
This process is extremely manual and, of course, very difficult to scale.

s41
00:04:01.878 --> 00:04:03.880
Isn't that right ?

s42
00:04:03.880 --> 00:04:16.371
Now the question is , you know, part of the founding of Axon and the question that I've been thinking about is how far AI can go in understanding distressed

s43
00:04:16.371 --> 00:04:19.303
credit and what paths can be taken.

s44
00:04:19.303 --> 00:04:27.871
And is it possible to systematize what is actually a rather complex process, right?

s45
00:04:27.871 --> 00:04:34.239
So, what does it mean to understand a non-performing loan, right?

s46
00:04:34.239 --> 00:04:41.738
As investors, we have many questions that we are trying to find answers to.

s47
00:04:41.738 --> 00:04:46.274
Things you think about from a financial perspective, right?

s48
00:04:47.309 --> 00:04:49.719
Do you know what liquidity is?

s49
00:04:49.719 --> 00:04:51.088
What is the capital structure?

s50
00:04:51.088 --> 00:04:54.830
And then there are legal questions.

s51
00:04:54.830 --> 00:04:57.979
How do you know what the structural priority of the debt is ?

s52
00:04:57.979 --> 00:04:58.786
Isn't that right?

s53
00:04:58.786 --> 00:05:02.449
That is, do you have a higher or lower priority?

s54
00:05:02.449 --> 00:05:05.835
And then there are operational questions, like: how do you understand business?

s55
00:05:05.835 --> 00:05:07.994
What are the revenues there?

s56
00:05:07.994 --> 00:05:14.734
Will revenues go up or down tomorrow, based on fundamental, you know, fundamental business analysis.

s57
00:05:14.734 --> 00:05:16.628
Is n't that right?

s58
00:05:16.628 --> 00:05:27.772
Right now, what we're seeing in the market, if you follow Pro Go, Harvey, or some of these big AI companies, they're actually thinking about how to compare all these different

s59
00:05:27.772 --> 00:05:29.731
issues, right?

s60
00:05:29.731 --> 00:05:37.436
For example, how well does AI answer the question of which portion of debt is structurally senior?

s61
00:05:37.436 --> 00:05:45.512
And then do thorough comparative testing and make it, let's call it, a big financial benchmark.

s62
00:05:45.512 --> 00:05:50.345
But I wanted to think about it a little differently, didn't I ?

s63
00:05:50.345 --> 00:06:00.593
Because I was thinking this way: the real understanding of credit, especially distressed credit, is whether you can predict what will happen next?

s64
00:06:00.593 --> 00:06:02.502
Isn't that right ?

s65
00:06:02.502 --> 00:06:08.368
Can I figure out and get a probability distribution of which way this will go in the future?

s66
00:06:08.368 --> 00:06:15.024
Um, so I thought this was a pretty interesting exercise.

s67
00:06:15.024 --> 00:06:20.873
The first step, you know, what you should do is take all the, you know, I chose the company.

s68
00:06:20.873 --> 00:06:27.601
So, I chose iHeartMedia, a company that filed for bankruptcy in 2018 .

s69
00:06:27.601 --> 00:06:31.597
So she, you know, she went through all these bureaucratic hoops.

s70
00:06:31.597 --> 00:06:36.466
Um, she also had another stressful transaction in 2024.

s71
00:06:36.466 --> 00:06:39.039
So this is a real-time situation.

s72
00:06:39.039 --> 00:06:41.499
This is happening today.

s73
00:06:41.499 --> 00:06:46.147
Nobody has a real, you know, answer to what the future holds, right?

s74
00:06:46.147 --> 00:06:49.515
Um, there is no absolute truth.

s75
00:06:49.515 --> 00:06:53.778
But what I did was this: let's throw this into the smartest, you know, models.

s76
00:06:53.778 --> 00:07:06.569
I used Astra 6 on Max Intelligence and Opus 5.5. And the conclusion that I found very interesting is that they disagreed.

s77
00:07:06.569 --> 00:07:17.011
They didn't agree on what the probabilities of the future would be, given the structure or context of the data, right?

s78
00:07:17.011 --> 00:07:28.657
And so, if you look here at the agreed restructuring, Astra thought it was the best option for bankruptcy, and Opus thought it was the best option, right?

s79
00:07:28.657 --> 00:07:34.986
Now, how do you tell who is right and who is wrong?

s80
00:07:34.986 --> 00:07:38.760
You can't, because I don't know what the probabilities are, right?

s81
00:07:38.760 --> 00:07:40.346
I don't know what will happen.

s82
00:07:40.346 --> 00:07:42.236
I don't know if it will be a bankruptcy.

s83
00:07:42.236 --> 00:07:45.097
I don't know if this will be an out-of-court restructuring.

s84
00:07:45.097 --> 00:07:48.194
I do n't know if this will be a refinancing.

s85
00:07:48.194 --> 00:07:50.667
I have my own judgment on these things.

s86
00:07:50.667 --> 00:07:56.325
I would say I'm leaning towards Astra 6, but, you know, we're all human.

s87
00:07:56.325 --> 00:07:58.570
And that's a pretty interesting conclusion, right?

s88
00:07:58.570 --> 00:08:03.142
Like, okay, wait, they disagree.

s89
00:08:03.142 --> 00:08:11.330
Um, so the second thing I thought about was: okay, how can I at least improve their process assessment, right?

s90
00:08:11.330 --> 00:08:18.224
So I can see whether I agree or disagree with different assumptions.

s91
00:08:18.486 --> 00:08:27.252
Um, and so, essentially, you've created a kind of benchmark where we test every level.

s92
00:08:29.627 --> 00:08:33.913
Obviously, mistakes can be made at every level of the LLM, right?

s93
00:08:33.913 --> 00:08:39.475
Regarding the level of data extraction, I would say that with public companies this issue is already pretty much resolved.

s94
00:08:39.475 --> 00:08:47.292
I would say with private loans, and also when you have Excel and PDF documents, it can still be a challenge, right?

s95
00:08:47.292 --> 00:08:50.286
Then you have the reconciliation or data output level.

s96
00:08:50.286 --> 00:08:52.982
So, extraction is, for example, obtaining the EBITDA indicator.

s97
00:08:52.982 --> 00:08:58.156
Uh, the conclusion would be : okay, now you have EBITDA and you have debt, so you have a credit load.

s98
00:08:58.156 --> 00:09:00.137
And then you have judgment.

s99
00:09:00.137 --> 00:09:02.098
Okay, now you have a credit load.

s100
00:09:02.098 --> 00:09:06.082
Is this a problem loan or not?

s101
00:09:07.178 --> 00:09:18.466
What is your Oh.

s102
00:09:19.562 --> 00:09:27.158
Uh, is there any...okay, I'll just, I'll just keep talking.

s103
00:09:27.158 --> 00:09:42.736
So, uh, actually, the question is , is it like my computer has shut down?

s104
00:09:43.066 --> 00:09:47.798
Uh, excuse me, just one second .

s105
00:09:47.798 --> 00:09:56.645
I'll tell you I have a charger, but it's not here.

s106
00:09:57.913 --> 00:10:03.181
Uh, yes.

s107
00:10:05.344 --> 00:10:08.022
Okay, I'll just, I'll just talk.

s108
00:10:08.022 --> 00:10:13.245
This is normal.

s109
00:10:13.245 --> 00:10:14.018
Oh, okay.

s110
00:10:14.018 --> 00:10:16.686
It works.

s111
00:10:19.863 --> 00:10:21.149
OK.

s112
00:10:21.149 --> 00:10:25.366
So, therefore, you have a level of judgment , right?

s113
00:10:25.366 --> 00:10:32.698
And then, uh, then, then you can start, okay , if you can evaluate each level, right?

s114
00:10:32.698 --> 00:10:45.246
When you get to the prediction level, you've made sure the data points are right, you've made sure the, uh, level of agreement is right, you've made sure there's some kind of reasonable judgment,

s115
00:10:45.246 --> 00:10:46.134
right?

s116
00:10:46.134 --> 00:10:57.848
And then , once you're confident in those things, then you can move to the forecasting level, which is really how, you know, you as a human analyst think about all of these steps, right?

s117
00:10:57.848 --> 00:11:07.346
You, you, if you base a judgment on a bad data point, it's not going to be a good judgment, and that's, that's just how it works, right?

s118
00:11:07.346 --> 00:11:14.087
So, uh, Op.

s119
00:11:14.087 --> 00:11:17.966
Here we are.

s120
00:11:17.966 --> 00:11:29.827
So, overall, what we found, and maybe this isn't surprising, but, uh, I think it was interesting, is that once we created a structured reasoning process, you , you actually got both intellects,

s121
00:11:29.827 --> 00:11:42.039
both models, to actually converge on what they thought the future probabilities for the business would be, right?

s122
00:11:42.039 --> 00:11:56.931
So now, instead of talking about a 62% bankruptcy with Opus and a 25% with Astra, now both of these models say that there is actually only a quarter, or a 25% chance of bankruptcy, and both models agreed that the path of concerted

s123
00:11:56.931 --> 00:12:02.085
restructuring is actually the more likely path for the business.

s124
00:12:02.085 --> 00:12:07.112
So, what are the main conclusions here, right ?

s125
00:12:07.112 --> 00:12:14.967
Uh, one, I would say, first conclusion: if you give questions that cannot be verified , right?

s126
00:12:14.967 --> 00:12:25.154
With the same structured context, organized context, as humans, I mean , do these models have an advantage in prediction?

s127
00:12:25.154 --> 00:12:28.394
And you have to say "no", right?

s128
00:12:28.394 --> 00:12:33.153
Because they are, in fact, even agents— they are people, right?

s129
00:12:33.153 --> 00:12:37.641
In some sense of the word, and I won't get into an argument about that .

s130
00:12:37.641 --> 00:12:55.943
Uh, the second thing is, but a structured reasoning process can force two models or, you know, different intellects to actually come to the same vision of the future and a similar understanding of the business and the path it

s131
00:12:55.943 --> 00:12:58.193
can take.

s132
00:12:58.193 --> 00:13:02.554
Hmm, but is this the right way ?

s133
00:13:02.554 --> 00:13:04.425
We don't know, do we ?

s134
00:13:04.425 --> 00:13:07.856
Um, and is that really the single most accurate measurement, right?

s135
00:13:07.856 --> 00:13:13.491
Otherwise, any historical analysis would create a certain distortion by predicting the future.

s136
00:13:13.491 --> 00:13:22.085
Um, but I think the really interesting thing that I was thinking about is that convergence is also quite interesting, right ?

s137
00:13:22.085 --> 00:13:30.290
Because if you can systematically build a process that actually converges to the same probabilistic outcomes.

s138
00:13:30.290 --> 00:13:33.607
This, in fact, is the market price.

s139
00:13:33.607 --> 00:13:38.730
Um, and that's why I thought that was an interesting conclusion.

s140
00:13:38.730 --> 00:13:42.079
Um, so this is the last one.

s141
00:13:42.079 --> 00:13:43.598
And now briefly about me.

s142
00:13:43.598 --> 00:13:46.454
So, my name is , as I said, Harini.

s143
00:13:46.454 --> 00:13:48.158
Um, building an Axon.

s144
00:13:48.158 --> 00:13:49.872
Um, a little bit about my background.

s145
00:13:49.872 --> 00:14:00.817
I spent 10 years in finance, um , the last few years as a distressed and special situations investor at a, uh, private equity/ distressed asset hedge fund.

s146
00:14:00.817 --> 00:14:04.123
Um, here's the QR code if you're interested in Axon.

s147
00:14:04.123 --> 00:14:06.188
Um, and that's all I had.

s148
00:14:06.188 --> 00:14:08.742
Thank you.