WEBVTT

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

NOTE One cue per sentence. Cue ids are the line anchors on /transcripts/tJFjeMBKbIY.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
00:00:12.920 --> 00:00:13.840
Good afternoon.

s3
00:00:13.840 --> 00:00:19.400
Thank you for being here, day four of our conference, a room with no windows.

s4
00:00:19.400 --> 00:00:21.440
Right just after lunch.

s5
00:00:21.440 --> 00:00:24.560
You are the strongest people in this building.

s6
00:00:24.560 --> 00:00:26.720
Let me start with a confession.

s7
00:00:26.720 --> 00:00:30.120
For 15 years, my job has been one thing.

s8
00:00:30.120 --> 00:00:43.920
I sitting in a room, a very smart, confident person hands me a piece of paper, and before my company spends a hundred million dollars, I have to decide, do I believe this paper?

s9
00:00:43.920 --> 00:00:48.800
This year, my job still exactly that.

s10
00:00:49.960 --> 00:00:58.360
Except now, some of time, the very confident person handing me the paper is a chatbot.

s11
00:00:58.360 --> 00:01:13.360
And honestly, the chatbot is often better written than humans, better grammar, nicer formatting, never get defensive when you ask a follow-up question, very very sure of itself.

s12
00:01:13.360 --> 00:01:19.440
The problem is being sure of yourself and being right are two different skills.

s13
00:01:19.440 --> 00:01:26.240
Some of most the confident people I have ever met in finance were also the most wrong.

s14
00:01:26.240 --> 00:01:31.640
AI just learned that trick faster than rest of us.

s15
00:01:31.640 --> 00:01:35.960
So, here my whole war whole talk in two sentence.

s16
00:01:35.960 --> 00:01:52.440
First, almost every AI finance product is built to impress people for five minutes, and almost none are built to survive a room whose entire job is to not be impressed.

s17
00:01:52.440 --> 00:02:09.000
Second, and this is the part I promised the organizers added the exact same skills that fix your product is the skills that gets investors like me to write you a check.

s18
00:02:09.000 --> 00:02:10.240
Same muscle.

s19
00:02:10.240 --> 00:02:11.800
I've proved both.

s20
00:02:11.800 --> 00:02:21.520
And I'm telling this now, this week, because a lot of you are about to get pulled into exactly this.

s21
00:02:21.520 --> 00:02:24.160
Your CEO saw a demo somewhere.

s22
00:02:24.160 --> 00:02:34.400
Your biggest comes a customer suddenly has a compliance department, or you are 3 months from raising your next round.

s23
00:02:34.400 --> 00:02:50.360
When any of those days arrive, I'd rather you to hear the hard part from someone who sit on the other side of the table, than discover them live in a room in front of the people

s24
00:02:50.360 --> 00:02:53.800
who bill by hour.

s25
00:02:55.200 --> 00:02:58.360
A quick quick a bit about me.

s26
00:02:58.360 --> 00:03:02.840
I promise this is the boring part, and I will keep it fast.

s27
00:03:02.840 --> 00:03:10.560
15 years of cross-border deals, Hong Kong, mainland China, the UK, the US.

s28
00:03:10.800 --> 00:03:26.239
Mergers, IPOs, big strategic investment on names you'd actually recognize, the kind of companies that go public, and your LinkedIn feed won't shut up and about it for a week.

s29
00:03:26.239 --> 00:03:32.959
Along the way, I've sit in about 200 investment committee meetings.

s30
00:03:32.959 --> 00:03:36.400
I have the grey hairs to prove it.

s31
00:03:36.400 --> 00:03:37.560
I checked.

s32
00:03:37.560 --> 00:03:40.040
It's not genetics.

s33
00:03:40.720 --> 00:03:47.120
And here the part that matters for second half of this talk.

s34
00:03:47.120 --> 00:03:55.880
I've also read hundreds and hundreds of pitch decks, founders' decks, bankers' decks.

s35
00:03:55.880 --> 00:03:59.440
47 slide seed decks.

s36
00:03:59.440 --> 00:04:04.317
I have seen fonts that should be illegal.

s37
00:04:04.317 --> 00:04:05.160
[sighs and gasps]

s38
00:04:05.160 --> 00:04:10.400
200 committee meetings taught me one thing.

s39
00:04:10.400 --> 00:04:12.520
No textbook says outlawed.

s40
00:04:12.520 --> 00:04:16.799
The number on the page is not what gets a deal approved.

s41
00:04:16.799 --> 00:04:18.720
Trust is a number.

s42
00:04:18.720 --> 00:04:20.760
What gets a deal approved?

s43
00:04:20.760 --> 00:04:24.120
Money doesn't follow intelligence.

s44
00:04:24.120 --> 00:04:39.520
Money follows trust, and the trust is fragile, especially when the thing that wrote the page has never once in its entire life said the words, "I'm not sure."

s45
00:04:39.520 --> 00:04:45.360
Let me give you one small taste of what those rooms feel like.

s46
00:04:45.360 --> 00:04:56.360
Early in my career, a very polished, very expensive banker present a beautiful slides full of confident wrong numbers.

s47
00:04:56.360 --> 00:05:03.360
One senior person in room asked a one quiet question, "Where does this number come from?"

s48
00:05:03.360 --> 00:05:08.600
The banker posed uh what felt like an entire physical quarter.

s49
00:05:08.600 --> 00:05:13.919
The pose taught me more about finance than 3 years of exams did.

s50
00:05:13.919 --> 00:05:20.080
So, I'm not sure as a So, I'm not here as a builder.

s51
00:05:20.080 --> 00:05:22.960
I don't build these systems.

s52
00:05:22.960 --> 00:05:32.080
I sit across the table from them and from the founders selling them, deciding whether to trust them.

s53
00:05:32.080 --> 00:05:37.680
Today, I'm going to give you both halves of that.

s54
00:05:37.680 --> 00:05:44.400
What breaks trust in your product and uh what builds trust in your pitch.

s55
00:05:46.400 --> 00:05:51.560
Let's define two machines that moves everyone keep confusing.

s56
00:05:51.560 --> 00:05:53.880
Machine one is a demo.

s57
00:05:53.880 --> 00:06:03.200
One clean document in, one fluent answer out is the whole job is to make a room go oh for 5 minutes.

s58
00:06:03.200 --> 00:06:06.040
Machine two is a is a memo.

s59
00:06:06.040 --> 00:06:11.480
The real document is a real committee race before real money moves.

s60
00:06:11.480 --> 00:06:22.120
Hundreds of pages, filings, transcripts, broken nose, scratches, someone's rushed nose from a call last Tuesday.

s61
00:06:22.120 --> 00:06:31.240
Half the sources disagree with each other and the memo's job is not to make you go oh.

s62
00:06:31.240 --> 00:06:34.200
Its job's to survive an argument.

s63
00:06:34.200 --> 00:06:39.480
Basically, a family dinner except somebody's uncle brought a spreadsheet.

s64
00:06:39.480 --> 00:06:42.840
Here's everyday worship of the gap.

s65
00:06:42.840 --> 00:06:46.000
Ask your phone to summarize a long email.

s66
00:06:46.000 --> 00:06:47.040
That's a demo.

s67
00:06:47.040 --> 00:06:52.680
It's not just sound plausible for 10 seconds.

s68
00:06:52.680 --> 00:07:04.080
Now, imagine that some same phone has to stand in front of a bank and defend out loud why you deserve a mortgage.

s69
00:07:04.080 --> 00:07:06.760
Suddenly, plausible isn't enough.

s70
00:07:06.760 --> 00:07:12.160
Now, it has to be right and it has to prove it.

s71
00:07:12.160 --> 00:07:17.440
That second situation is what a memo actually is.

s72
00:07:19.440 --> 00:07:24.400
Now, you might think the demo world and the memo world never touch.

s73
00:07:24.400 --> 00:07:29.040
Let me tell you about the most expensive typo in history.

s74
00:07:29.040 --> 00:07:39.600
February 2023, one of the biggest tech company on the earth launches its shiny new AI assistant with a promotional demo.

s75
00:07:39.600 --> 00:07:49.800
In that demo, the assistant answers a simple question about a space telescope and it gets a wrong the wrong answer.

s76
00:07:49.800 --> 00:08:03.320
One sentence one wrong fact about a telescope in um marketing demo, the market noticed the company's stock dropped around 8% in a day.

s77
00:08:03.320 --> 00:08:10.120
That's roughly $100 billion of value gone because of unchecked sentence.

s78
00:08:10.120 --> 00:08:13.560
$100 billion for one sentence.

s79
00:08:13.560 --> 00:08:19.080
Nobody in that company asked the one question.

s80
00:08:19.080 --> 00:08:27.919
Every junior analyst on my team is trained to ask before anything leaves the building, wait.

s81
00:08:27.919 --> 00:08:30.360
Where does this claim come from?

s82
00:08:30.360 --> 00:08:32.520
Did everyone check it?

s83
00:08:32.520 --> 00:08:34.840
So, here's the punchline.

s84
00:08:34.840 --> 00:08:38.280
Even the demo failed the memo test.

s85
00:08:38.280 --> 00:08:44.839
The moment real money is watching and the real money is always watching.

s86
00:08:44.839 --> 00:08:48.200
Every sentence become a memo sentence.

s87
00:08:48.200 --> 00:08:51.480
There is no safe demo anymore.

s88
00:08:51.480 --> 00:09:05.520
Keep that story in your head because of the same trust breaking moment happens in six smaller, quieter, very predictable ways inside of your product every day.

s89
00:09:05.520 --> 00:09:08.839
Let's go through them fast.

s90
00:09:10.560 --> 00:09:15.080
Six ways trust quietly breaks.

s91
00:09:15.080 --> 00:09:17.520
Which source do you believe?

s92
00:09:17.520 --> 00:09:19.440
Do the numbers agree?

s93
00:09:19.440 --> 00:09:23.920
Do you hide contradictions or show them?

s94
00:09:23.920 --> 00:09:27.160
Is that the a fact or a guess?

s95
00:09:27.160 --> 00:09:30.680
Can you prove it in 30 seconds?

s96
00:09:30.680 --> 00:09:34.280
And whose name is actually on the decision?

s97
00:09:34.280 --> 00:09:35.640
Six.

s98
00:09:35.640 --> 00:09:37.720
Keep count with me.

s99
00:09:37.720 --> 00:09:42.920
I will keep each one short and I will bring receipts.

s100
00:09:45.600 --> 00:09:47.240
Model one.

s101
00:09:47.240 --> 00:09:55.080
Not every source deserves the same trust, but most AI system treat them like they do.

s102
00:09:55.080 --> 00:10:04.160
Think of like this, a number from an audit filing is your accountant speaking under oath.

s103
00:10:04.160 --> 00:10:19.680
A number from a analyst note is a friend at a party confident probably from someone's internal email is a thing you overheard in in a elevator.

s104
00:10:19.680 --> 00:10:25.200
Most of retrieval systems can't tell them these apart.

s105
00:10:25.200 --> 00:10:32.880
They grab whichever text is close to your question and hand it over like a gospel.

s106
00:10:32.880 --> 00:10:34.240
True story.

s107
00:10:34.240 --> 00:10:35.080
Anonymized.

s108
00:10:35.080 --> 00:10:48.760
I know I once watched a um very expensive AI tool confidently called a number from a group chat someone's rough guess texted six months earlier.

s109
00:10:48.760 --> 00:10:51.839
The model loved the confident freezing.

s110
00:10:51.839 --> 00:10:58.920
The real audit number was the three rows away in the actual filings.

s111
00:10:58.920 --> 00:11:02.880
The AI just liked the group chat version better.

s112
00:11:02.880 --> 00:11:05.440
It sounded more enthusiastic.

s113
00:11:05.440 --> 00:11:16.880
If your system can can't tell an accountant under oath from a rumor in a group chat, it it is not ready for real money.

s114
00:11:22.360 --> 00:11:23.440
Model two.

s115
00:11:23.440 --> 00:11:30.200
The numbers have to agree each other everywhere, every time.

s116
00:11:30.200 --> 00:11:33.840
Here's a memo that already died.

s117
00:11:33.840 --> 00:12:01.400
Page one says revenue growth 18% 18% page 11 little table nobody reads for fun says 17.4 nobody in the room cares about the missing 0.6. They care about what it what it means if this person didn't check the easy

s118
00:12:01.400 --> 00:12:08.080
mathematics what did they not check on the hard stuff.

s119
00:12:08.080 --> 00:12:16.080
That memo didn't pass not because of the number because of what number you implied.

s120
00:12:16.080 --> 00:12:31.080
And if you want the industrial strength version of this failure remember the giant American real estate company that let algorithm to buy houses at scale.

s121
00:12:31.080 --> 00:12:40.720
The algorithm was extremely confident about the house prices the house disagreed.

s122
00:12:41.080 --> 00:12:52.839
The company ended up writing off around half a billion dollars shut the whole unit down and let a quarter of the staff go.

s123
00:12:52.839 --> 00:13:08.760
The model wasn't stupid the model was unsupervised nobody built a boring mechanic machinery that forces the numbers to keep agreeing with reality after lunch day.

s124
00:13:09.560 --> 00:13:15.200
Fluent and a confident remember is not the same as right.

s125
00:13:19.440 --> 00:13:23.800
Model three surprises people.

s126
00:13:24.160 --> 00:13:28.440
A contradiction is not a bug a contradiction is a gift.

s127
00:13:28.440 --> 00:13:42.480
If the CEO says one gross number on the earning call and the official filing says the different a different one, the gap is the single most interesting thing in the real

s128
00:13:42.480 --> 00:13:44.960
story highs.

s129
00:13:44.960 --> 00:13:48.760
Real diligence leaves for that gap.

s130
00:13:48.760 --> 00:13:53.440
AI does the opposite.

s131
00:13:53.440 --> 00:13:57.720
It is a trained to sound smooth and helpful.

s132
00:13:57.720 --> 00:14:07.720
So, when it hears a conflict, it quietly picks whichever version reads nicer and moves on.

s133
00:14:07.720 --> 00:14:12.000
You never even learn there was a disagreement.

s134
00:14:12.000 --> 00:14:23.040
I've sit through exactly these CEO's number and the filing numbers meaningfully different.

s135
00:14:23.040 --> 00:14:27.280
Neither one uh Nobody Nobody flagged it.

s136
00:14:27.280 --> 00:14:31.080
Everyone just used the nicer one.

s137
00:14:31.080 --> 00:14:37.880
We caught it because one person happened to have both documents open at once.

s138
00:14:37.880 --> 00:14:39.080
Pure luck.

s139
00:14:39.080 --> 00:14:42.080
Luck is not a control.

s140
00:14:42.080 --> 00:14:47.000
Your job as a builder isn't resolve the argument.

s141
00:14:47.000 --> 00:14:57.000
It's to make sure that the argument happens in front of a human instead of quietly along inside box.

s142
00:14:59.440 --> 00:15:01.080
Model four.

s143
00:15:01.080 --> 00:15:07.560
Facts and the guesses have to live in separate box.

s144
00:15:08.080 --> 00:15:12.839
And the fluent AI loves melting them into one smooth sentence.

s145
00:15:12.839 --> 00:15:18.960
Example, the company will likely receive approval next quarter.

s146
00:15:18.960 --> 00:15:21.360
Reads like a fact.

s147
00:15:21.360 --> 00:15:23.160
Sounds like a fact.

s148
00:15:23.160 --> 00:15:24.520
it is a guess.

s149
00:15:24.520 --> 00:15:29.920
Somebody buys the estimator wearing fact shaped the clothing.

s150
00:15:29.920 --> 00:15:36.440
A committee's entire job is to agree with the guesses while trusting the facts.

s151
00:15:36.440 --> 00:15:52.160
If your system melts them together, the committee can't find them seems, and then all all they can do is to is is approve or reject the genuine map of the document.

s152
00:15:52.160 --> 00:15:56.360
You should not spend a hundred million dollars on maps.

s153
00:15:56.360 --> 00:16:00.200
I watched approval accept soon.

s154
00:16:01.760 --> 00:16:06.680
Turn across three drafts of a demo into approval received.

s155
00:16:06.680 --> 00:16:07.960
Nobody lied.

s156
00:16:07.960 --> 00:16:21.800
The guess just rose its fact costume a little longer each rewrite until nobody remembered it then it started as a guess.

s157
00:16:21.800 --> 00:16:25.800
The approval did didn't arrive on schedule.

s158
00:16:25.800 --> 00:16:28.520
That was an uncomfortable phone call.

s159
00:16:28.520 --> 00:16:31.440
This fix almost embarrassingly cheap.

s160
00:16:31.440 --> 00:16:44.080
Label your guesses, a tag, a color, anything that survives being copy pasted into someone else's slides three weeks later.

s161
00:16:47.440 --> 00:16:58.360
Model five, if nobody can find where a claim come from it doesn't matter how right it is.

s162
00:16:58.800 --> 00:17:03.400
You've all heard about the New York's lawyer.

s163
00:17:03.400 --> 00:17:08.920
Uh he filed a legal brief written with a chat bot's help.

s164
00:17:08.920 --> 00:17:19.439
The brief cited six court cases, beautiful citations proper proper formatting, very convincing.

s165
00:17:19.439 --> 00:17:22.720
One small issue, the cases didn't exist.

s166
00:17:22.720 --> 00:17:27.680
The AI invented invented all six.

s167
00:17:27.680 --> 00:17:33.480
And here, my favorite detail, the part that should be taught in school.

s168
00:17:33.480 --> 00:17:41.360
Before filing, the lawyer got suspicious, so he asked the chatbot, "Are these cases real?"

s169
00:17:41.360 --> 00:17:44.040
And the chatbot said, "Yes."

s170
00:17:44.040 --> 00:17:51.840
That is like asking the guy who sold you the watch whether the watch is real.

s171
00:17:51.840 --> 00:17:54.680
The judge fined him.

s172
00:17:54.680 --> 00:18:07.240
The story went around the world, and my second favorite detail, the fake cases even had a real real estate surrounding names and the page numbers.

s173
00:18:07.240 --> 00:18:11.520
The AI didn't just lie, it just the formatting.

s174
00:18:11.520 --> 00:18:16.040
It cited itself beautifully, wrong but beautifully.

s175
00:18:16.040 --> 00:18:19.880
The license is a 30-second test.

s176
00:18:19.880 --> 00:18:27.520
When someone points at a sentence and says, "Show me where this come from."

s177
00:18:27.520 --> 00:18:40.280
You either click once or land on exact source paragraph, or you open seven browser tabs and start swiping.

s178
00:18:40.280 --> 00:18:49.240
I have personally been the guy with seven tabs in a real meeting while the room full of people watched my scroll.

s179
00:18:49.240 --> 00:18:51.920
10 out of 10.

s180
00:18:51.920 --> 00:18:54.640
Wouldn't would not recommend.

s181
00:18:54.640 --> 00:18:59.560
If you remember only one sentence from this whole talk.

s182
00:18:59.560 --> 00:19:02.440
The click-through is a product.

s183
00:19:02.440 --> 00:19:07.160
Everything else is well-written packaging.

s184
00:19:11.560 --> 00:19:19.680
Model six, my favorite, because it's the most human someone has to sign.

s185
00:19:21.240 --> 00:19:24.240
Here's the story you probably know.

s186
00:19:24.240 --> 00:19:39.640
An airline website chatbot told a grieving customer he could book a full-price ticket now and claim a bereavement discount afterwards, that the policy didn't exist.

s187
00:19:39.640 --> 00:19:46.520
The chatbot made it up politely, fluently, confidently.

s188
00:19:46.520 --> 00:19:52.400
The customer took the airlines to a tribunal.

s189
00:19:52.760 --> 00:20:07.080
And the airlines' defense, this is real, was was that the chatbot is called a separate legal entity responsible for its own action.

s190
00:20:07.080 --> 00:20:12.360
That is the corporate version of my dog ate my homework.

s191
00:20:12.520 --> 00:20:14.960
The tribunal didn't buy it.

s192
00:20:14.960 --> 00:20:22.200
The airline paid, and every one of us in our boardroom quietly took a note that day.

s193
00:20:22.200 --> 00:20:27.280
You can't outsource accountability to your own software.

s194
00:20:27.280 --> 00:20:32.760
At the bottom of every real decision, a human signs.

s195
00:20:32.760 --> 00:20:43.360
If your architecture doesn't have a fundable human at the end of it, you you have not built a product.

s196
00:20:43.360 --> 00:20:47.360
You have have built an excuse generator.

s197
00:20:47.360 --> 00:20:54.280
So, build your AI around that accountable person, not instead of them.

s198
00:20:56.360 --> 00:20:58.800
Okay, the fix.

s199
00:20:58.800 --> 00:21:00.400
Five things.

s200
00:21:00.400 --> 00:21:06.320
Each one is a direct cure for a story you have just heard.

s201
00:21:06.320 --> 00:21:10.320
This is almost a word for word.

s202
00:21:10.320 --> 00:21:16.840
What I demand from our vendor before lighting their system near a live deal.

s203
00:21:16.840 --> 00:21:21.960
One, every claim comes with a receipt.

s204
00:21:21.960 --> 00:21:31.240
Each sentence linked us straight to its source paragraph with the source trust level attached.

s205
00:21:31.240 --> 00:21:34.320
Not a citation tab at the end.

s206
00:21:34.320 --> 00:21:41.080
Two, fact and the guesses stay visibility separate.

s207
00:21:41.080 --> 00:21:50.040
At glance at the page, I insistently see what's proven and what's somebody's best estimate.

s208
00:21:50.040 --> 00:21:55.120
Three, numbers agree with each other automatically.

s209
00:21:55.120 --> 00:22:00.840
The system refuse to ship a memo where the figures don't match.

s210
00:22:00.840 --> 00:22:04.320
No human checking at 2:00 in the morning.

s211
00:22:04.320 --> 00:22:09.840
Four, contradictions get surfaced, never smoothed over.

s212
00:22:09.840 --> 00:22:19.160
When sources disagree, the system raises its hand instead of picking the front layer answer.

s213
00:22:19.160 --> 00:22:24.600
Five, a real human approval gate and is logged.

s214
00:22:24.600 --> 00:22:29.280
Who's Who reviewed what change when they signed.

s215
00:22:29.280 --> 00:22:32.000
That log is audit trail.

s216
00:22:32.000 --> 00:22:34.640
Notice what's not on the list.

s217
00:22:34.640 --> 00:22:44.360
A smart promoter not One of these is a bigger brain problems.

s218
00:22:44.360 --> 00:22:48.600
All five plumbing and honesty problems.

s219
00:22:48.600 --> 00:22:55.480
The winners in the in this category won't win on benchmark points.

s220
00:22:55.480 --> 00:23:10.800
They will win because a tired, skeptical finance person finance person can trust that their output at 11:00 at night without opening seven tabs.

s221
00:23:13.640 --> 00:23:17.040
Now, the part I promised, the money.

s222
00:23:17.040 --> 00:23:21.560
Many of you are not just the building AI products.

s223
00:23:21.560 --> 00:23:26.080
You are raising for them or you will be.

s224
00:23:26.080 --> 00:23:33.120
So, let me tell you what actually happens after you leave the pitch meeting.

s225
00:23:33.120 --> 00:23:35.440
Your deck becomes a memo.

s226
00:23:35.440 --> 00:23:43.880
Literally, literally, literally, someone like me sits down and writes an internal memo about you.

s227
00:23:43.880 --> 00:23:59.000
Every number you said out loud gets checked get get checked against your data room, what which means everything I just told you about the documents apply to you personally.

s228
00:23:59.000 --> 00:24:00.680
Four lessons.

s229
00:24:00.680 --> 00:24:01.800
Okay, my time's up.

s230
00:24:01.800 --> 00:24:04.080
Thank you.

s231
00:24:04.291 --> 00:24:06.291
[applause]