WEBVTT

NOTE Sentence-level transcript of https://www.youtube.com/watch?v=CLve4Nx_RwI

NOTE One cue per sentence. Cue ids are the line anchors on /transcripts/CLve4Nx_RwI.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
00:00:12.360 --> 00:00:13.640
Good morning everyone.

s3
00:00:13.640 --> 00:00:20.840
I'm Sam Green, team lead for applied AI at Bridgewater Associates, and this is my colleague Erin Lensky.

s4
00:00:20.840 --> 00:00:31.600
And today we're here to talk to you about the Vault, the data platform that we've built at Bridgewater to turn unstructured data into a first-class asset in our investment process.

s5
00:00:31.600 --> 00:00:37.200
And as part of the talk this morning, we'd like to share three lessons that we've learned building the Vault.

s6
00:00:37.200 --> 00:00:43.120
The first is that humans and agents have very different needs from a data platform.

s7
00:00:43.120 --> 00:00:46.920
And so you have to build for those different needs separately.

s8
00:00:46.920 --> 00:00:53.560
The second is that the greatest impact occurs when a single tool can meet both of these sets of needs at the same time.

s9
00:00:53.560 --> 00:01:06.080
And the third is that the real prize is when you can build the foundations needed to make agents a proactive participant in your investment process or whatever you're doing with your with your your tools

s10
00:01:06.080 --> 00:01:11.640
instead of a reactive reactive tool for your humans to use.

s11
00:01:11.640 --> 00:01:17.280
Before I I dive deeper into talking about the Vault, I wanted to start with some some background about Bridgewater.

s12
00:01:17.280 --> 00:01:25.800
Over the last 50 years, Bridgewater has been in the constant pursuit of the world's best understanding of how markets and economies work.

s13
00:01:25.800 --> 00:01:32.200
And turning that into a systematic encoding that we can use to stress test our logic over time.

s14
00:01:32.200 --> 00:01:44.640
And what originally began as a set of if-then rules on a yellow legal pad evolved into a continuous compounding process with humans at the center constantly looking for the next thing that's changing in markets.

s15
00:01:44.640 --> 00:01:53.520
And more recently, we've been working on building AiA, the artificial investment associate, which puts a machine at the center of this continuous process.

s16
00:01:53.520 --> 00:02:03.520
Across all these phases of evolution, we've had a constant philosophy in mind, which is that we want to be constantly looking at how the world is changing and what's going on in the world.

s17
00:02:03.520 --> 00:02:14.360
Noticing what what dynamics are not yet encoded in our understanding and use that learning to improve and compound for the future.

s18
00:02:14.840 --> 00:02:21.400
For most of Bridgewater's history, that meant building and overseeing time series models on structured data.

s19
00:02:21.400 --> 00:02:35.720
And that structured data needed to be carefully built so that it could be safe for use in our research process and aligned to our internal ontology and world model so that we could get great leverage out of using the the data that we had built.

s20
00:02:35.720 --> 00:02:42.600
With the emergence of language models, we had a huge new opportunity to engage with a with a much wider world of data.

s21
00:02:42.600 --> 00:02:51.160
Um, you know, news and research, transcripts, policy documents, and all of the internal conversation that was previously just in the ether.

s22
00:02:51.160 --> 00:02:56.280
So, the internal memos and meeting transcripts and debate happening inside the firm.

s23
00:02:56.280 --> 00:03:01.800
We quickly realized that all of that data needed the exact same treatment to be fit for purpose.

s24
00:03:01.800 --> 00:03:13.320
It needed to be carefully cleaned, governed, and made ready for use in both point-in-time research questions today and also our systematic study over time.

s25
00:03:13.320 --> 00:03:25.880
And when we looked out at the landscape of data vendors and data platform providers, there was nothing that provided us with all of the control that we needed to build and maintain our data advantage.

s26
00:03:25.920 --> 00:03:33.880
And so, we built the Vault, which is our repository of external and internal unstructured data.

s27
00:03:33.880 --> 00:03:41.360
Um, and the platform that indexes, organizes, and makes all of it ready for use by our human investors and agents.

s28
00:03:41.360 --> 00:03:52.080
Today, the Vault has more than 100 million documents in it with thousands more added every day, uh, minute by minute, from 135 external sources and again growing all the time,

s29
00:03:52.080 --> 00:03:55.360
as well as all of the data from inside of Bridgewater.

s30
00:03:55.360 --> 00:03:59.760
And crucially, all that data is not just dumped on our investors' desks.

s31
00:03:59.760 --> 00:04:13.840
It is carefully aligned to our internal internal ontology with enrichments like the countries, concepts, markets, and most importantly, time period discussed in the document.

s32
00:04:14.080 --> 00:04:17.200
To bring this to life, I want to start with a quick motivating example.

s33
00:04:17.200 --> 00:04:25.919
So, what I did here is I asked a fairly basic representative investment question to an enterprise chat client.

s34
00:04:25.919 --> 00:04:33.320
I asked, "How is AI infrastructure investment impacting the Korean won, and do market participants expect this to continue?"

s35
00:04:33.320 --> 00:04:44.200
And as as you're all I'm sure familiar with, the agent goes to the web and looks for resources it can draw on because I've explicitly asked it to go find

s36
00:04:44.320 --> 00:04:46.760
other people's perspectives on the question.

s37
00:04:46.760 --> 00:04:56.280
And you can see it it zooms in and it finds 15 relevant-looking resources to draw on before it gives me back its answer.

s38
00:04:56.280 --> 00:04:59.919
And when I when I zoomed in on the answer it gave back, a couple of things stood out to me.

s39
00:04:59.919 --> 00:05:08.440
The first is that even though it found 15 relevant-looking sources, it only picked one or two to answer to answer my question.

s40
00:05:08.440 --> 00:05:15.400
And the second is down at the end of this response where the agent itself acknowledges that it hasn't answered my question.

s41
00:05:15.400 --> 00:05:24.320
It didn't find enough documents on the open web to establish a market consensus and meet the meet the parameters that I laid out.

s42
00:05:24.320 --> 00:05:28.000
When I double-clicked even further, the problem got even worse.

s43
00:05:28.000 --> 00:05:41.880
One of the notes that it primarily drew on to answer my question was published way back in July of 2026, even though I asked this question and captured this answer just at the end of last week in early October.

s44
00:05:41.880 --> 00:05:47.320
In the age of AI and in investing, four three months is a massive amount of time.

s45
00:05:47.320 --> 00:05:53.840
It's not even close to good enough to draw uh draw on this resource to answer my question.

s46
00:05:54.000 --> 00:05:59.840
When I ask this exact same query attached to the vault, something totally different happens.

s47
00:05:59.840 --> 00:06:06.520
The first is that the connection to the vault helps to prompt the agent that it should look for a specific period of time when answering my question.

s48
00:06:06.520 --> 00:06:09.760
So, it zooms in on October 1st to October 8th.

s49
00:06:09.760 --> 00:06:20.760
It finds a much wider array of resources, 52 candidate documents that are relevant to the query over this specific period of time because of the vault's much more significant

s50
00:06:20.760 --> 00:06:24.400
corpus of content that's relevant to investing questions.

s51
00:06:24.400 --> 00:06:30.680
And when it ultimately answers my question, it draws on seven instead of only one or two.

s52
00:06:30.680 --> 00:06:37.200
And gives me back a rich answer that the agent is more confident is going to address my question.

s53
00:06:37.200 --> 00:06:45.160
And so, for us, this is the the big lesson of from building the vault to kind of motivate the talk today, which is that it takes great data and the right connectors

s54
00:06:45.160 --> 00:06:47.520
to make agents impactful.

s55
00:06:47.520 --> 00:06:53.800
And I'm going to talk hand it over to Aaron now to talk about some of the product engineering takeaways from building the vault.

s56
00:06:54.080 --> 00:06:55.760
Thanks, Sam.

s57
00:06:55.760 --> 00:06:56.120
So, hi.

s58
00:06:56.120 --> 00:06:56.640
I'm Aaron.

s59
00:06:56.640 --> 00:06:58.520
I run product engineering at Bridgewater.

s60
00:06:58.520 --> 00:07:02.880
I'm going to take us through a few of our learnings from building the vault over the past few years.

s61
00:07:02.880 --> 00:07:10.760
I'm going to start with one that might seem obvious, but I still think it's worth saying, which is search is hard.

s62
00:07:11.560 --> 00:07:15.520
But, it gets even harder if we're designing for two different consumers.

s63
00:07:15.520 --> 00:07:18.880
And in this case, we're designing for humans and agents.

s64
00:07:18.880 --> 00:07:21.320
I'm going to go over the difference here starting with us.

s65
00:07:21.320 --> 00:07:27.160
So, humans need precision and a high signal-to-noise ratio because we can only read so much.

s66
00:07:27.160 --> 00:07:33.440
Even a fast reader manages about 300 words a minute, and personally, I'm much slower.

s67
00:07:33.440 --> 00:07:36.120
Agents are a different kind of reader.

s68
00:07:36.120 --> 00:07:38.480
They're far less sensitive to precision.

s69
00:07:38.480 --> 00:07:40.720
If the data is there, they'll find it.

s70
00:07:40.720 --> 00:07:44.920
And basically, anything you hand to them will get attended to.

s71
00:07:44.920 --> 00:07:53.320
They scale super quickly and they come and they consume about uh a hundred times faster, 30,000 words per minute.

s72
00:07:53.320 --> 00:07:56.920
And so we need to build Vault for both of these.

s73
00:07:56.920 --> 00:08:02.320
And when we're thinking about the dimensions that matter, there are these five on on the board.

s74
00:08:02.320 --> 00:08:07.360
Recall, precision, filtering, entitlements, and scale.

s75
00:08:07.360 --> 00:08:12.000
Some of them matter to people, some matter more to agents, many to both.

s76
00:08:12.000 --> 00:08:13.760
I'm going to go deep on three of these.

s77
00:08:13.760 --> 00:08:19.280
Precision for people, scale for agents, and then filtering for both.

s78
00:08:19.280 --> 00:08:21.800
So, precision.

s79
00:08:21.840 --> 00:08:30.400
Once we've filled the Vault, our which was a heroic effort by our data and legal teams, recall was no longer a pressing issue.

s80
00:08:30.400 --> 00:08:35.080
However, the number one complaint was Vault search is bad.

s81
00:08:35.080 --> 00:08:40.200
As a product owner, it's painful to hear that, but my response was let's benchmark it.

s82
00:08:40.200 --> 00:08:42.120
And that's what we did.

s83
00:08:42.120 --> 00:08:48.640
We collected questions and found the best documents by hand, including ones that looked helpful but weren't.

s84
00:08:48.640 --> 00:08:52.400
In all In total it was about 200 examples and it took us less than a week.

s85
00:08:52.400 --> 00:08:57.360
It was not fun or very glamorous, but it was very helpful.

s86
00:08:57.360 --> 00:08:59.600
And after we ran it, it was very clear.

s87
00:08:59.600 --> 00:09:01.440
Search was bad.

s88
00:09:01.440 --> 00:09:09.760
The key metric we were looking at is precision at 10, which is of the top 10 results, how many are the right documents.

s89
00:09:09.760 --> 00:09:12.520
And we were hitting about 60%.

s90
00:09:12.520 --> 00:09:15.480
So, the people were right, it was bad.

s91
00:09:15.480 --> 00:09:18.720
But now we had a way to measure all of the improvements.

s92
00:09:18.720 --> 00:09:21.120
And here on the board are five of them.

s93
00:09:21.120 --> 00:09:30.680
And by the time we got done with Agentic Search, we were up in the 90-95% uh recall at 10 or precision at 10.

s94
00:09:31.120 --> 00:09:32.960
All right, so that was precision.

s95
00:09:32.960 --> 00:09:35.200
Now on to filtering.

s96
00:09:35.200 --> 00:09:38.360
Like anyone, we want the most relevant documents.

s97
00:09:38.360 --> 00:09:50.600
But when we looked at the production traces, what we were seeing is that almost always there was some condition, either as of a certain date or from a certain source or about a specific asset.

s98
00:09:50.600 --> 00:09:58.440
And all of that requires advanced filtering, and that's where we ran into our first major performance issue.

s99
00:09:58.440 --> 00:10:07.200
So I'm going to quickly go through this of a little bit technical here of there's classically two ways to fill filter a vector database.

s100
00:10:07.200 --> 00:10:12.640
Post filtering will use that vector index first and then filter the top results by metadata.

s101
00:10:12.640 --> 00:10:20.200
It's fast, but recall could be pretty low because the globally relevant results might not match your filters.

s102
00:10:20.200 --> 00:10:27.520
On the other side, we have pre-filtering where we'll filter first and then calculate distances for every one of those matches.

s103
00:10:27.520 --> 00:10:33.840
The advantage there is that the recall is very good, but it can be very slow, and we needed something in between.

s104
00:10:33.840 --> 00:10:37.400
So that's what led us to try and then ultimately switch to Turbo Puffer.

s105
00:10:37.400 --> 00:10:44.280
Their approach, which they called native filtering, looks only in the nearest clusters that contain matches and then scores just those.

s106
00:10:44.280 --> 00:10:52.520
What you end up getting is something that's fast like post filtering, but with close to the same recall as pre-filtering.

s107
00:10:53.120 --> 00:10:56.520
And lastly, I'm going to talk about scale.

s108
00:10:56.600 --> 00:11:00.160
Human scale for us is pretty pedestrian, 100 queries an hour.

s109
00:11:00.160 --> 00:11:01.720
We have 300 users.

s110
00:11:01.720 --> 00:11:04.080
They're going to search, you know, every so often.

s111
00:11:04.080 --> 00:11:08.960
It takes our users minutes to read the docs that they're that they're receiving.

s112
00:11:08.960 --> 00:11:13.080
Agents, on the other hand, are a few orders of magnitude beyond that.

s113
00:11:13.080 --> 00:11:18.160
Easily 100 plus queries a second, and agent swarms can hit thousands easily.

s114
00:11:18.160 --> 00:11:22.000
And they obviously, will read, as I said earlier, will read docs in seconds.

s115
00:11:22.000 --> 00:11:28.600
So we needed something that could scale as well, and again, Turbo Puffer handled that easily, too.

s116
00:11:28.600 --> 00:11:30.880
So we were covering all five dimensions.

s117
00:11:30.880 --> 00:11:34.400
I didn't talk about entitlements, but I'll be talking about that later today.

s118
00:11:34.400 --> 00:11:35.720
Um but we covered it all.

s119
00:11:35.720 --> 00:11:38.080
It unlocked new workflows.

s120
00:11:38.080 --> 00:11:41.520
But it didn't change how our researchers worked.

s121
00:11:41.520 --> 00:11:44.320
They still weren't coming to Vault every morning.

s122
00:11:44.320 --> 00:11:53.040
And that's my second lesson that I want to share here is that no matter how amazing the tools are, we can't just drop them on our investors' desks and expect them to be used.

s123
00:11:53.040 --> 00:11:58.640
We need to migrate their old habits into an a new form.

s124
00:11:58.640 --> 00:12:02.640
And so we looked at what they were already doing.

s125
00:12:02.640 --> 00:12:04.120
Many of them were living in email.

s126
00:12:04.120 --> 00:12:06.440
The research came straight from the brokers.

s127
00:12:06.440 --> 00:12:08.360
They had subscribed lists.

s128
00:12:08.360 --> 00:12:11.040
They had rules that filtered everything into folders.

s129
00:12:11.040 --> 00:12:11.880
It worked.

s130
00:12:11.880 --> 00:12:13.360
It was familiar.

s131
00:12:13.360 --> 00:12:16.000
And they knew that they weren't missing anything.

s132
00:12:16.000 --> 00:12:19.400
So we had to build all of these features into the Vault.

s133
00:12:19.400 --> 00:12:20.680
That's exactly what we did.

s134
00:12:20.680 --> 00:12:21.920
Everything's searchable.

s135
00:12:21.920 --> 00:12:25.120
The subscribe lists became saved searches.

s136
00:12:25.120 --> 00:12:27.400
And folders became saved collections.

s137
00:12:27.400 --> 00:12:30.040
It's the same habits, just a new home.

s138
00:12:30.040 --> 00:12:33.240
And once we had them there, we can open up new habits.

s139
00:12:33.240 --> 00:12:37.360
So we can ask they can ask questions across all the data as Sam showed.

s140
00:12:37.360 --> 00:12:39.680
They can follow dynamic feeds by topic.

s141
00:12:39.680 --> 00:12:43.760
And they can hand those collections over to their agents to take further action.

s142
00:12:43.760 --> 00:12:47.600
And that's when we clearly had product market fit.

s143
00:12:47.839 --> 00:12:50.760
And once we had that fit, the users wanted it everywhere.

s144
00:12:50.760 --> 00:12:52.200
So we put it in chat.

s145
00:12:52.200 --> 00:12:53.240
We made a phone app.

s146
00:12:53.240 --> 00:12:54.560
It's in Slack in their IDE.

s147
00:12:54.560 --> 00:12:57.640
And of course, we availed all of that to the agents as well.

s148
00:12:57.640 --> 00:13:02.720
Um including a a CLI because agents love CLIs.

s149
00:13:02.760 --> 00:13:06.800
And so the main question here is okay, so what does this enable?

s150
00:13:06.800 --> 00:13:09.320
Sam's going to talk about what we're building right now.

s151
00:13:09.320 --> 00:13:14.800
But first and foremost, it's helping us to move around that research circle faster.

s152
00:13:14.800 --> 00:13:24.120
With our pocket analyst, which is our first AI analyst, it can do it can do hours it can do hours of research in just a few minutes for a few dollars.

s153
00:13:24.120 --> 00:13:26.520
And that's relying on the Vault.

s154
00:13:26.520 --> 00:13:30.120
And then we're expanding now with the researcher, which is the next generation.

s155
00:13:30.120 --> 00:13:34.720
It's the same idea stretched across days or weeks.

s156
00:13:34.920 --> 00:13:45.480
We're also able to train custom models to more intelligently reason over the vault's data and use those together with other foundation models to feed systematic decision-making processes.

s157
00:13:45.480 --> 00:13:51.640
The forecaster is our first one, which produces AI-generated macro-relevant predictions.

s158
00:13:51.640 --> 00:14:00.280
And then, we also have a built investment signals from the vault, turning unstructured data into structured, tradable strategies.

s159
00:14:00.280 --> 00:14:05.160
Now, I'll hand it to Sam to take us through Vault's latest capability.

s160
00:14:05.360 --> 00:14:07.360
Thanks.

s161
00:14:07.560 --> 00:14:18.360
So, like I alluded to at the beginning of the talk, the real prize is when you can take agents from a reactive resource that can be ready to answer a a human's question

s162
00:14:18.360 --> 00:14:27.160
and instead make them proactive participants helping to steer your high-value professionals towards the most important areas for their attention.

s163
00:14:27.160 --> 00:14:42.400
And the way that we're solving that problem is in an experience we're calling the investor workbench, where the agent is always on tracking what's important to that investor and presenting them with the areas that are most in need of their attention.

s164
00:14:42.400 --> 00:14:55.640
It looks like an experience like this that relies on the vault, where a topic is proposed and several pieces of information are presented to the investor to decide whether they should dig deeper or ignore that topic entirely.

s165
00:14:55.640 --> 00:15:05.360
You can see here the card that an investor sees in order to decide whether to dig in deeper into the same AI impact on currencies question that we looked at earlier in the talk.

s166
00:15:05.360 --> 00:15:12.960
And the investor receives everything they need, an agentic summary if they want a quick scan across the agent's findings.

s167
00:15:12.960 --> 00:15:21.280
They also get the exact evidence that the agent was drawing on, so they can assess for themselves what was leading to this recommendation.

s168
00:15:21.280 --> 00:15:29.160
And if they choose to dig deeper, they have a direct connection to the research output from the analytical tools that Aaron was just describing.

s169
00:15:29.160 --> 00:15:44.160
And when they dig deeper alongside the agent to look at the dynamics driving this question, the learnings are automatically captured, the results of their collaboration are automatically captured back in the vault.

s170
00:15:44.320 --> 00:15:45.960
And so this for us is the key lesson.

s171
00:15:45.960 --> 00:16:02.760
The vault enables compounding because it provides the right evidence, the right extra internal and external information at the fingertips of our investors and agents, and it automatically captures the synthesis that they generate and puts it back in the corpus for the next turn.

s172
00:16:02.760 --> 00:16:06.520
And that makes us go around the circle faster and better.

s173
00:16:06.520 --> 00:16:16.240
And so when we think back to our the the question we began with, what it takes to make an agent work is the right context and the right understanding of what's going on in your world.

s174
00:16:16.240 --> 00:16:19.720
And so our question for you is, does your AI know what's going on in the world?

s175
00:16:19.720 --> 00:16:22.360
And does it remember what your firm has learned about it?

s176
00:16:22.360 --> 00:16:24.560
Because we are teaching ours too.

s177
00:16:24.560 --> 00:16:28.040
And if these sounds like an sound like an exciting challenge, we'd love to talk.

s178
00:16:28.040 --> 00:16:30.200
Thank you.

s179
00:16:30.371 --> 00:16:32.371
[applause]