WEBVTT

NOTE Sentence-level transcript of https://www.youtube.com/watch?v=B0fjR3yaZFU

NOTE One cue per sentence. Cue ids are the line anchors on /transcripts/B0fjR3yaZFU.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
00:00:12.480 --> 00:00:13.280
Hi everyone.

s3
00:00:13.280 --> 00:00:14.240
I'm Varun.

s4
00:00:14.240 --> 00:00:21.920
I'm one of the co-founders at Long Lake and I'm excited to share a little bit about what we've been up to for the last 2 years.

s5
00:00:21.920 --> 00:00:27.280
It all comes back to a question all of us have asked time and time again.

s6
00:00:27.280 --> 00:00:35.360
The models are getting better, but the real question is how do you actually deploy the AI into the real world?

s7
00:00:35.360 --> 00:00:41.120
How do you get the models to complete economically relevant tasks?

s8
00:00:43.120 --> 00:00:46.720
Let me start by saying everyone has seen the demo.

s9
00:00:46.720 --> 00:00:55.720
Think of the agent automatically booking a flight, the agent automatically completing a ticket in some kind of customer service portal.

s10
00:00:55.720 --> 00:01:00.560
Think of an agent completing a block of code ready to commit and go.

s11
00:01:00.560 --> 00:01:05.400
The reality is we've all seen this and it feels like magic.

s12
00:01:05.400 --> 00:01:09.320
2 years ago any of this would have been complete science fiction.

s13
00:01:09.320 --> 00:01:12.400
The capabilities are real.

s14
00:01:12.400 --> 00:01:16.920
Now, walk with me into a 200-person property management firm.

s15
00:01:16.920 --> 00:01:24.680
Real people, real properties, real dollars, real customers all across the US.

s16
00:01:24.760 --> 00:01:32.800
You would expect AI to show up by now, but the reality is nothing has changed at all.

s17
00:01:34.600 --> 00:01:36.120
Here's the thing.

s18
00:01:36.120 --> 00:01:42.720
This is totally normal and maybe in fact I'd argue this is what we should expect.

s19
00:01:42.960 --> 00:01:45.960
This is true for every general-purpose technology.

s20
00:01:45.960 --> 00:01:48.560
You know, take electricity for example.

s21
00:01:48.560 --> 00:01:57.520
Electricity was invented in the 1880s and it was first demoed at Edison's Pearl Street Station Dynamo Room over in Manhattan.

s22
00:01:57.520 --> 00:02:01.040
This was the magic demo of its time.

s23
00:02:01.560 --> 00:02:07.040
The reality is it took a long time for electricity to be fully adopted.

s24
00:02:07.040 --> 00:02:09.600
Consider a Ford factory.

s25
00:02:09.600 --> 00:02:12.400
It's not enough to just have electricity.

s26
00:02:12.400 --> 00:02:15.680
You have to rip out the existing motors and equipment.

s27
00:02:15.680 --> 00:02:16.959
You have to bring in the new equipment.

s28
00:02:16.959 --> 00:02:21.120
You have to go and train everybody to use that very same equipment.

s29
00:02:21.120 --> 00:02:26.200
Here's a picture of a Ford electrified moving assembly in 1924.

s30
00:02:26.200 --> 00:02:28.720
These things take time.

s31
00:02:28.959 --> 00:02:33.200
Diffusion of any technology takes a generation.

s32
00:02:33.200 --> 00:02:44.240
And since everyone here in this room today is talking about AI, I would argue AI diffusion is perhaps the single most important problem for the next 20 years.

s33
00:02:44.240 --> 00:02:46.120
The models are going to keep getting better.

s34
00:02:46.120 --> 00:02:57.160
The big question is how do we actually get these models to be in the real world, complete real tasks, uh and make people more efficient, happier, and provide better service.

s35
00:02:57.400 --> 00:03:00.080
So taking a quick step step back, who are we?

s36
00:03:00.080 --> 00:03:01.600
Uh we are Long Lake.

s37
00:03:01.600 --> 00:03:10.239
Over the last 2 years, we've raised over $3 billion from Elad Gil, General Catalyst, and AlphaWave since our founding.

s38
00:03:10.239 --> 00:03:11.480
Here's the strange part.

s39
00:03:11.480 --> 00:03:13.200
We we don't sell software.

s40
00:03:13.200 --> 00:03:18.440
We actually go out and acquire and partner with real services businesses in the world.

s41
00:03:18.440 --> 00:03:26.320
Uh we've acquired 35 businesses across HOA and property management, architecture, HR services, and a lot more.

s42
00:03:26.600 --> 00:03:33.800
To give you a little bit more flavor, we have roughly a 40% team right now split between technology, finance, and operations.

s43
00:03:33.800 --> 00:03:42.360
More than half our team is part of the technology team focused on uh building products, data, and deploying the core products into the field.

s44
00:03:42.360 --> 00:03:54.680
Uh we're in a collected group of folks, a bunch of ex-founders who've worked in the services before, ex-military, folks from Palantir, Ramp, Glean, uh and from the finance side, Blackstone, H.I.G.,

s45
00:03:54.680 --> 00:03:55.880
et cetera.

s46
00:03:55.880 --> 00:04:00.480
We are we are not selling them to these companies above from the outside.

s47
00:04:00.480 --> 00:04:05.400
We're actually deploying into these companies and figuring out how to get the technology to work.

s48
00:04:05.400 --> 00:04:16.400
And just to show you the scale we're playing at, we announced recently our $6.3 billion take private of American Express Global Business Travel, the world's largest corporate travel platform.

s49
00:04:16.959 --> 00:04:19.480
We own these businesses.

s50
00:04:19.480 --> 00:04:23.000
So, when the AI doesn't work, it's not their problem.

s51
00:04:23.000 --> 00:04:24.400
We're not the vendor.

s52
00:04:24.400 --> 00:04:26.800
It's our problem.

s53
00:04:27.080 --> 00:04:29.840
Concretely, again, we are not the vendor.

s54
00:04:29.840 --> 00:04:31.440
We are the operator owners.

s55
00:04:31.440 --> 00:04:37.240
And we work very closely with our teams within the businesses to drive real outcomes.

s56
00:04:37.600 --> 00:04:41.040
Now, I want to step back and get to the concrete about the how.

s57
00:04:41.040 --> 00:04:44.280
What are the lessons we've learned over the last 2 and 1/2 years?

s58
00:04:44.280 --> 00:04:49.520
And what we've learned from deploying AI into companies we've owned.

s59
00:04:51.600 --> 00:04:53.000
Three quick lessons.

s60
00:04:53.000 --> 00:04:57.360
One, how we move agents from co-pilots to co-workers.

s61
00:04:57.360 --> 00:05:02.120
Two, how we leverage real-world data within these businesses.

s62
00:05:02.120 --> 00:05:07.360
Remember, we're seeing all of the work that's being done in these real services businesses.

s63
00:05:07.360 --> 00:05:11.880
There's a lot of interesting problems and solutions embedded within that.

s64
00:05:11.880 --> 00:05:19.919
And then finally, perhaps the most interesting and exciting is how do you actually get all of this technology to compound over time by learning loops in the enterprise.

s65
00:05:19.919 --> 00:05:23.680
We'll get to that at the end over here.

s66
00:05:23.680 --> 00:05:27.000
So, starting off from co-pilots to co-workers.

s67
00:05:27.000 --> 00:05:31.000
There's a spectrum of how much autonomy you can give an agent.

s68
00:05:31.000 --> 00:05:32.560
On the left here, you see a co-pilot.

s69
00:05:32.560 --> 00:05:36.080
This is, you know, your simple rag chatbot from 2 years ago.

s70
00:05:36.080 --> 00:05:36.840
It's very quick.

s71
00:05:36.840 --> 00:05:37.800
You can ask a question.

s72
00:05:37.800 --> 00:05:39.680
Maybe it's integrated with some systems.

s73
00:05:39.680 --> 00:05:43.440
It can give you information back very, very quickly.

s74
00:05:44.080 --> 00:05:45.960
The second step is a synchronous agent.

s75
00:05:45.960 --> 00:05:49.760
Consider something like Claude code, Codex, Claude co-work.

s76
00:05:49.760 --> 00:05:50.680
It's real-time.

s77
00:05:50.680 --> 00:05:52.160
There's this two-way interaction.

s78
00:05:52.160 --> 00:05:54.160
It's a bit more sophisticated than a co-pilot.

s79
00:05:54.160 --> 00:05:56.920
You can go let it run off for 1 to 5 minutes.

s80
00:05:56.920 --> 00:05:59.400
Uh it'll call tools, maybe use its skills.

s81
00:05:59.400 --> 00:06:00.320
Uh it's still synchronous.

s82
00:06:00.320 --> 00:06:02.240
You still need to step in and ask a query.

s83
00:06:02.240 --> 00:06:06.600
So, the next obvious rung of the ladder is the asynchronous agent.

s84
00:06:06.600 --> 00:06:08.960
You can come in here, still ask a query.

s85
00:06:08.960 --> 00:06:12.760
The agent will go off into the background, do some work, and then come back.

s86
00:06:12.760 --> 00:06:19.040
Uh and what's really interesting about asynchronous agents is that the user does not have to be the one that triggers them.

s87
00:06:19.040 --> 00:06:21.240
You can have external triggers as well.

s88
00:06:21.240 --> 00:06:31.919
Maybe someone completes a certain task and there is an async job queue uh that allows the async agent to pull off from and proactively offer advice to the end user.

s89
00:06:31.919 --> 00:06:35.520
Then, I'd argue the next step is a long-running agent.

s90
00:06:35.520 --> 00:06:40.760
How do you get these agents to work for hours, days, weeks, months, etc.?

s91
00:06:40.760 --> 00:06:47.480
I think this is currently a very core problem that a lot of the labs are focused on, as are we.

s92
00:06:48.000 --> 00:06:53.080
And then finally, at the end, the holy grail, an AI co-worker.

s93
00:06:53.080 --> 00:06:55.200
This is where most people start off.

s94
00:06:55.200 --> 00:07:00.280
You want a proactive partner that gets work done just alongside you.

s95
00:07:00.280 --> 00:07:09.720
This is what everyone wants to sell you, but what we've learned from owning the outcomes in this business is you have to earn the right to do more.

s96
00:07:09.720 --> 00:07:13.560
It's it's not enough to jump to the co-worker immediately, right?

s97
00:07:13.560 --> 00:07:14.720
For for a bunch of reasons.

s98
00:07:14.720 --> 00:07:18.520
One, for certain tasks, the models might not quite be there yet.

s99
00:07:18.520 --> 00:07:32.040
And two, you actually have to work with these companies in the field, interact and iterate very, very closely, so that they understand that this is the beginning of AI, and you can work up the rungs over time.

s100
00:07:33.240 --> 00:07:38.640
I think a really unique lens to look at this problem through is the that of the jagged frontier.

s101
00:07:38.640 --> 00:07:42.600
We all know that agents are incredibly good at writing code.

s102
00:07:42.600 --> 00:07:47.000
So, what does the, for example, synchronous agent for code generation look like?

s103
00:07:47.000 --> 00:07:48.000
This is super simple.

s104
00:07:48.000 --> 00:07:52.640
This is just your coding agent, maybe it's Codex, Cloud Code, just running on your desktop.

s105
00:07:52.640 --> 00:07:54.200
It has access to a file system.

s106
00:07:54.200 --> 00:07:55.800
You collaborate within real time.

s107
00:07:55.800 --> 00:07:59.160
You get instant feedback and you iterate.

s108
00:07:59.440 --> 00:08:03.800
The next step is, you know, if you look at code code generation, what is the async agent?

s109
00:08:03.800 --> 00:08:06.200
This is also fairly straightforward and largely solved.

s110
00:08:06.200 --> 00:08:11.240
You take the exact same coding agent, you wrap it in a sandbox, and you just let it go run.

s111
00:08:11.240 --> 00:08:16.880
It can build, it can test, and once it's done with its work, it can provide the code in the form of a PR.

s112
00:08:16.880 --> 00:08:23.760
One thing that's really unique about engineers is folks are incredibly good at already paralyzing their work.

s113
00:08:23.760 --> 00:08:30.880
It's very commonplace to launch 10 jobs and be comfortable with the fact that job seven might finish before job three.

s114
00:08:30.880 --> 00:08:35.479
So, engineers are incredibly good at using these async agents.

s115
00:08:35.719 --> 00:08:41.640
Now, when we come to services, the equivalent of a synchronous agent, what we talked about a little bit earlier, it's a co-working agent.

s116
00:08:41.640 --> 00:08:44.480
It's an agent that has deep context about your enterprise.

s117
00:08:44.480 --> 00:08:53.160
It interacts potentially with MCPs, custom tools, custom integrations, uh and you can chat with it synchronously just like any of these other products.

s118
00:08:53.160 --> 00:08:56.160
I think this is a frontier here in the bottom right.

s119
00:08:56.160 --> 00:09:00.520
What does it mean to build an asynchronous agent for the services?

s120
00:09:00.520 --> 00:09:08.120
What does it mean to paralyze work in industries where work is traditionally done in a very, very serial manner?

s121
00:09:08.120 --> 00:09:17.560
This is where we spend a lot of time and this is what I wake up every morning really excited thinking about, you know, we've we've figured out what the async and forking mechanism for code is.

s122
00:09:17.560 --> 00:09:20.000
You just spin up a bunch of sandboxes and do work.

s123
00:09:20.000 --> 00:09:23.880
What does that look like for the rest of the world?

s124
00:09:24.839 --> 00:09:27.440
So, here's a couple questions we think about pretty seriously.

s125
00:09:27.440 --> 00:09:32.400
One, you know, the models are trained on code, they want to write code, they're incredibly good at writing code.

s126
00:09:32.400 --> 00:09:35.360
How do we leverage these coding agents for actual knowledge work?

s127
00:09:35.360 --> 00:09:44.560
You You rather than wait for the models to catch up on doing services knowledge work, what if we just use that code knowledge and represent knowledge work as code?

s128
00:09:44.560 --> 00:09:47.520
Two, as I mentioned, engineers are used to paralyzing work.

s129
00:09:47.520 --> 00:09:49.360
How do you paralyze work that's traditionally serial?

s130
00:09:49.360 --> 00:09:54.280
You know, people clean out their inbox one email by one email, not 10 emails at once.

s131
00:09:54.280 --> 00:09:59.480
And finally, how do you move up the ladder here both in terms of product and user enablement?

s132
00:09:59.480 --> 00:10:01.040
What are the right form factors?

s133
00:10:01.040 --> 00:10:05.680
And I'd argue this varies dramatically from industry to industry.

s134
00:10:05.680 --> 00:10:13.080
Just because you have one way of launching an async agent for code, doesn't mean that same way is going to work for architecture or property management.

s135
00:10:13.440 --> 00:10:17.560
The second point I want to cover today is leveraging real-world data.

s136
00:10:17.560 --> 00:10:18.520
We all know this.

s137
00:10:18.520 --> 00:10:26.040
Frontier models have learned from everything humanity has written down, but the most valuable tasks are not on the internet.

s138
00:10:26.040 --> 00:10:29.462
How do you actually close the books when you're missing receipts?

s139
00:10:29.462 --> 00:10:29.560
[snorts]

s140
00:10:29.560 --> 00:10:35.040
How do you scope a building for construction in a blueprint, potentially collaboratively?

s141
00:10:35.040 --> 00:10:39.120
How do you coordinate vendors for fixing a broken roof?

s142
00:10:39.120 --> 00:10:47.760
All of this knowledge lives in people's heads, in 20-year-old software, uh in the way that one senior person on one of these teams just knows how to do it.

s143
00:10:47.760 --> 00:10:52.640
How do you make this information explicit and create tasks that you can actually learn from?

s144
00:10:52.640 --> 00:10:54.839
So, we've constructed a little bit of a flywheel.

s145
00:10:54.839 --> 00:10:59.080
We get our agents to collaborate with our employees to do real work.

s146
00:10:59.080 --> 00:11:02.839
And this allows us to generate rich traces of data and information.

s147
00:11:02.839 --> 00:11:08.560
Tool calls, the hiccups, the papercuts, everything that goes wrong with doing real work.

s148
00:11:08.560 --> 00:11:12.320
This in turn allows us to build real-world evals.

s149
00:11:12.320 --> 00:11:14.000
There is a ground truth here.

s150
00:11:14.000 --> 00:11:19.560
In the case of the roofing example, the question is, did the roof get repaired?

s151
00:11:19.560 --> 00:11:21.680
Did the books get closed?

s152
00:11:21.680 --> 00:11:26.640
And this allows us to hill climb and build better agents, which leads to more and more impact.

s153
00:11:26.640 --> 00:11:29.560
And what's really exciting is it ratchets up.

s154
00:11:29.560 --> 00:11:34.160
Every week our hill climbing benchmarks become a regression test.

s155
00:11:34.160 --> 00:11:37.640
So, our agents get better and better over time.

s156
00:11:37.680 --> 00:11:44.480
Just to drive a little bit deeper here on the traces, there's three upshots of being able to collect these rich traces.

s157
00:11:44.480 --> 00:11:49.120
One, we get to generate amazing evals that are built and scored automatically.

s158
00:11:49.120 --> 00:11:52.960
Uh and we're able to gather both implicit and explicit feedback.

s159
00:11:52.960 --> 00:11:59.400
Explicit feedback in the sense of thumbs ups and thumbs down, maybe people provide a note telling us whether this response was good or not.

s160
00:11:59.400 --> 00:12:01.320
Uh and also implicit feedback.

s161
00:12:01.320 --> 00:12:01.560
Right?

s162
00:12:01.560 --> 00:12:02.800
Again, we have the ground truth.

s163
00:12:02.800 --> 00:12:10.440
Maybe there's some data that the AI generated and there's a real diff between the data that the AI generated and what was ultimately submitted.

s164
00:12:10.440 --> 00:12:14.520
That's rich information that almost no one else has.

s165
00:12:14.520 --> 00:12:23.240
Two, we've started post training models internally on all of the data that these businesses operate on and produce, generally speaking.

s166
00:12:23.240 --> 00:12:27.640
This is all data that is completely out of distribution for most frontier labs.

s167
00:12:27.640 --> 00:12:29.880
Think of the task I showed at the beginning.

s168
00:12:29.880 --> 00:12:40.280
A lot of the models A lot of the frontier models today just can't do these tasks yet and we're trying to post train our own models internally to be able to do that on the rich source of data that we own.

s169
00:12:40.280 --> 00:12:43.839
And then finally, the actual agents themselves.

s170
00:12:43.839 --> 00:12:48.720
The real world is incredibly hairy and messy and you want customization per company.

s171
00:12:48.720 --> 00:12:51.040
Every company does things very differently.

s172
00:12:51.040 --> 00:12:52.640
Customization per user.

s173
00:12:52.640 --> 00:12:55.720
The way each user does their work is very unique.

s174
00:12:55.720 --> 00:12:57.480
And customization per client.

s175
00:12:57.480 --> 00:12:59.800
The way you work with every client is different.

s176
00:12:59.800 --> 00:13:04.000
It's a services business and you want to uphold those standards.

s177
00:13:04.440 --> 00:13:08.920
I love this picture because it's the whole thing in a single image.

s178
00:13:08.920 --> 00:13:13.320
Um the the way we usually talk about LLM tasks is the top panel.

s179
00:13:13.320 --> 00:13:13.600
Right?

s180
00:13:13.600 --> 00:13:15.680
You just It's It's a slope.

s181
00:13:15.680 --> 00:13:16.360
You got a bike.

s182
00:13:16.360 --> 00:13:19.520
And but there's clear sight to success.

s183
00:13:19.520 --> 00:13:22.440
The reality is most work is not like that.

s184
00:13:22.440 --> 00:13:24.839
And And you and I both know that.

s185
00:13:24.839 --> 00:13:27.200
Uh there are hills and ravines.

s186
00:13:27.200 --> 00:13:29.760
Uh there's death by a thousand paper cuts.

s187
00:13:29.760 --> 00:13:32.280
But But that's what real work looks like.

s188
00:13:32.280 --> 00:13:33.720
That's the entire job.

s189
00:13:33.720 --> 00:13:37.200
The exceptions are the job.

s190
00:13:37.720 --> 00:13:39.640
That's That's the demo.

s191
00:13:39.640 --> 00:13:42.400
That's the actual job.

s192
00:13:43.040 --> 00:13:49.080
Now, on to the final thing I want to chat with you guys today is learning loops within the enterprise.

s193
00:13:49.080 --> 00:13:54.040
I'd argue there's two hot trends everyone's talking about in 2026.

s194
00:13:54.040 --> 00:13:56.040
One, it's continual learning.

s195
00:13:56.040 --> 00:13:58.680
How do you make an agent better over time with feedback?

s196
00:13:58.680 --> 00:14:06.120
I think there are plenty of sessions uh this week on how you can use continual learning, whether it's in the prompt or in the weights.

s197
00:14:06.120 --> 00:14:07.880
And two, enablement.

s198
00:14:07.880 --> 00:14:12.920
How do you get in these enterprises and actually get them to adopt and use AI?

s199
00:14:12.920 --> 00:14:17.640
Traditionally speaking, these two initiatives are owned by two separate teams.

s200
00:14:17.640 --> 00:14:18.000
Right?

s201
00:14:18.000 --> 00:14:21.640
The continual learning is owned by your research team, your platform engineering team.

s202
00:14:21.640 --> 00:14:25.080
Enablement's owned by growth or deployment or customer experience.

s203
00:14:25.080 --> 00:14:29.520
Uh usually pretty siloed, not much interaction between the two.

s204
00:14:29.720 --> 00:14:32.200
We think these are part of the exact same loop.

s205
00:14:32.200 --> 00:14:37.000
The agent only improves if people actually use it.

s206
00:14:37.000 --> 00:14:41.160
And people only use the agent if it's worth adopting.

s207
00:14:41.160 --> 00:14:43.440
So, here's a little graphic of a snowball.

s208
00:14:43.440 --> 00:14:49.160
More usage drives continual learning, which drives a better agent, which drives more usage again.

s209
00:14:49.160 --> 00:14:53.760
All this to say, there's still a really big elephant in the room.

s210
00:14:53.760 --> 00:14:55.839
How do you get the initial usage?

s211
00:14:55.839 --> 00:15:04.280
I think a lot of people, you know, will use Claude Code or or give it to their whole enterprise, expect folks to just start using it.

s212
00:15:04.280 --> 00:15:08.360
Everyone assumes the usage just shows up.

s213
00:15:08.360 --> 00:15:11.080
But as we all know, that's simply not the case.

s214
00:15:11.080 --> 00:15:12.200
It never does.

s215
00:15:12.200 --> 00:15:12.560
Right?

s216
00:15:12.560 --> 00:15:17.680
Getting a 100-year-old firm to change its processes is hard.

s217
00:15:17.680 --> 00:15:22.040
You could have the best AI coworker on the internet or on Earth.

s218
00:15:22.040 --> 00:15:29.480
And if the people if the person who's closed the books for the last 20 years continues to do things the same way, nothing changes.

s219
00:15:29.480 --> 00:15:31.840
Nothing happens.

s220
00:15:32.240 --> 00:15:33.880
So, what can you actually do about it?

s221
00:15:33.880 --> 00:15:37.120
What you know, this this seems like incredibly hard.

s222
00:15:37.120 --> 00:15:38.240
What what's the upshot?

s223
00:15:38.240 --> 00:15:39.960
How do you actually get this stuff to work?

s224
00:15:39.960 --> 00:15:47.840
Well, I think a lot about Jensen and how he dominated the market in his words with extreme hardware software co-design.

s225
00:15:47.840 --> 00:15:52.960
Designing the chips and the software together as one system.

s226
00:15:52.960 --> 00:15:57.200
We look at this through the lens of extreme software service co-design.

s227
00:15:57.200 --> 00:16:02.560
How do you co-design our products with the people and the processes at our businesses?

s228
00:16:02.560 --> 00:16:08.440
And I'd argue this is only possible from being within under the same roof.

s229
00:16:08.480 --> 00:16:20.480
We need to meet the people within these companies both metaphorically, for example, bringing products to their systems so that the energy required for enablement is kept low, and also physically.

s230
00:16:20.480 --> 00:16:25.600
Get on a plane, show up, say hi, learn what people actually do.

s231
00:16:25.600 --> 00:16:38.040
You know, maybe you build a product that's natively embedded into Excel or into their ERP system, maybe their 3D design software, or or maybe even their Microsoft products like Outlook, Gmail, etc.

s232
00:16:38.040 --> 00:16:39.080
Or you show up in person.

s233
00:16:39.080 --> 00:16:41.720
You do a lunch and learn with a bunch of folks at one of the companies.

s234
00:16:41.720 --> 00:16:45.640
You go to their conferences and you create cotton candy and run a stand for them.

s235
00:16:45.640 --> 00:16:50.920
You go mountain biking and ask them about all the difficulties that they have with their actual day-to-day jobs.

s236
00:16:50.920 --> 00:17:02.120
Or you show up in person one-on-one or sometimes even two-on-one in this case and just show them how to use the tools and learn from the feedback because this is what the rest of the world really looks like.

s237
00:17:02.120 --> 00:17:04.600
It's not like the folks in this room or in San Francisco.

s238
00:17:04.600 --> 00:17:06.280
It's a lot more like this.

s239
00:17:06.280 --> 00:17:11.360
You cannot co-design software with the services business over Zoom or over a support ticket.

s240
00:17:11.360 --> 00:17:12.480
You you have to be there.

s241
00:17:12.480 --> 00:17:14.040
You have to be in person.

s242
00:17:14.040 --> 00:17:17.240
And I'd argue this is the part that actually makes it work.

s243
00:17:17.240 --> 00:17:22.160
In order to get AI diffusion to work, you have to touch some grass.

s244
00:17:22.160 --> 00:17:23.199
Thank you so much.

s245
00:17:23.199 --> 00:17:25.880
I'll be around for the rest of day if there's anything I can help with.

s246
00:17:25.880 --> 00:17:27.680
My email is up there.

s247
00:17:27.680 --> 00:17:30.400
And yeah, thank you.

s248
00:17:43.732 --> 00:17:45.732
[music]