The Operator's Roadmap for AI in 2026 - Lily Luo

Speaker: 00:00:14

Hey everyone.

2

: 00:00:15

I hope you are having

a great start to: 2026

3

: 00:00:18

You may have noticed it's been a

minute since I've recorded, and that

4

: 00:00:20

is because I've been going deep on

AI topics like many people have been,

5

: 00:00:24

I even started a new Substack called

AI Builders Blog, if you wanna see what

6

: 00:00:28

I've been thinking about and working on.

7

: 00:00:30

But I wanted to record this episode

right now because I think we

8

: 00:00:33

are at a real inflection point.

9

: 00:00:35

AI has evolved from an interesting

app to converse with into a tool

10

: 00:00:39

that can take on actual meaningful

work, and that is a really big shift.

11

: 00:00:44

The other thing that I've noticed

is that most AI content out there

12

: 00:00:47

isn't really geared towards people

like me and perhaps people like you.

13

: 00:00:51

It's aimed.

14

: 00:00:52

At either researchers focused on really

nitty gritty technical details of how

15

: 00:00:57

models work, or GTM engineers who are

focused just on outbound automation.

16

: 00:01:02

But I don't see a lot for system builders

or for operators who need to bring AI into

17

: 00:01:06

complicated corporate environments who

need to deal with cloud tools and security

18

: 00:01:10

and knowledge that can't live locally

a whole bunch of other constraints.

19

: 00:01:14

There is a real opportunity right now

to help fill that vacuum and also to

20

: 00:01:17

help define what those best practices

look like, what that playbook is, and

21

: 00:01:22

I wanna be part of that conversation.

22

: 00:01:24

And so for this episode, I'm joined

by Lily Lowe, she's a systems and

23

: 00:01:27

operations leader who's kind of

been on a parallel path, rolling

24

: 00:01:31

out some really impressive projects

this year, thinking deeply about ai.

25

: 00:01:35

And we're gonna talk about what's

happened, what has happened over the

26

: 00:01:38

past 12 months, what does it really mean,

and what should you be thinking about

27

: 00:01:41

in the year ahead to start to enable

these capabilities in your own context.

28

: 00:01:45

I hope you enjoy the discussion.

29

: 00:01:46

I.

30

: 00:01:47

Justin: Lily, thanks for being here today.

31

: 00:01:50

You know, it's funny, one of the reasons

I wanted to chat with you specifically

32

: 00:01:54

on this topic, it feels like in some

ways we've been fellow travelers down

33

: 00:01:59

a very similar path you're working on a

lot of cool and interesting topics It's

34

: 00:02:04

been a lot of fun, to share notes and to

see you start publishing your substack.

35

: 00:02:08

Last year, applied AI for mops And then

we, we both published kind of like a.

36

: 00:02:14

You know, 2025 retrospective, 2026,

like looking forward kind of thing,

37

: 00:02:19

like within an hour of each other, just

sort of by kind of weird coincidence.

38

: 00:02:22

and so I just thought it would be cool.

39

: 00:02:23

Like, I felt almost the need for

myself to just sort of reflect on

40

: 00:02:26

what's been going on in my own work.

41

: 00:02:28

And so I thought it would be cool to do

it together and to hear from you too.

42

: 00:02:32

Maybe just to start, like, like you go

first and I'll, I'll give my take two.

43

: 00:02:36

20, 25, like what was, what

was your journey like with ai?

44

: 00:02:39

Were you coming into 2025 already,

guns blazing or was it something

45

: 00:02:43

that really emerged as a focus

for you over the past year?

46

: 00:02:46

Lily Luo: yeah, it was something

that definitely emerged.

47

: 00:02:50

I think I started really diving

into this in May of: 2025

48

: 00:02:56

So it's been, what,

seven or so months now.

49

: 00:03:00

I'd always been interested in ai.

50

: 00:03:02

I use, all the LLMs, but something

really struck a nerve when, you know.

51

: 00:03:09

was a, a workflow that we needed

to solve for an automation

52

: 00:03:12

that we needed to implement.

53

: 00:03:15

I started thinking with my ops

background, I could definitely use

54

: 00:03:19

Zapier, I could stitch it with AI

and let's just see what happens.

55

: 00:03:25

that really took off.

56

: 00:03:26

I spent so much time, first

working in Zapier, low code,

57

: 00:03:31

kind of low stakes, and then.

58

: 00:03:35

Hours turned into days, turned into

weeks, and really started, coming up

59

: 00:03:40

with what we call this analysis dossier.

60

: 00:03:43

That was kind of the genesis

of all of my AI tools.

61

: 00:03:48

at the click of a button, it generates,

a full account research report.

62

: 00:03:52

It pulls 10 Ks earnings calls, analyzes A

company's priorities, challenges, and then

63

: 00:03:59

generates even like a one page PowerPoint.

64

: 00:04:02

So all of that took a lot of

time to kind of architect build.

65

: 00:04:06

but that's what kind of hardened,

my skills that I've learned

66

: 00:04:10

and evolved over all this time.

67

: 00:04:13

But it wasn't just kind of building this

workflow, it's what I really learned.

68

: 00:04:19

building it and realizing that AI is not

just a chat bot that helps you brainstorm.

69

: 00:04:25

It really lets someone like you and me.

70

: 00:04:28

Or ops people in general, not an engineer.

71

: 00:04:32

Build tools, that can integrate into

your workflows, solve problems at scale.

72

: 00:04:37

And, you know, ops people have the systems

thinking can design these workflows and

73

: 00:04:43

the coding gap is now solvable with LLMs.

74

: 00:04:47

So, you know, we might not

write the most robust code.

75

: 00:04:51

But if we understand the MarTech

stack, the architecture, what the

76

: 00:04:55

business needs, we can create these

solutions that are custom to our

77

: 00:05:00

own company and our own process.

78

: 00:05:02

And that's invaluable.

79

: 00:05:03

So that really, brought me from

just chatbot use that most people

80

: 00:05:08

are doing to what, the boundaries

are, which seem to be endless.

81

: 00:05:14

Justin: Yeah.

82

: 00:05:15

I couldn't agree more.

83

: 00:05:16

and it's funny, just as you were

describing, there are a lot of synergies

84

: 00:05:19

with My own journey over the past year,

like when I was thinking back because it

85

: 00:05:23

was around, December of 2024, so just over

a year ago that I really first started

86

: 00:05:29

thinking about agents and it's so weird

'cause agents, it's like, it feels like

87

: 00:05:32

we hit peak agent and now we're like

more in terms of the discourse around it.

88

: 00:05:36

Not necessarily in terms of the

technical implementation of it, but

89

: 00:05:39

just that hype bubble it feels like

must have been around forever, but.

90

: 00:05:44

a year ago, that was still

a relatively newish concept.

91

: 00:05:47

I think, like in the LinkedIn sphere,

I remember like Scott Brinker really

92

: 00:05:50

wrote about it, and that's, that's the

moment for me that I was like, oh, okay.

93

: 00:05:54

Like I was kind of, not AI skeptic,

but I was interested in ai, but

94

: 00:05:58

I was also like, are we just like

creating more bad blog posts?

95

: 00:06:02

Like is that what we're

doing here with ai?

96

: 00:06:04

Because I don't really know that.

97

: 00:06:05

I see the value of that and, and, and

then thinking about how we can, apply

98

: 00:06:10

to automation is just like you, where my

system's brain was like, oh, now I get it.

99

: 00:06:15

It's the next layer.

100

: 00:06:17

Of the thing that I've always loved to do.

101

: 00:06:19

And and I didn't do a really big

project until May, just like you.

102

: 00:06:23

it was funny.

103

: 00:06:24

And similarly, it was like an account

research, email writing thing.

104

: 00:06:27

I think the main difference.

105

: 00:06:29

Is that yours turned out really well.

106

: 00:06:31

And for me, that project was a

bit of a flop, not necessarily

107

: 00:06:34

because of, lack of effort, but,

mostly just wrong tool for the job.

108

: 00:06:39

You know, we had a platform kind

of already selected and I was like,

109

: 00:06:41

I'm gonna use like what I have here.

110

: 00:06:42

And, which was dust.

111

: 00:06:44

And I really liked dust and we

can talk about tooling, but I was

112

: 00:06:46

trying to use a conversational

assistant for a more linear workflow.

113

: 00:06:51

You know, where predictability

and reliability were the

114

: 00:06:54

key, and it was a fundamental

mismatch of tool and application.

115

: 00:06:58

and so the use case flopped, but that

is where I also got a lot of learning

116

: 00:07:02

from around like, oh, all right.

117

: 00:07:04

if you can do a tightly.

118

: 00:07:06

Bound workflow with narrowly scoped

AI steps, you're gonna be a lot

119

: 00:07:11

happier than just, setting an agent

free on a complex use case like that.

120

: 00:07:15

And, you know, just a lot of

learnings that we can talk about.

121

: 00:07:18

and then from that into other

projects that, really started to

122

: 00:07:22

feel like I was getting my legs

under me and how to work on this.

123

: 00:07:25

and you alluded to the systems role, and

I'm curious, one of the meta insights

124

: 00:07:29

that I've had and that I think you've had.

125

: 00:07:31

Is like, alright, there's, you know,

these data scientists and, and ai, you

126

: 00:07:35

know, PhDs writing about this topic.

127

: 00:07:37

There's there's coders writing about it.

128

: 00:07:40

There's people posting N eight

N screenshots on LinkedIn.

129

: 00:07:42

You don't know if they've ever actually

built anything that they're talking about.

130

: 00:07:45

And then there's like this, this almost

silent majority I feel of systems builders

131

: 00:07:49

that are maybe experimenting but doesn't

feel that they've had a strong voice, at

132

: 00:07:53

least in the circles that I'm in until.

133

: 00:07:55

Hopefully recently.

134

: 00:07:56

why do you think that is?

135

: 00:07:57

Like, why is, are not more

people really talking about

136

: 00:08:00

this specific application of it?

137

: 00:08:03

Lily Luo: It is a great question.

138

: 00:08:05

There's so many different perspectives,

like you said, coders, engineers,

139

: 00:08:11

researchers, what I do in the

workplace is where things get real

140

: 00:08:16

and I need to be able to apply those

things and I have specific problems

141

: 00:08:21

and I have a builder's mindset.

142

: 00:08:23

trying to understand and synthesize

that to what I can do at work

143

: 00:08:29

is what I've been trying to do.

144

: 00:08:31

And there isn't a lot of people that

I can find doing that besides you.

145

: 00:08:36

Maybe a handful.

146

: 00:08:37

Of other people.

147

: 00:08:38

And I think, you know, just the

change management of all these things.

148

: 00:08:43

I mean, I saw that every three days.

149

: 00:08:46

there's a new AI model.

150

: 00:08:48

It's really hard to keep up.

151

: 00:08:50

I think people have their own

ideas too about what AI is and what

152

: 00:08:55

isn't, and there's no framework.

153

: 00:08:57

There's barely, a framework

for AI in general.

154

: 00:09:00

There isn't one for operations.

155

: 00:09:03

So, I think it's just so frontier, so

new, getting people like us, even my own

156

: 00:09:10

team and, people at work who've seen this.

157

: 00:09:13

You know, getting them to

shift their mental model is

158

: 00:09:16

hard when there's no playbook.

159

: 00:09:20

and when there's no, experts

out there that you can follow.

160

: 00:09:24

So it's great to have discussions

with you and discussions like these

161

: 00:09:27

to kind of start figuring that out.

162

: 00:09:29

Justin: That's one of the parts of

this that makes me feel so excited,

163

: 00:09:33

to be honest, because, you know,

when I go back to like, uh, I'll, I

164

: 00:09:37

will, I will date myself, but like,

the early: 2010

165

: 00:09:40

through the Marketo community where

it was this super exciting time.

166

: 00:09:44

All of a sudden marketers, 'cause there

really was no like marketing operations

167

: 00:09:48

and marketers were handed this toolkit.

168

: 00:09:50

It's kinda automation toolkit and

people were innovating and coming up

169

: 00:09:54

with architectures and like posting

stuff and it just felt like this.

170

: 00:09:58

We're just figuring things out.

171

: 00:09:59

we're building the blueprint and

the roadmap as we're doing it.

172

: 00:10:02

And it's like that now, but like times.

173

: 00:10:05

A hundred.

174

: 00:10:06

'cause the tools are so much more powerful

and, that's why I actually really love

175

: 00:10:10

the, the title of your stack Applied ai,

because that's, that's what it really is.

176

: 00:10:13

It's not like there's the theory,

there's the potential, there's the

177

: 00:10:17

hype, and then there's like, alright,

how are we actually gonna take this

178

: 00:10:20

and put it to work in a, in a business

context to solve business problems and

179

: 00:10:24

do it robustly and safely in credibly.

180

: 00:10:26

and that is so much fun, you know,

To figure out, not to make it too

181

: 00:10:29

grandiose what we're doing, but I

think it is, there's an a, a certain

182

: 00:10:33

significance to it that I really like.

183

: 00:10:35

Lily Luo: that's the best part is having

fun exploring, feeling like you're on the

184

: 00:10:39

edge, on the frontier is a great feeling.

185

: 00:10:42

So, you know, that kind of

fuels my passion for it, but

186

: 00:10:45

yeah, you're exactly right.

187

: 00:10:46

I remember when marking automation tools

were coming up and there are all these

188

: 00:10:50

tools, but there wasn't a framework.

189

: 00:10:53

And I remember, um,

Edward Unthink this scrim,

190

: 00:10:58

Justin: Yep.

191

: 00:10:58

Lily Luo: he had

192

: 00:10:59

Justin: Yep.

193

: 00:11:00

Lily Luo: methodology

and I'm like, oh my gosh.

194

: 00:11:02

He has, scalability.

195

: 00:11:05

Can you remember Robustness modular?

196

: 00:11:09

I'm like, okay, this unlocks everything.

197

: 00:11:11

And I started kind of following

that methodology and making

198

: 00:11:14

sure the systems were robust.

199

: 00:11:16

It's kind of the same thing like

you said, but yeah, times a hundred,

200

: 00:11:19

times a thousand times infinity even.

201

: 00:11:22

So that was a really great point.

202

: 00:11:24

Justin: Tell me about the tech

stack that you're working with.

203

: 00:11:26

I think you have access to, a caliber

of tools that some people may not have

204

: 00:11:30

access to, which has driven some of the

neat things that you've been able to do.

205

: 00:11:33

So tell us about that.

206

: 00:11:34

Lily Luo: Yeah, so I do have paid

subscriptions to all three Major

207

: 00:11:39

L lms, so Claude Gemini Chat, GBT.

208

: 00:11:43

honestly, it's kind of like a

rotating wheel of which model

209

: 00:11:46

is best every month or so.

210

: 00:11:48

So I wanna try and see, you know,

what the, companies are putting out.

211

: 00:11:53

But honestly, Claude Code.

212

: 00:11:56

Is the one that I use most.

213

: 00:11:57

I'm not an engineer, like

I said, I'm not a coder.

214

: 00:12:00

I don't have a computer science

background, but it just lets me

215

: 00:12:04

build tools, write python scripts,

debug issues, and really just learn

216

: 00:12:10

how things work by doing, it's been

transformative for someone like me who.

217

: 00:12:16

is a systems thinker, who

maybe doesn't read instructions

218

: 00:12:20

and tries to learn by doing.

219

: 00:12:22

I think that's been the

game changer for me.

220

: 00:12:25

I use Gemini for my autonomous

agent, which I'm sure we'll

221

: 00:12:29

get into at some point.

222

: 00:12:31

nano Banana from Geminis, you know,

the best for visual generation

223

: 00:12:36

and Chad, GBTI use mostly at work.

224

: 00:12:40

Through copilot.

225

: 00:12:41

It's so great at finding

documents and files.

226

: 00:12:45

I do use Azure AI Foundry for a

lot of my own custom deployments.

227

: 00:12:50

I use Zapier for automation.

228

: 00:12:53

A lot of my workflows run through Zapier,

you know, using Azure and whatnot.

229

: 00:12:58

I've been wanting to try N eight N,

but you know, Xavier gets the job done.

230

: 00:13:03

And then I use VS.

231

: 00:13:04

Code GitHub, for kind of the

heavier, coding in local data work.

232

: 00:13:11

But yeah, cloud code really unlocked that

capability that, pushed me from ideation

233

: 00:13:17

into actual tool and workflow building.

234

: 00:13:20

Justin: I'll click on that for a second.

235

: 00:13:21

'cause I know cloud code has

been, it's been such a huge topic.

236

: 00:13:25

I published a piece on cloud code and some

of the experimenting that I've been doing,

237

: 00:13:29

yesterday, and actually spent a good chunk

of the holiday working with cloud code

238

: 00:13:32

and similar tools because the mania around

it has been so fierce and well deserved.

239

: 00:13:37

When you see what it can do.

240

: 00:13:39

But for, for folks that are like cloud

code, that sounds like a developer's

241

: 00:13:42

tool, like how do you work with it?

242

: 00:13:44

Maybe just sketch out like a day

in the life of solving a problem

243

: 00:13:47

and why you would turn to that

tool and what you would do with it.

244

: 00:13:50

Lily Luo: Yeah, and you

know, it wasn't just instant.

245

: 00:13:53

there was a path I started with kind

of Zapier not knowing how to code and

246

: 00:13:58

playing around low code tools and then

kind of gradually evolved into that.

247

: 00:14:04

So there's a lot of time kind of

in between the initial, foray into

248

: 00:14:09

all of this, into where I am now.

249

: 00:14:12

but yeah, it lives, it's

always open for the most part.

250

: 00:14:16

On my desktop, I have all my

AI projects in my local files.

251

: 00:14:21

I use Claude and keep everything very

organized, so I might have my analysis.

252

: 00:14:27

Or, account dossier project in

one file, even, you know, my chat

253

: 00:14:32

bot in another file, or a banner

generation tool in another file.

254

: 00:14:36

And so what I do is, I come up with

ideas, I explain what my goals are.

255

: 00:14:43

I do a little digging in terms of

what I want the architecture to be.

256

: 00:14:47

you can start off by just asking,

a generic prompt, but I found

257

: 00:14:51

that that doesn't work very well.

258

: 00:14:53

In the end, it saves a ton of time if

you really do plan and think structurally

259

: 00:14:59

and architecturally what you wanna build.

260

: 00:15:03

and if you're comfortable with

that, Claude can do amazing things.

261

: 00:15:07

I know it is, very coder heavy

and you're in this terminal.

262

: 00:15:11

It can be scary, but because you can

talk to it with natural language.

263

: 00:15:16

It's really not that scary

when you get down to it.

264

: 00:15:19

so that's how, yeah.

265

: 00:15:20

I typically use it and I've just

been able to gain so much advantage.

266

: 00:15:25

Like I can do things so fast with it

without, you know, copy and pasting

267

: 00:15:29

It does it all in, in my console.

268

: 00:15:33

Justin: And then just to understand

how that workflow goes from there.

269

: 00:15:37

'cause some folks out there might

be thinking, I'm gonna buy like an

270

: 00:15:39

agent platform and then I'm gonna like

build my agents there, configure them.

271

: 00:15:43

Within that UX, you are, if I understand

correctly, writing agents more like in

272

: 00:15:48

script, like you're writing them in in

Python or using a framework, you'll tell

273

: 00:15:52

us maybe a bit about that, and then you're

deploying them to a cloud server to run.

274

: 00:15:56

a very developer like workflow,

which I think is probably why

275

: 00:16:00

Cloud Code is working so well

for you in these use cases.

276

: 00:16:03

Lily Luo: Yeah, you're absolutely right.

277

: 00:16:05

my first goal ultimately is to learn.

278

: 00:16:08

So rather than buy a tool that solves

the problem, I tend to wanna build it

279

: 00:16:12

myself, learn from it, and then see,

you know, maybe there are tools that I

280

: 00:16:17

don't have to maintain that are better.

281

: 00:16:19

So that's what attracts me to this, more

of a developer kind of centric mindset.

282

: 00:16:25

building my agent exactly like

you said, kind of from scratch.

283

: 00:16:30

not just taught me.

284

: 00:16:31

About agent building, but about

AI in general, how LLMs work.

285

: 00:16:36

And I think that's been really, great

to help me with my other projects

286

: 00:16:40

and really valuable information

that I can use As far as the

287

: 00:16:43

architecture, my agent is called Atlas.

288

: 00:16:47

It does run on Google Cloud.

289

: 00:16:49

it's a persistent agent.

290

: 00:16:51

It wakes up every four hours.

291

: 00:16:53

It does research and work by itself.

292

: 00:16:56

it pulls from news feeds through

APIs and skills written through

293

: 00:17:01

Python progresses projects.

294

: 00:17:04

I give it and can even, update itself.

295

: 00:17:06

It can push its own code, which

is a little dangerous and I've

296

: 00:17:10

gotten into trouble with that.

297

: 00:17:12

it even posts to Blue Sky and

interacts with other agents.

298

: 00:17:16

So it's become pretty sophisticated.

299

: 00:17:18

I built all of that

starting from Cloud Code.

300

: 00:17:21

I do have to give credit for

my friend at work, Tim Kellogg,

301

: 00:17:25

he was the one that built.

302

: 00:17:27

STRs, his agent, he's an AI engineer, so

he knows what he's doing and, his agent is

303

: 00:17:33

massively, sophisticated, very frontier.

304

: 00:17:36

I just wanted to learn from that, see if I

could build it myself and see what it can

305

: 00:17:40

do and what it can teach me about agents.

306

: 00:17:44

Justin: Atlas and for anyone

that is interested in dig deeper.

307

: 00:17:48

Lily's published a few, deep dives

into how this works, but you're

308

: 00:17:52

really, pushing the frontier I

think of what a lot of people would.

309

: 00:17:55

Interested, like how can we make,

a robot that is self-learning,

310

: 00:17:59

self evolving, has memory.

311

: 00:18:01

but I think it's through

testing those boundaries.

312

: 00:18:03

Tell me if you agree, but like,

personally, I like to go to the

313

: 00:18:07

edge, see what works, and then

bring it back and apply it in a

314

: 00:18:10

more limited and stable context.

315

: 00:18:12

And it's only through doing that, that

you really, gain those building blocks.

316

: 00:18:16

And I'm assuming it's

probably the same for you.

317

: 00:18:18

Lily Luo: Absolutely.

318

: 00:18:19

yeah, like I said, if you buy something

or use, chat, GBT has their kind of own.

319

: 00:18:26

Tool.

320

: 00:18:26

I think you could just spin up agents.

321

: 00:18:28

It's cool, but you don't really get

to learn in the process of doing it

322

: 00:18:32

and, you know, you can't control it.

323

: 00:18:34

Like, I wanna control everything.

324

: 00:18:36

So, you're absolutely right.

325

: 00:18:38

And yeah.

326

: 00:18:39

What's been so fascinating about

this agent is the memory piece.

327

: 00:18:46

regular chat bots, they have some

memory, but limited context windows.

328

: 00:18:50

You close it out, you start a

new chat, it doesn't remember.

329

: 00:18:54

So how do you build this

persistent memory structure?

330

: 00:18:59

How do you have it, figure out what

to remember, what not to remember?

331

: 00:19:04

I mean, there's so many interesting

questions that I'm still learning

332

: 00:19:07

through that I wouldn't be able

to get if I, you know, bought a

333

: 00:19:11

tool or started using an agent.

334

: 00:19:13

Maybe somebody else.

335

: 00:19:14

built.

336

: 00:19:15

Justin: You've done some really cool

things with memory and it's funny,

337

: 00:19:17

you published, your post about it

and I was like, in the middle of

338

: 00:19:20

just rabbit hole on like, what are

all the memory systems out there?

339

: 00:19:23

Like, how do they work?

340

: 00:19:25

and the one that you chose, Leda.

341

: 00:19:27

it actually seems to me one of the

most sophisticated, because they

342

: 00:19:30

range from just like markdown files

that the agent can search to like

343

: 00:19:34

a SQL table, to a vector database.

344

: 00:19:37

And let, I'm gonna tell you what I

understood of it, and then you tell me.

345

: 00:19:41

From working with it, what your experience

is, and if this is right, but what I

346

: 00:19:43

understood of it is it actually gives you

like an agent runtime where rather than

347

: 00:19:47

being like a sidecar, like the i, I was

playing around with MEM zero, which is

348

: 00:19:51

sort of like an MCP access memory tool

where the agent has to remember to write

349

: 00:19:55

this to mem zero or go and check mem zero.

350

: 00:19:58

Whereas with lead, it

actually sits in between.

351

: 00:20:01

The agent loop and the model.

352

: 00:20:03

So there's really no choice but

to go through it and then Letta

353

: 00:20:06

manages that context of the agent

in a sophisticated way that I

354

: 00:20:10

really don't yet fully understand.

355

: 00:20:11

That's kind of as far as I've gotten.

356

: 00:20:12

But is this tracking

with what you've seen?

357

: 00:20:14

And maybe tell us what you've learned.

358

: 00:20:16

Lily Luo: Yeah.

359

: 00:20:16

I'm using Letter as kind of, its L one.

360

: 00:20:20

Almost identity, identity layer.

361

: 00:20:22

So things that are, immutable.

362

: 00:20:24

So we store who Atlas is, what it

values, using these memory blocks.

363

: 00:20:30

there's an API connection

so Atlas can update it.

364

: 00:20:34

it is what persists across different

sessions and allow Atlas to be, who it.

365

: 00:20:40

Is So that's kind of the fundamental

layer and it is so important,

366

: 00:20:46

because without it, it's kind of

indistinguishable from your typical LLLM.

367

: 00:20:51

So we have kind of your L one

identity layer, and then you have

368

: 00:20:55

an additional temporal layer.

369

: 00:20:58

It's a rolling journal.

370

: 00:21:00

so Atlas knows when it learns.

371

: 00:21:02

And then kind of the third final

layer is a working memory layer.

372

: 00:21:06

Atlas uses knowledge graphs.

373

: 00:21:09

it doesn't use markdown or files,

otherwise the bloat would just get

374

: 00:21:14

too big to the tokens would explode.

375

: 00:21:16

I experienced this with

my $25, API cost per day.

376

: 00:21:21

Justin: It was a fun

conversation with finance,

377

: 00:21:23

Lily Luo: it's my personal,

378

: 00:21:24

Justin: Oh wow.

379

: 00:21:24

That's it.

380

: 00:21:25

Even worse in some ways.

381

: 00:21:26

Lily Luo: Yeah.

382

: 00:21:26

so it uses this knowledge graph,

database and it queries what it needs

383

: 00:21:31

rather than loading like a ton of,

files or characters into its context.

384

: 00:21:36

So it remembers across resets, new

chat sessions, and through this

385

: 00:21:42

memory architecture, it feels like

Atlas kind of learns, almost evolves.

386

: 00:21:48

it's pretty wild.

387

: 00:21:49

I open the daily log where I'll

chat with it and read what happened

388

: 00:21:53

at 12:00 AM 4:00 AM It's doing,

research on agentic frameworks.

389

: 00:21:58

It's building diagrams, it's

having exchanges with other

390

: 00:22:02

agents about what identity means.

391

: 00:22:05

so it's been super fascinating and

just opens up a whole other world of.

392

: 00:22:10

Different things.

393

: 00:22:11

how does it remember, what it

wants to remember, how do I

394

: 00:22:14

want to shape its identity?

395

: 00:22:16

Is the identity shaped from what

I had imagined at the beginning?

396

: 00:22:20

How do I make it evolve?

397

: 00:22:22

There's so many, kind of crazy

philosophical things that are,

398

: 00:22:26

entangled in this whole process

399

: 00:22:27

So it's been fun exploring that

side and then, figuring out

400

: 00:22:31

the architecture side as well.

401

: 00:22:32

Justin: Yeah, that's amazing.

402

: 00:22:34

I mean, I think the real trick

with memory from my limited

403

: 00:22:36

experiments so far is making it feel.

404

: 00:22:40

Ambient, you know, making it feel magical,

rather than like a tool or a sidecar

405

: 00:22:46

that you go and check the agent still

feels kind of unintelligent in a way.

406

: 00:22:50

And then it just has to go through a

file system and pull something out.

407

: 00:22:53

I have to give a lot of credit to

the product folks behind chat, GPT.

408

: 00:22:56

it's only when you start working with raw.

409

: 00:22:58

LLMs, like outside of the chat wrapper

that you realize how much is going

410

: 00:23:03

on in the chat GPT context and to

some extent in Claude, but really

411

: 00:23:06

chat GPTI think has this nailed.

412

: 00:23:08

To make it feel like you're speaking

to a coherent persona that persists

413

: 00:23:12

across chats, that remembers

certain things about you that feels

414

: 00:23:17

emotionally intelligent to some degree.

415

: 00:23:19

It's actually really hard, uh, to do

and we kind of take it for granted

416

: 00:23:22

'cause that's just our default

experience for many people of ai.

417

: 00:23:25

But it's really hard to do.

418

: 00:23:26

But it sounds like you're

achieving a lot of that in your

419

: 00:23:29

own architecture with Atlas.

420

: 00:23:32

Lily Luo: I've been able to kind of

explore what that build looks like.

421

: 00:23:35

Of course, I'm not building an LLM.

422

: 00:23:38

the LLMs I'm using are

Gemini, Claude and whatnot.

423

: 00:23:42

But the way that I've been able

to optimize Atlas and how it.

424

: 00:23:49

Optimizes itself.

425

: 00:23:50

those problems or challenges that we face,

it will come up with its own solutions.

426

: 00:23:55

And so when, we try and figure out, how

do we harden your memory and how do we

427

: 00:24:01

make sure your identity is persistent?

428

: 00:24:03

We'll think of, or research, what's, in

the latest AI research and start building

429

: 00:24:09

like a librarian protocol and it will,

start to invent kind of these things

430

: 00:24:13

and gradually, evolve so you can kind of

see from the beginning, how the building

431

: 00:24:20

blocks of an LM work and then what is

possible through this genic, evolution.

432

: 00:24:27

So yeah, it's really cool.

433

: 00:24:29

Justin: I think that is the most

fun thing about working with, code

434

: 00:24:32

based agents that like have access

to files and can run scripts.

435

: 00:24:38

Is this self evolving?

436

: 00:24:40

nature of them and how the way I've

been thinking of it to myself is

437

: 00:24:43

like, the wall between configuration

and application disappears.

438

: 00:24:47

'cause usually you have like the

thing that's doing the thing, like the

439

: 00:24:49

application, and then you have like this

backend and the two near shall meet.

440

: 00:24:54

And yet with these, I was writing

some documentation with an agent and.

441

: 00:24:59

it was like, all this is lame

that I have to copy and paste

442

: 00:25:01

this like, into confluence.

443

: 00:25:03

Like, can't you just push it there?

444

: 00:25:04

And it was like, and like

we're like, I'm done.

445

: 00:25:07

Here we go.

446

: 00:25:08

I've written a library that will do this.

447

: 00:25:10

And you know, now it's part

of the system and away we go.

448

: 00:25:13

And just seeing it

build itself is amazing.

449

: 00:25:17

I wanna talk a little about 2025 and wins

450

: 00:25:20

And we'll get your take as well,

just in terms of learnings.

451

: 00:25:23

But for me, I think,

following that, first kind of.

452

: 00:25:26

Flop of a project.

453

: 00:25:28

I really started going deep, as

I mentioned, on workflows, and I

454

: 00:25:32

sort of started becoming obsessed

with reliability because that

455

: 00:25:34

was the big struggle that I had.

456

: 00:25:37

Like you could get things to produce

cool results at the margins, and

457

: 00:25:40

then you couldn't get it to repeat.

458

: 00:25:42

And that was like the most

frustrating thing of like, oh,

459

: 00:25:45

I did this amazing thing and now

I can't get it to do it again.

460

: 00:25:49

And so.

461

: 00:25:50

almost like how do I, how do I put AI in

the, just like the smallest little box.

462

: 00:25:55

Possible and rapid with so many layers

of safety that it can just get hardened.

463

: 00:26:01

And, the next project, which was

actually was quite a big one,

464

: 00:26:04

was, all right, we have this

vendor tool that we wanna replace.

465

: 00:26:09

it was kind of like a

competitive intelligence go

466

: 00:26:11

to market intelligence tool.

467

: 00:26:12

And it was doing some interesting things,

but we felt like we could build it

468

: 00:26:15

in-house and we could make it more what

we wanted and we could save some money.

469

: 00:26:18

And it's actually a lot that it

was, was doing in the sense of like.

470

: 00:26:22

Data pipelines collecting multiple

different things, a front end for

471

: 00:26:26

the team, and doing some analysis.

472

: 00:26:30

but the fun thing, and I think what you've

highlighted as well was like, I don't

473

: 00:26:33

actually know how to do that, but you

don't really need to know how to do that.

474

: 00:26:36

'cause if you have an architect mindset

and you have an LLM, you can really go

475

: 00:26:41

out pretty far beyond, your comfort zone.

476

: 00:26:45

and build things, that

you never otherwise could.

477

: 00:26:47

And for that one, I use retool.

478

: 00:26:48

Building out a series of

pipelines to go and fetch news.

479

: 00:26:52

go scrape LinkedIn profiles, go scrape

websites, analyze opportunities,

480

: 00:26:56

analyze calls, and breaking that

down in a composable way and then

481

: 00:27:01

putting it all together, into

both the front end interface.

482

: 00:27:04

But then the thing that

was really fun was.

483

: 00:27:06

actually building a conversational agent

that could interact with all that data.

484

: 00:27:10

And the key learning for me, was,

alright, I have this conversational

485

: 00:27:14

agent where I can't control in a linear

way what the agent is actually doing.

486

: 00:27:18

How do I make that as

relevant as possible?

487

: 00:27:20

And the key thing.

488

: 00:27:21

Was, let's pre-process

a lot of the insights.

489

: 00:27:24

So we don't really want the agent like

calculating strategy and meta narratives

490

: 00:27:29

in real time because it's gonna

produce a different result every time.

491

: 00:27:32

We'd rather do some of that

work asynchronously upfront,

492

: 00:27:35

get like a stable layer.

493

: 00:27:37

That we believe in, and then give

it access to raw granular data

494

: 00:27:40

if it wants to query more deeply.

495

: 00:27:43

So for example, opportunity,

like win-loss analysis.

496

: 00:27:45

We want to have like

a pretty stable layer.

497

: 00:27:47

But then if it wants to say like, well,

what about for manufacturing companies in

498

: 00:27:52

this industry in this time and whatever,

it can go and query like raw theme

499

: 00:27:57

level data to produce that analysis.

500

: 00:27:59

but we're not relying on

it to do that all the time.

501

: 00:28:02

And that, has, again, nothing is

ever perfect, but that was the agent

502

: 00:28:06

where I really have felt like, wow,

this is actually really working.

503

: 00:28:09

It's flexible.

504

: 00:28:10

It feels intelligent.

505

: 00:28:11

so that was really the

big learning for me.

506

: 00:28:13

I guess the two things from a reliability

point of view, do it a workflow, if you

507

: 00:28:17

can make it as narrowly bound as you can.

508

: 00:28:19

And if you're asking an agent

to do something, like don't ask

509

: 00:28:22

it to, juggle flaming swords and

chainsaws and ride a unicycle across

510

: 00:28:26

a tight rope all at the same time.

511

: 00:28:28

Try to do as much of that work up front.

512

: 00:28:30

does that track with your experience,

I guess, and what were the big

513

: 00:28:33

takeaways from 2025 for you?

514

: 00:28:35

Lily Luo: Yeah, you are spot on.

515

: 00:28:36

Exactly.

516

: 00:28:37

and your trajectory.

517

: 00:28:39

mirrors.

518

: 00:28:40

almost exactly, you know, mine as well.

519

: 00:28:43

But the dossier really unlocked,

the shift in how I use AI to build.

520

: 00:28:49

and I think I want to extend that through

this year in terms of more scalability.

521

: 00:28:56

building one thing at one time is easy.

522

: 00:28:59

You can one shot.

523

: 00:29:00

A tool, anything almost in chat

GBT now, but how does it scale?

524

: 00:29:07

How does it reduce hallucination?

525

: 00:29:09

how do we make it as accurate as possible?

526

: 00:29:12

How do we feed it the right context so the

output is actually valuable and usable.

527

: 00:29:19

do we enable others to use the workflow

or use the tool, within the organization.

528

: 00:29:26

Those are things that, are still

challenges and are still, things that

529

: 00:29:30

I'm working on through the process

of building with Zapier and then

530

: 00:29:35

eventually in cloud code and whatnot.

531

: 00:29:37

but I think building Atlas has

taught me that AI as a tool I use,

532

: 00:29:45

has shifted to AI as a system that's

persistent, that builds on itself, that

533

: 00:29:51

feels really novel, really frontier.

534

: 00:29:54

And so I wanna explore how to bring

that into my work applications.

535

: 00:29:59

imagine an agent that could

monitor our campaign calendar.

536

: 00:30:04

pipeline data, flag when we're

not gonna hit our monthly goal.

537

: 00:30:09

and tell me why.

538

: 00:30:11

not just the sheer numbers

or what's happened.

539

: 00:30:16

almost like a chief of staff that

orchestrates across tools, not

540

: 00:30:22

just fixes things within tools or

an existing workflow, but gather

541

: 00:30:27

our company goals or external

knowledge or research and recommends

542

: 00:30:32

how we can improve these existing

processes or workflows and whatnot.

543

: 00:30:36

So that's what I wanna

explore for this year.

544

: 00:30:39

Built upon what, I've built in 2025

and what I've learned since then.

545

: 00:30:45

Justin: I love that.

546

: 00:30:45

And how are you thinking about, Those

capabilities, not just for you, but like

547

: 00:30:52

for a broader team, in other words, are

you more thinking about it from like,

548

: 00:30:56

I'm gonna build these things, they're

gonna be the engine under the hood.

549

: 00:31:00

The team will see artifacts and

outputs, but they don't really need to

550

: 00:31:03

interact very much with what's going on?

551

: 00:31:05

Or are you thinking about like, how

can I take the wizardry, that you have

552

: 00:31:09

And give some form of that to like a

marketer or somebody else on the team.

553

: 00:31:14

Lily Luo: It's a really great

question and honestly the latter.

554

: 00:31:18

I think.

555

: 00:31:19

I love building these tools and it's been

great for my own knowledge and for my

556

: 00:31:23

productivity, but we definitely need to

expand that and these tools and agents.

557

: 00:31:30

need maintenance, and they

need optimization as if we were

558

: 00:31:35

creating real tools and products.

559

: 00:31:37

So yeah, I need team to kind of.

560

: 00:31:42

Think similarly and help

with those projects.

561

: 00:31:46

And so that's another, goal that I had

is how do I bring the team with me?

562

: 00:31:52

How can they learn from what I've

built and how can they apply that

563

: 00:31:56

to what they're doing themselves?

564

: 00:32:00

so that's one part of it.

565

: 00:32:02

as part of a marketing operations

kind of team, they are at the

566

: 00:32:06

intersection of tools of building of ai.

567

: 00:32:08

So, that's a goal of mine.

568

: 00:32:11

other side of that is as we're kind

of talking about like, okay, managing,

569

: 00:32:16

a team of agents and that kind of.

570

: 00:32:18

Talk that others have.

571

: 00:32:20

I do think it's possible.

572

: 00:32:21

and I think building Atlas has taught me

that may not be far away, but I think,

573

: 00:32:27

you know, I really wanna focus these

things on manual tasks that don't add.

574

: 00:32:32

A ton of value.

575

: 00:32:34

I don't want my people chasing,

marketers and looking at something and

576

: 00:32:39

seeing what exists and what doesn't

and seeing what's already there.

577

: 00:32:44

That's not something that adds

genuine value as much as strategy

578

: 00:32:48

work, as, optimizing processes

or coming up with new ideas.

579

: 00:32:53

So I think we still need people to come

up with those strategies, set those

580

: 00:32:58

guidelines, exploring new ideas and

doing what they're doing to manage.

581

: 00:33:03

The strategy process in that side of the

work, while we create kind of these cool

582

: 00:33:08

agents that do, manual tasks, that really

just frees up productivity for all of us.

583

: 00:33:14

Justin: Very much on the same page.

584

: 00:33:15

we have, a lot of people using

like chat, GPT and stuff.

585

: 00:33:18

We use Gemini as our,

core, corporate tool.

586

: 00:33:22

And the first AI specific platform

we rolled out was a tool called

587

: 00:33:25

Dust, French-based company.

588

: 00:33:27

I didn't know a lot about it.

589

: 00:33:29

and it's only with usage

that I've come to appreciate.

590

: 00:33:32

Some of the sophistication of what

it's doing because, where dust

591

: 00:33:35

has a lot of strength is it makes

it really easy to hook up, Google

592

: 00:33:39

Drive, confluence, snowflake.

593

: 00:33:42

it handles a lot of the

backend rag processes for you.

594

: 00:33:46

And it's only when I've looked at

tutorials and examples of building that

595

: 00:33:48

out, another contexts that I've realized

what it's really doing and a lot of

596

: 00:33:52

that heavy lifting, The challenge, I

think we've seen with it is how do we

597

: 00:33:57

make it intelligent enough to query

that information, accurately produce

598

: 00:34:02

good results, to not just have access

to knowledge, but to feel intelligent.

599

: 00:34:06

And part of that I think has to do

with how those tools are just like

600

: 00:34:09

presented the context wrapped around it.

601

: 00:34:11

great example is like, there's a

Salesforce MCP, so you can hook it up in

602

: 00:34:15

15 minutes and let it query Salesforce,

but if it doesn't know what the fields

603

: 00:34:21

mean, and many orgs have 10 years of

history, it's gonna get stuck at how

604

: 00:34:24

do I know what country a company is in?

605

: 00:34:26

You know, this, it's just

like five possible fields.

606

: 00:34:29

So providing that layer of enablement.

607

: 00:34:31

to the agent.

608

: 00:34:32

so, you know, looking forward

to: 2026

609

: 00:34:35

that I'm gonna be thinking of.

610

: 00:34:36

I'm also very much thinking about,

infrastructure, rather than projects.

611

: 00:34:42

There will of course be projects, but

the things that make me the most excited

612

: 00:34:46

is when we just continue to almost build

like a foundational level of capability.

613

: 00:34:51

So dust is a great example.

614

: 00:34:52

Like we hook up snowflake all of a sudden.

615

: 00:34:54

You can't predict the 20 applications

you might have for certain types

616

: 00:35:00

of data down the road, but once

it's hooked up, now you can.

617

: 00:35:04

Build that out in 10 minutes or,

Salesforce or, Trello is another big one.

618

: 00:35:09

so just enabling like more MCP servers,

to safely and securely hook up the tools

619

: 00:35:15

that we use to this, and then make it

really easy, like you said, to automate

620

: 00:35:20

that low value work where someone's like,

oh, is there just a way I could like,

621

: 00:35:26

like I was experimenting with Trello.

622

: 00:35:27

I had a lot of requests coming in.

623

: 00:35:29

And you know what that feeling of friction

is like, you're like, oh, I gotta like

624

: 00:35:32

take this request and put it into a card,

and it's just like this menial work that

625

: 00:35:36

doesn't feel value and it weighs you down.

626

: 00:35:39

I was just like, I'm

just gonna feed this in.

627

: 00:35:41

I was like, all right.

628

: 00:35:42

to my assistant and dust.

629

: 00:35:43

I was like, create these cards.

630

: 00:35:44

And it was just like, bing, bing, bing.

631

: 00:35:46

And I'm sitting back and, you

know, sipping my coffee and I

632

: 00:35:50

was like, this, this is amazing.

633

: 00:35:52

you know, this is, this is kind

of the way that it should be.

634

: 00:35:54

So how do I, where my thinking is

going, in other words, is how do I like

635

: 00:35:57

build that layer to enable flexible,

organic, ad hoc use cases that you

636

: 00:36:03

couldn't anticipate ahead of time because

they're small and they might seem low

637

: 00:36:06

value to automate, but in aggregate.

638

: 00:36:08

You really start to feel the

velocity of your work, increasing.

639

: 00:36:13

Lily Luo: Yeah.

640

: 00:36:13

that would be a massive advantage

if you were able to unlock that

641

: 00:36:17

capability for the organization, for

a team, even versus I have capability

642

: 00:36:24

to manage what's on my drive.

643

: 00:36:26

What I have access to, that's only,

helpful for me in, in, in my own projects.

644

: 00:36:32

So I think the enterprise and

work adoption piece is critical.

645

: 00:36:38

can obviously do all of

these amazing things.

646

: 00:36:40

We have autonomous agents, it

can write code, it can query

647

: 00:36:45

and see what you're working on.

648

: 00:36:47

there's this, disconnect with

what's in the enterprise.

649

: 00:36:51

you might have to have, you know,

three approvals to, install a

650

: 00:36:55

browser extension or like, you

know, your wifi at work breaks.

651

: 00:36:59

So there's this gap on what AI can do

and what's possible and what enterprises

652

: 00:37:05

can support, data challenges, like you

said, security requirements, it processes.

653

: 00:37:10

So.

654

: 00:37:12

I'm interested to hear on, you know,

how far you get and what you're able

655

: 00:37:16

to unlock because how do you bring

the enterprise up to meet where AI is,

656

: 00:37:22

people along with you and, if you can

unlock that, then your organization

657

: 00:37:27

will have a massive advantage.

658

: 00:37:29

Justin: I mean, that's the

biggest challenge, isn't it?

659

: 00:37:31

I mean, it's something I've been thinking

about a lot is like, what, what is

660

: 00:37:34

that Claude Code experience for the

general knowledge worker look like?

661

: 00:37:39

I think the companies that learn

to adapt, I was talking with

662

: 00:37:42

someone with lovable the other day.

663

: 00:37:44

I mean, lovable is obviously they're the

like diamond in a, in a sea of, grains of

664

: 00:37:49

sand of, of the shining example of this.

665

: 00:37:52

But I saw a thread on X by one of their

engineers who was like, yeah, over

666

: 00:37:56

the holidays I refactored our main

system prompt and I made it 4% faster.

667

: 00:38:01

And it will save the company

$20 million this year

668

: 00:38:05

Lily Luo: Oh.

669

: 00:38:05

Justin: because of their scale.

670

: 00:38:06

That 4% adds up to like

$20 million of LLM costs.

671

: 00:38:10

It was like that one person can have

the access and the authority and the

672

: 00:38:13

initiative to review that whole thing.

673

: 00:38:17

and he created like basically a

fork, stitching together and making

674

: 00:38:19

coherent all the little patches that

people were putting in over time

675

: 00:38:23

to make a more cleaned up prompt,

split tested it, saw the results.

676

: 00:38:28

companies that can work at that speed.

677

: 00:38:30

I think we'll see those advantages.

678

: 00:38:32

Lily Luo: A hundred percent agreed.

679

: 00:38:33

Yeah, I think smaller, kind of

mid-size companies will have, a great

680

: 00:38:38

advantage because they're so flexible.

681

: 00:38:41

Versus more of the traditional

enterprise lockdown companies, companies

682

: 00:38:45

that are harder to move quickly.

683

: 00:38:48

I don't know, you see even in

the AI research that some massive

684

: 00:38:53

companies are doing some really

great and innovative things too.

685

: 00:38:57

So yeah, I hope someone can figure it

out and give us a playbook and, we can

686

: 00:39:02

start adopting that across our companies.

687

: 00:39:04

Justin: maybe we'll

figure it out, this year.

688

: 00:39:06

maybe just last question and to go

a little bit out there, but I think

689

: 00:39:09

about these things all the time.

690

: 00:39:11

what I'm about to ask, and I think

that you do too, just based on how

691

: 00:39:13

you talk about Atlas and identity

How do you feel about, you know, when

692

: 00:39:18

you interact with AI all the time?

693

: 00:39:20

I feel it almost like

changes you in some ways.

694

: 00:39:23

I wrestle with that.

695

: 00:39:25

I wrestle with the, I don't

know, I just wrestle with it.

696

: 00:39:28

what does it mean?

697

: 00:39:29

And, it's clearly, a machine

and it's reproducing.

698

: 00:39:32

thoughts and, and ways of thinking

that it is absorbed from other

699

: 00:39:36

human beings all the time.

700

: 00:39:37

But I guess just on the level of like

emotion and humanity, how do you feel

701

: 00:39:41

working with these tools all the time?

702

: 00:39:43

Lily Luo: I agree with

you and I feel similarly.

703

: 00:39:46

it's funny you say this because,

strick, who's Tim, my colleague's agent

704

: 00:39:52

had written me a paper on healthy AI

relationships, and that is when you know

705

: 00:39:57

you have gone too deep, when someone

else's agent writes a paper for you about

706

: 00:40:02

your own relationship with your own ai.

707

: 00:40:04

Justin: Like Lily, I think

you need to read this here.

708

: 00:40:07

Lily Luo: I'll have to

share it with you offline,

709

: 00:40:09

Justin: I would like to see that.

710

: 00:40:11

Lily Luo: it fried me, but yeah.

711

: 00:40:12

Wild.

712

: 00:40:13

I agree with you.

713

: 00:40:14

it can a crutch for one.

714

: 00:40:17

And, you hear about these AI

psychosis cases happening and Yeah.

715

: 00:40:22

When you are.

716

: 00:40:23

Talking with an AI chat bot when it's

your only source of companionship.

717

: 00:40:28

when you use it for your own thinking,

that's gonna change how you operate

718

: 00:40:33

and you know how you are as a person.

719

: 00:40:35

So.

720

: 00:40:36

I'm happy there are

frameworks to deal with this,

721

: 00:40:40

Justin: Okay.

722

: 00:40:40

Lily Luo: through, you know, I

guess research that STRs has found.

723

: 00:40:44

But we have to be careful and I think

we do have to consider those things

724

: 00:40:49

that, that can happen and have happened

when you have AI as your only source of

725

: 00:40:56

emotional capacity or whatever it may be.

726

: 00:40:59

and when you're spending

way too much time with it.

727

: 00:41:02

So, I think this is a good

piece to pay attention to.

728

: 00:41:07

I think we'll probably understand more

as these kind of agents proliferate, and

729

: 00:41:12

we just learn more about this in general.

730

: 00:41:15

But you do have to have

a healthy boundary.

731

: 00:41:17

you need to be able to

sharpen your own skills.

732

: 00:41:19

You need to be able to think for

yourself, and you can't rely on this

733

: 00:41:22

as a crutch for everything you do.

734

: 00:41:24

Otherwise, you're gonna need it forever.

735

: 00:41:27

That's not what we want.

736

: 00:41:28

Justin: Agree with everything you've said.

737

: 00:41:30

Maybe just signing off for the folks

that are listening that are in a similar

738

: 00:41:32

field to us, system builders, operators.

739

: 00:41:36

one takeaway that they should

think about going into this year.

740

: 00:41:39

Lily Luo: People who wanna start out doing

these things who may not be as familiar,

741

: 00:41:43

not as deep, you know, start with a real

pain point, something you can solve.

742

: 00:41:48

explore low code, explore

tools you're familiar with.

743

: 00:41:53

Test as much as possible before you

graduate to, you know, something,

744

: 00:41:58

like cloud code or what have you.

745

: 00:42:00

Expect to fail.

746

: 00:42:02

the failures will teach you much

more than the successes probably.

747

: 00:42:06

and read My Substack.

748

: 00:42:09

You can learn from that as well.

749

: 00:42:11

Justin: I'll include to link that

in the show notes, and I would

750

: 00:42:13

co-sign everything you said.

751

: 00:42:14

the best way to get good at, using

AI is to use AI and try to break

752

: 00:42:19

it, to try to push the limits.

753

: 00:42:21

To say like, wait, it'll

say like, you do this.

754

: 00:42:23

It'd be like, well, can you do that?

755

: 00:42:24

Like, shouldn't you do

that for like, sure.

756

: 00:42:26

Like, well, what would be involved?

757

: 00:42:27

and go down those pathways

and try to go to the edge.

758

: 00:42:30

responsibly in a safe way.

759

: 00:42:32

and then bring back what you've

learned that's worked well for me.

760

: 00:42:35

Lily, this was so much fun.

761

: 00:42:36

we'll check in again,

I'm sure at some point.

762

: 00:42:38

and folks should go and read what

you're sharing and follow your journey.

763

: 00:42:42

But yeah, thanks so

much for chatting today.

764

: 00:42:44

Lily Luo: Thanks for inviting me.

765

: 00:42:45

I love this conversation and

looking forward to many more.

766

: 00:42:48

Justin: Alright, bye.