1
00:00:03,144 --> 00:00:05,214
Leo Dion (host): Welcome to
another episode of Empower Apps.

2
00:00:05,284 --> 00:00:06,954
I'm your host, Leo Dion.

3
00:00:07,054 --> 00:00:08,994
Today, I'm joined by Joe Fabisevich.

4
00:00:10,154 --> 00:00:12,194
Joe, thank you for coming back on the show

5
00:00:13,240 --> 00:00:13,440
Joe Fabisevich (guest): Yeah.

6
00:00:13,470 --> 00:00:14,320
Thank you for having me.

7
00:00:14,400 --> 00:00:14,830
I appreciate it

8
00:00:15,448 --> 00:00:18,518
Leo Dion (host): This is about your
third time on the show, and a lot has

9
00:00:18,518 --> 00:00:21,068
changed since you were last on for Plinky.

10
00:00:21,558 --> 00:00:25,288
But we'll go ahead and let you introduce
yourself for those who don't know

11
00:00:26,804 --> 00:00:27,054
Joe Fabisevich (guest): Yeah.

12
00:00:27,074 --> 00:00:29,054
So, I'm an iOS developer by trade.

13
00:00:29,094 --> 00:00:30,774
I work on an indie app called Plinky.

14
00:00:31,124 --> 00:00:33,494
It's the easiest way to
save links for later.

15
00:00:33,594 --> 00:00:36,824
It basically is just like Pocket
or Raindrop or any of those other

16
00:00:36,824 --> 00:00:38,224
apps, but with a different take.

17
00:00:38,704 --> 00:00:42,984
I also have been teaching AI for the
last year and a half, so I've taught at

18
00:00:42,984 --> 00:00:47,684
a nonprofit, and I also run workshops
where people bring an idea that they

19
00:00:47,684 --> 00:00:49,114
want to build, like a real idea.

20
00:00:49,114 --> 00:00:52,824
It could be an app, it could be a
prototype, it could be, like, an agent

21
00:00:52,824 --> 00:00:58,104
or something, and then they walk out
after three hours with that in hand,

22
00:00:58,124 --> 00:01:01,594
but they've learned a whole lu- bunch
of lessons about AI along the way.

23
00:01:02,014 --> 00:01:04,914
And then, you know, for the next couple
weeks, I get messages from them saying,

24
00:01:04,914 --> 00:01:08,304
like, "I built this thing that I've wanted
for years," or, "My wife has wanted this

25
00:01:08,304 --> 00:01:09,864
idea, and I was able to deliver it."

26
00:01:09,894 --> 00:01:13,174
And it's just, like, a really cool
experience to see people learn a

27
00:01:13,174 --> 00:01:16,784
whole lot in a short period of time,
and then yeah, like enable them

28
00:01:16,784 --> 00:01:19,024
to continue building upon that.

29
00:01:20,650 --> 00:01:22,580
Leo Dion (host): So
let's get into it today.

30
00:01:22,580 --> 00:01:26,400
The focus of our episode is AI
and AI-assisted development.

31
00:01:26,940 --> 00:01:31,480
How has your life changed in, as far
as a software developer in the last few

32
00:01:31,480 --> 00:01:34,280
years with the rise of these AI tools?

33
00:01:35,606 --> 00:01:37,416
Joe Fabisevich (guest): Yeah, I mean,
how's my life changed is the big

34
00:01:37,416 --> 00:01:42,076
question, but specifically in software,
I'd say it's not to say that, like, I

35
00:01:42,126 --> 00:01:45,116
don't think about code, I don't do code.

36
00:01:45,206 --> 00:01:47,006
I just don't write the code manually.

37
00:01:47,006 --> 00:01:52,566
Like 95 to 99% of my code is AI
generated, but I am constantly

38
00:01:52,566 --> 00:01:54,436
shaping the code that comes out.

39
00:01:54,506 --> 00:01:59,066
Which is to say that, like, I care
as deeply about the code that I

40
00:01:59,076 --> 00:02:05,586
ship to production today as I did
before AI, but my ability to ship

41
00:02:05,616 --> 00:02:07,096
has become much more ambitious.

42
00:02:07,146 --> 00:02:11,396
So it's not an exaggeration to say that,
like, through the process of instead of

43
00:02:11,396 --> 00:02:16,896
sitting at a computer and almost, like,
drawing on a blank canvas the way you

44
00:02:16,896 --> 00:02:20,806
would, like, one stroke at a time for a
painting, I'm very much often just, like,

45
00:02:21,376 --> 00:02:25,796
molding the code into what I want it to
be, and that process is genuinely faster.

46
00:02:25,846 --> 00:02:30,746
It can feel more exhausting at times, but
the payoff is that I do a day's worth of

47
00:02:30,746 --> 00:02:34,776
work and I get a week's worth of output
that looks like the code that I want.

48
00:02:34,836 --> 00:02:37,706
And especially as the models
continue to get better and as

49
00:02:37,706 --> 00:02:42,146
I build up infrastructure in my
projects that the code that I want

50
00:02:42,146 --> 00:02:44,916
it to produce, the results just
keep getting better and better.

51
00:02:44,916 --> 00:02:51,086
And so I am still thinking as
deeply about code as I was before.

52
00:02:51,336 --> 00:02:54,236
And in fact, like, I just went
through, like, a bunch of interviews

53
00:02:54,236 --> 00:02:59,136
for, you know, potential jobs and
I had to code manually, right?

54
00:02:59,136 --> 00:03:02,326
And so I was very worried about, like,
my skills atrophying or deteriorating.

55
00:03:02,766 --> 00:03:07,266
And it turns out that no, as a person
who is a really good engineer, and B,

56
00:03:07,446 --> 00:03:12,006
is thinking about code all the time,
my ability to type out that code and

57
00:03:12,016 --> 00:03:16,236
think about the code the way that I want
it to has stayed the same as it was.

58
00:03:16,236 --> 00:03:19,316
And honestly I would say that somewhere
around six months ago I started to

59
00:03:19,316 --> 00:03:23,576
get to the point with working with
AI models where I'm learning more

60
00:03:23,916 --> 00:03:27,226
by continuing to dive in and asking
it, like, "Why did you do this?

61
00:03:27,236 --> 00:03:30,956
How did you do that?" Sometimes the
answer is, "Oh, that's a bad idea."

62
00:03:30,996 --> 00:03:36,076
And honestly, as a developer, I'm able
to identify that 90, 95% of the time.

63
00:03:36,566 --> 00:03:40,736
But every once in a while, like, it'll
surprise me and I'll be like, "Ah, that is

64
00:03:40,826 --> 00:03:45,786
actually a better pattern than I would've
come up with." And, you know, the, like,

65
00:03:46,396 --> 00:03:49,756
the closed-minded way of looking at that
is to say, "Ugh, like, you know, the AI

66
00:03:49,756 --> 00:03:52,866
is better than me." But the open-minded
way is to say like, "Hey, I have an

67
00:03:52,956 --> 00:03:55,266
opportunity to learn from AI here."

68
00:03:55,576 --> 00:03:59,306
And I continue to think that there's
just, like, a huge amount of educational

69
00:03:59,306 --> 00:04:02,016
value in working with AI with an open mind

70
00:04:03,270 --> 00:04:05,440
Leo Dion (host): I really like that,
yeah, 'cause I'm in the same boat

71
00:04:05,460 --> 00:04:10,870
where I've done some interviews where
I've had to code, and yeah, I was

72
00:04:11,190 --> 00:04:14,240
impressed with myself that I actually
knew what the heck I was doing.

73
00:04:14,500 --> 00:04:18,000
I didn't get the syntax all right
'cause auto-complete wasn't great

74
00:04:18,010 --> 00:04:23,000
and stuff, but like, as far as, like,
the ideas, yes, I, I understood it.

75
00:04:23,450 --> 00:04:30,070
I think, so there's a lot to unfold
there, and I think I wanna start with

76
00:04:30,150 --> 00:04:34,080
the misconceptions people have with
software development with AI, and one

77
00:04:34,080 --> 00:04:39,290
of the things you talked about, and
I'll just call it work slop, right?

78
00:04:39,830 --> 00:04:47,350
This idea of I think a l- lot of people
want to extract economic value from

79
00:04:47,350 --> 00:04:54,380
it, and so they're willing to take
shortcuts and use AI without understanding

80
00:04:54,380 --> 00:04:56,120
what the heck is going on underneath.

81
00:04:56,560 --> 00:05:02,800
Do you see that problem a lot in
the industry right now and possibly

82
00:05:02,800 --> 00:05:04,570
a problem in the near future?

83
00:05:05,868 --> 00:05:06,148
Joe Fabisevich (guest): Yeah.

84
00:05:06,148 --> 00:05:08,068
So I mean, I think work
slop represents two things.

85
00:05:08,078 --> 00:05:12,198
One is the desire to take
shortcuts, as you put it, but

86
00:05:12,198 --> 00:05:15,218
it's like to prioritize short-term
returns over long-term benefits.

87
00:05:15,498 --> 00:05:19,908
And so even when you worked at
companies that before… in a world

88
00:05:19,908 --> 00:05:25,118
before AI, you would see people would
ship crap because they knew that

89
00:05:25,128 --> 00:05:30,038
they had to make a deadline or they
had pressure from a PM or something.

90
00:05:30,328 --> 00:05:33,578
And so human nature hasn't
changed in that regard.

91
00:05:33,578 --> 00:05:36,858
We've just kind of accelerated the
ability to produce the output that

92
00:05:36,858 --> 00:05:38,778
generates the returns or the results.

93
00:05:39,078 --> 00:05:42,818
So I have always made the argument
that, like, we should be prioritizing

94
00:05:43,118 --> 00:05:46,458
or balancing really the short term
and the long term, but prioritizing

95
00:05:46,458 --> 00:05:47,938
our own long-term success.

96
00:05:48,078 --> 00:05:49,608
AI doesn't change that equation.

97
00:05:49,608 --> 00:05:54,868
I think it's really important that when
I use AI to generate code, I go and I

98
00:05:54,918 --> 00:05:58,998
actually mold it and make sure, like, if
I'm building a platform or I'm building

99
00:05:58,998 --> 00:06:03,298
something that is going to stand the test
of time or I want it to stand the test of

100
00:06:03,298 --> 00:06:07,788
time or to serve many users or to scale
up, you're not gonna vibe code your way

101
00:06:07,788 --> 00:06:11,468
to the next Google, but you can vibe code
your way to, like, your own personal app.

102
00:06:11,788 --> 00:06:16,138
And I think that we should be encouraging
that creativity and human flourishing

103
00:06:16,148 --> 00:06:20,028
where people really do wanna see, like,
the ideas in their head come to life.

104
00:06:20,628 --> 00:06:23,818
But as a professional software
engineer, I think that it's still on

105
00:06:23,818 --> 00:06:26,918
you to say the thing that we ship to
production, the thing that we ship to

106
00:06:26,918 --> 00:06:29,778
customers, in the same way that you
care about user experience and user

107
00:06:29,778 --> 00:06:32,538
design, this matters just as much.

108
00:06:33,568 --> 00:06:40,568
The other thing that I see with work
slop is that people are just, you

109
00:06:40,568 --> 00:06:45,518
know, they have different incentive
structures that something is pushing

110
00:06:45,518 --> 00:06:50,454
them internally to say, "This doesn't
matter." And I think it's kind of

111
00:06:50,454 --> 00:06:54,794
unearthing a real thing in the world,
which is a lot of the work that we do or

112
00:06:54,804 --> 00:06:58,954
have done is not economically important.

113
00:06:59,324 --> 00:07:01,684
It is just the way that we've
always proved that we're doing

114
00:07:01,694 --> 00:07:03,294
economically valuable things.

115
00:07:03,464 --> 00:07:08,894
And so, you know, a PRD is valuable if
you're working on a complex feature,

116
00:07:09,284 --> 00:07:14,034
but an AI-generated PRD that is slop is
not valuable, which means that the shape

117
00:07:14,034 --> 00:07:16,534
of the PRD wasn't the valuable thing.

118
00:07:17,134 --> 00:07:20,754
It's that the fact that you spent the
time thinking about it, documenting

119
00:07:20,754 --> 00:07:22,524
it, and communicating it was the value.

120
00:07:22,914 --> 00:07:28,764
And so we oftentimes create
proxies for what's important by

121
00:07:28,764 --> 00:07:30,144
the way that it looks in the end.

122
00:07:30,504 --> 00:07:34,614
But, you know, when someone writes
a book, they spend years writing a

123
00:07:34,614 --> 00:07:36,524
book that someone reads in six hours.

124
00:07:37,504 --> 00:07:41,114
That means that years of work went
into putting together something

125
00:07:41,114 --> 00:07:43,014
that was digestible in six hours.

126
00:07:43,034 --> 00:07:44,884
It's not six hours worth of knowledge.

127
00:07:44,884 --> 00:07:46,234
It's just the way that we consume it.

128
00:07:46,514 --> 00:07:53,084
And so when people make work slop, they
are oftentimes equating it as one-to-one.

129
00:07:53,084 --> 00:07:55,594
Like, the thing that matters
is that we have a PRD.

130
00:07:55,804 --> 00:07:59,204
No, the thing that matters is that
you put the thought into this that is

131
00:07:59,204 --> 00:08:02,094
required to ship something that is good

132
00:08:03,542 --> 00:08:03,842
Leo Dion (host): Yeah.

133
00:08:03,852 --> 00:08:06,452
So there's a couple of things I
was thinking when you said that.

134
00:08:07,112 --> 00:08:12,682
One of the parallels I see a lot is,
like, the way businesses use outsourcing

135
00:08:13,392 --> 00:08:20,552
where it's not the quality of the
workers, it's we just, we saved money

136
00:08:20,552 --> 00:08:25,492
by hiring cheap, unskilled engineers
and not actually vetting them or

137
00:08:26,112 --> 00:08:27,732
you know, doing the right work.

138
00:08:27,732 --> 00:08:31,222
It's kind of like the whole nobody
got fired if they hired IBM kind

139
00:08:31,222 --> 00:08:34,422
of thing, where it's like, yeah,
the work might be crap, but, like,

140
00:08:34,442 --> 00:08:36,202
we did what everybody else does.

141
00:08:36,512 --> 00:08:42,982
And also, like, there is a lot … I'm a
remote worker, but I've been in offices,

142
00:08:43,032 --> 00:08:50,242
and there's a lot of ceremony in working
at an office that is unnecessary and just

143
00:08:50,412 --> 00:08:54,942
it's what you do 'cause your boss tells
you what to do, and I see a lot of that.

144
00:08:55,202 --> 00:08:56,832
So, what was the other thing?

145
00:08:56,842 --> 00:08:58,272
There was something else I was gonna say.

146
00:08:59,012 --> 00:09:00,012
No, I think I covered it.

147
00:09:00,062 --> 00:09:04,942
But, like, o- one of the things in my
talks on vibe coding that I talk about

148
00:09:04,962 --> 00:09:10,262
is, like, the human in the loop idea, and
the idea being, like, it comes from the

149
00:09:10,262 --> 00:09:16,702
military, but the idea being, like, you
have a human somewhere along the AI chain

150
00:09:17,162 --> 00:09:21,842
that is looking at the code, reviewing
the PR, like, that kind of stuff.

151
00:09:22,382 --> 00:09:26,152
And yeah, I mean, that stuff
is still really important.

152
00:09:26,272 --> 00:09:32,522
Where, like, what are some ways
that can help a team not get into a

153
00:09:32,522 --> 00:09:35,922
rut of just depending on AI without
actually putting the work into it?

154
00:09:36,302 --> 00:09:39,592
'Cause, like, you could
do an AI-built PRD.

155
00:09:39,602 --> 00:09:42,182
Like, I've done that, but
then I, of course, I look

156
00:09:42,182 --> 00:09:43,482
at it and I double-check it.

157
00:09:43,482 --> 00:09:48,022
I don't just go, "Good enough." What
are some ways to make sure that, like,

158
00:09:49,192 --> 00:09:53,512
you're not just being lazy and just
depending on the AI to do everything for

159
00:09:53,512 --> 00:09:55,182
you without actually checking the work?

160
00:09:56,354 --> 00:09:56,634
Joe Fabisevich (guest): Yeah.

161
00:09:56,634 --> 00:09:59,794
So I mean, the first thing I'll touch on
is you mentioned this kind of like human

162
00:09:59,794 --> 00:10:03,824
in the loop, and I think what we notice
is that AI continues to push the human

163
00:10:03,824 --> 00:10:05,484
in the loop further and further back.

164
00:10:05,564 --> 00:10:07,324
Like, where in the
process do we need to be?

165
00:10:07,554 --> 00:10:08,774
Or it changes in general.

166
00:10:08,784 --> 00:10:11,424
So like I like to be
very involved upfront.

167
00:10:11,464 --> 00:10:15,004
I like to plan meticulously, and
I don't mean by like creating a

168
00:10:15,004 --> 00:10:16,624
planning doc that's AI generated.

169
00:10:16,924 --> 00:10:20,284
I mean, I have like a skill where I write
up like a, you know, I spend like 30

170
00:10:20,284 --> 00:10:23,254
minutes kinda like building the context
for the AI about the feature that I'm

171
00:10:23,254 --> 00:10:27,864
trying to build, and then I work with
it to like ask clarifying questions, and

172
00:10:27,864 --> 00:10:30,474
this is kind of like a relentless process.

173
00:10:30,684 --> 00:10:34,494
And it doesn't generate like a plan
per se, it just provides all the

174
00:10:34,494 --> 00:10:37,434
context that it needs to make sure
that all the ambiguities are resolved.

175
00:10:37,734 --> 00:10:44,834
And then by doing that, I then get like,
you know, a result that is 90, 95, 99%

176
00:10:44,834 --> 00:10:46,004
of the way to what I want it to be.

177
00:10:46,304 --> 00:10:49,064
But I spent like one to two,
maybe three hours doing that.

178
00:10:49,144 --> 00:10:53,834
And so when I said like, "Oh, the
purpose of that PRD is to communicate

179
00:10:53,834 --> 00:10:57,474
the value," it's not the creation of
the PRD, it's that the reason you can

180
00:10:57,474 --> 00:11:02,114
automate away a large chunk of work
is because you spent a lot of time

181
00:11:02,154 --> 00:11:03,864
actually thinking through the problem.

182
00:11:04,254 --> 00:11:08,434
And then, of course, let's say it only
gets you 90% of the way, it's still your

183
00:11:08,444 --> 00:11:10,644
job to figure out what the last 10% is.

184
00:11:10,914 --> 00:11:14,284
And so there's this idea called
Amdahl's law, which is when…

185
00:11:15,194 --> 00:11:19,284
Let's say that something is like
90% parallelizable and 10% serial.

186
00:11:19,394 --> 00:11:22,154
Let's call it being humans in the
loop as the serial part of it.

187
00:11:22,604 --> 00:11:25,894
You can't just like get 100% return.

188
00:11:26,224 --> 00:11:27,944
You're going to have a
bottleneck somewhere.

189
00:11:28,184 --> 00:11:32,564
And so the second part of the question is
like how do teams kind of like work with

190
00:11:32,564 --> 00:11:37,084
this, is to remember that certain tasks
are automatable, but that doesn't mean

191
00:11:37,084 --> 00:11:40,444
that the entire process is automatable
just because the tasks are automatable.

192
00:11:40,694 --> 00:11:44,274
And so what you should be doing is looking
and saying like, "What can we automate?

193
00:11:44,304 --> 00:11:48,144
What do we still need humans for? What do
we think in six months, in a year we will

194
00:11:48,144 --> 00:11:53,194
not need humans for, and how do we kind
of race towards that?" There's this idea

195
00:11:53,194 --> 00:11:56,634
that like you should be skating to where
the puck is going in terms of the models.

196
00:11:56,934 --> 00:12:01,204
And so even just looking three months
out, I oftentimes am like, "Oh, like

197
00:12:01,204 --> 00:12:05,194
this idea may not work today, but I'll
keep the idea in my pocket or I'll pr-

198
00:12:05,244 --> 00:12:08,654
you know, keep the initial prompt,"
like I'll write a big prompt for it, and

199
00:12:08,654 --> 00:12:12,204
then try it again in three months or six
months and then see, yep, that worked.

200
00:12:12,374 --> 00:12:17,264
And so part of this is teams need to,
one, understand what models can and

201
00:12:17,264 --> 00:12:24,614
cannot do, two, plan for when they can
do that, and then three, just continue

202
00:12:24,614 --> 00:12:26,114
to use your best judgment, right?

203
00:12:26,114 --> 00:12:31,024
It's like whether a human is in the loop
or not, it's ultimately your job to figure

204
00:12:31,024 --> 00:12:34,644
out is this going to be value additive.

205
00:12:34,674 --> 00:12:38,834
I don't think that models are there
yet for kind of determining what is

206
00:12:38,874 --> 00:12:43,154
uniquely useful to you, even though
sometimes models can feel like uniquely

207
00:12:43,154 --> 00:12:45,034
scary at almost how well they know you.

208
00:12:45,782 --> 00:12:46,622
Leo Dion (host): Right, right.

209
00:12:46,812 --> 00:12:48,502
So let's get into that.

210
00:12:48,562 --> 00:12:55,232
Like, as of today, what… I don't
know if I should ask what are models

211
00:12:55,232 --> 00:12:59,302
good at first, or maybe just get into
where do they fall short as of today?

212
00:13:00,302 --> 00:13:02,892
Joe Fabisevich (guest): I think the
best way I put it is h- models are

213
00:13:02,892 --> 00:13:06,202
very good at verifiable tasks, and that
means that's where they fall short is

214
00:13:06,202 --> 00:13:09,432
where tasks are unverifiable, and they
succeed where tasks are verifiable.

215
00:13:09,732 --> 00:13:15,282
So if you ask models to be creative, they
can do their best guess approximation

216
00:13:15,282 --> 00:13:18,182
at being creative, which is often
more creative than many people are.

217
00:13:18,492 --> 00:13:24,722
But writing a good essay
is not a verifiable task.

218
00:13:24,722 --> 00:13:28,132
Like, I tried to have AI generate
some documentation for an open

219
00:13:28,132 --> 00:13:32,062
source library I'm working on,
and I just found it was crap

220
00:13:32,082 --> 00:13:33,772
Leo Dion (host): What
specifically was crap about it?

221
00:13:35,160 --> 00:13:36,840
Joe Fabisevich (guest): I think the main
thing is it didn't really understand why

222
00:13:36,850 --> 00:13:38,350
it was writing the documentation, right?

223
00:13:38,390 --> 00:13:41,930
It was like, "This is what documentation
is shaped like, so I'm going to write

224
00:13:41,930 --> 00:13:44,520
documentation." On first glance, I was
like, "Wow, this is really impressive."

225
00:13:44,840 --> 00:13:48,130
And then I started to get into it,
and it started talking about how I

226
00:13:48,150 --> 00:13:52,560
implemented testing in this library,
and I got better results when I said,

227
00:13:52,910 --> 00:13:56,890
"You should actually just go read the
tests as opposed to guessing what tests

228
00:13:56,900 --> 00:13:58,580
should look like," if that makes sense

229
00:13:59,374 --> 00:14:02,844
Leo Dion (host): or like it's very
much I find like my READMEs can be

230
00:14:02,844 --> 00:14:08,184
very salesy when it's like, "No, this
is an open source library. You don't

231
00:14:08,184 --> 00:14:12,154
need to tell me what's so much better
about my library than the other way

232
00:14:12,154 --> 00:14:14,394
to do it." It's like, stop doing that.

233
00:14:14,434 --> 00:14:16,854
And that's what it's trained
on, so that's what it does

234
00:14:17,730 --> 00:14:19,320
Joe Fabisevich (guest): Yeah, and
I actually kind of to the note of

235
00:14:19,320 --> 00:14:23,580
training on is I found better success,
not perfect by any means, but I found

236
00:14:23,580 --> 00:14:27,520
better success by I pointed AI at all
of my writing on build.ms, and then

237
00:14:27,520 --> 00:14:32,430
I also pointed it at readmes from
my open source libraries where I've,

238
00:14:32,800 --> 00:14:34,690
you know, handcrafted meticulously

239
00:14:34,786 --> 00:14:35,066
Leo Dion (host): Yeah.

240
00:14:35,146 --> 00:14:38,316
Joe Fabisevich (guest): And said like,
"Basically copy this, but go do this."

241
00:14:38,476 --> 00:14:42,676
And the parts that it copied were,
like mechanically were great, and the

242
00:14:42,676 --> 00:14:46,746
parts where I tried to like translate
what worked in Boutique into what

243
00:14:46,826 --> 00:14:50,246
Broadcast should be like Broadcast
the name of the library, and I was

244
00:14:50,246 --> 00:14:53,176
like, "Oh, this kind of sucks," right?

245
00:14:53,176 --> 00:14:55,706
I'm gonna have to go in
there and do this myself.

246
00:14:55,906 --> 00:15:00,296
And so this kind of gets at what I mean
by verifiable tasks are automatable

247
00:15:00,306 --> 00:15:03,826
or are models are good at because
that's what they can automate.

248
00:15:04,126 --> 00:15:09,366
So coding is like the canonical
example of a task that is verifiable.

249
00:15:09,376 --> 00:15:10,666
You have a compiler in the loop.

250
00:15:10,686 --> 00:15:11,456
You can run it.

251
00:15:11,736 --> 00:15:15,256
You can… and we can talk about
this later, but like you can write

252
00:15:15,296 --> 00:15:19,716
enough logs that show the entire
journey that a user goes down through

253
00:15:19,736 --> 00:15:21,286
combinations of action and state.

254
00:15:21,666 --> 00:15:25,306
The more observability you can add,
you know, that's why people like tests,

255
00:15:25,316 --> 00:15:28,036
like AI is good at writing tests,
but when it doesn't know what tests

256
00:15:28,066 --> 00:15:32,776
to write, it writes mediocre tests
because the question of what test

257
00:15:32,786 --> 00:15:37,576
should you write is not verifiable, but
did a test pass is a verifiable task.

258
00:15:37,966 --> 00:15:40,926
And so models are really good
at verifiable tasks, and they're

259
00:15:40,936 --> 00:15:46,466
really not there yet without human
judgment or human intuition or, you

260
00:15:46,466 --> 00:15:50,686
know, a human in the loop at things
that still require verification.

261
00:15:51,790 --> 00:15:53,760
Leo Dion (host): Yeah, I think
what you're touching upon too is

262
00:15:53,760 --> 00:16:03,920
like how AI is not great at like
video or art or like writing books.

263
00:16:04,490 --> 00:16:06,920
Especially the video, like
especially visual stuff.

264
00:16:07,410 --> 00:16:09,600
It seems like that's
where it falls short and

265
00:16:09,660 --> 00:16:14,060
i'm not pessimistic that AI is gonna
take over and like r- do every,

266
00:16:14,340 --> 00:16:16,220
all the video work or all the art.

267
00:16:16,510 --> 00:16:20,750
Like I think there's a place for it when
it comes to like, oh, we just need you

268
00:16:20,750 --> 00:16:24,240
to generate a bunch of people in the
background maybe or something like that.

269
00:16:24,680 --> 00:16:25,920
But like if it's

270
00:16:26,212 --> 00:16:28,252
Joe Fabisevich (guest): really good at,
like, the mechanical stuff though, right?

271
00:16:28,252 --> 00:16:33,342
'Cause it's like, one, creating video
takes forever, and two, And that

272
00:16:33,342 --> 00:16:34,772
is, like, a task that it can do.

273
00:16:35,032 --> 00:16:40,892
But if you start to break down videos
into frames, then you may be able to get

274
00:16:40,912 --> 00:16:44,282
things that are actually tangibly useful.

275
00:16:44,692 --> 00:16:49,042
So, like, the more that you can break
down a problem, code, for example, is,

276
00:16:49,232 --> 00:16:53,752
like, like I said, it's the perfect
example because every character

277
00:16:53,772 --> 00:16:58,292
is well thought out and checked
by grammar, that if you can get…

278
00:16:58,632 --> 00:17:01,582
I think we're probably gonna end
up designing new systems that

279
00:17:01,602 --> 00:17:03,262
are meant to be automatable.

280
00:17:03,282 --> 00:17:06,922
So yes, video generation in the way that
you think about it today doesn't make

281
00:17:06,922 --> 00:17:11,852
sense, but you may actually have an LLM
that is trying to follow certain video

282
00:17:11,852 --> 00:17:16,802
generation rules that are well-defined
in a grammar, and then you can have

283
00:17:16,802 --> 00:17:20,612
an LLM that's smarter than it continue
to check, like a, they call it a model

284
00:17:20,622 --> 00:17:23,972
judge, to make sure that it's following
the grammar as you would expect.

285
00:17:24,262 --> 00:17:30,272
And so while I am kind of saying that,
like, anywhere where creativity is still

286
00:17:30,292 --> 00:17:37,182
needed is a place where LLMs can't really
fully solve the problem, we may cut down

287
00:17:37,302 --> 00:17:41,362
drastically because we may be able to
invent new ways of rec- representing

288
00:17:41,372 --> 00:17:43,312
what is the thing that we're going for.

289
00:17:44,254 --> 00:17:45,914
Leo Dion (host): Yeah, I
don't think we disagree.

290
00:17:46,104 --> 00:17:50,194
I think, like, I do think, like, video,
we think too much video generation.

291
00:17:50,204 --> 00:17:53,574
We just think, you know, videos
where grandma has six ha-

292
00:17:53,764 --> 00:17:55,104
fingers and stuff like that.

293
00:17:55,124 --> 00:17:59,404
But I'm thinking more like it's part,
it'll be part of the process of making

294
00:17:59,404 --> 00:18:02,794
a video, not like, "Oh, we're gonna
make an entire video with a prompt."

295
00:18:02,834 --> 00:18:06,934
Like, I just don't see that
happening anytime in my lifetime

296
00:18:07,078 --> 00:18:08,328
Joe Fabisevich (guest): that's
where I think we disagree

297
00:18:08,944 --> 00:18:09,604
Leo Dion (host): Oh, okay.

298
00:18:10,204 --> 00:18:13,954
So you think you can make a whole
movie from a prompt at some point?

299
00:18:14,020 --> 00:18:15,500
Joe Fabisevich (guest): I think
you'll be able to make whole movies

300
00:18:15,510 --> 00:18:16,830
by prompting, if that makes sense.

301
00:18:16,860 --> 00:18:21,650
The difference being that, like,
you don't have to make video- movies

302
00:18:22,620 --> 00:18:25,070
120 minutes at a time, you know.

303
00:18:25,140 --> 00:18:28,200
In the same way that I'm not
actually suggesting that professional

304
00:18:28,200 --> 00:18:32,180
software engineers should write
one prompt and expect an app,

305
00:18:32,840 --> 00:18:33,040
Leo Dion (host): Okay

306
00:18:33,070 --> 00:18:35,830
Joe Fabisevich (guest): may write
series of prompts and then end up

307
00:18:35,830 --> 00:18:39,150
with a movie that is, I don't know
what the right term is, but like

308
00:18:39,150 --> 00:18:41,500
satiable or like enjoyable to people.

309
00:18:41,830 --> 00:18:46,240
Because I also kind of think that there
is this dichotomy that we don't really

310
00:18:46,240 --> 00:18:51,240
talk about, but like what do consumers
enjoy versus what is professional

311
00:18:51,240 --> 00:18:53,570
grade output is very different.

312
00:18:53,590 --> 00:18:58,190
Like, I will admit that I think AI…
I hate AI video as a general concept.

313
00:18:58,210 --> 00:19:01,390
Like, it's just like it… I'm also…
I've never been a video person, which

314
00:19:01,390 --> 00:19:02,680
is weird being on video right now.

315
00:19:02,910 --> 00:19:05,490
But I just doesn't… it just
doesn't do it for me is probably

316
00:19:05,490 --> 00:19:06,200
the best way I can put it.

317
00:19:06,430 --> 00:19:10,040
But at the same time, there have been
like two AI videos that I can recall

318
00:19:10,040 --> 00:19:14,960
where I was like, "Damn, that was funny
and like, I actually enjoyed this."

319
00:19:14,970 --> 00:19:20,000
And how much of a human was in the loop
is questionable, but I suspect that I

320
00:19:20,010 --> 00:19:24,200
think those videos were like 60 seconds
long, and there was a human riffing

321
00:19:24,200 --> 00:19:25,660
on it basically at some point, right?

322
00:19:25,670 --> 00:19:29,600
Like, they came up with the silly
idea or whatever, but they were not

323
00:19:30,110 --> 00:19:34,270
video producing it in any way, and
I think that will continue to expand

324
00:19:35,102 --> 00:19:35,442
Leo Dion (host): Yeah.

325
00:19:35,552 --> 00:19:35,902
Okay.

326
00:19:36,642 --> 00:19:42,602
So we had t-talked about what are the
good parts, what are the bad parts.

327
00:19:43,242 --> 00:19:49,822
Like, as far as you're teaching
folks right now what are some of the

328
00:19:49,832 --> 00:19:55,922
biggest misconceptions and also the
big, like, aha moments that they were

329
00:19:55,922 --> 00:19:57,692
like, "Oh my gosh, this is amazing"?

330
00:19:59,206 --> 00:20:00,136
Joe Fabisevich (guest): Oh,
yeah, that's a good question.

331
00:20:00,176 --> 00:20:02,046
I mean, I think the biggest
misconception is that everyone thinks

332
00:20:02,046 --> 00:20:05,846
they're good at prompting, and I
think that prompting is a skill.

333
00:20:05,886 --> 00:20:10,006
It's like a, it's a skill of context
building, and I think that people

334
00:20:10,496 --> 00:20:13,746
oftentimes have an idea in their head
of what they want, and they are not

335
00:20:13,986 --> 00:20:15,686
particularly good at communicating it.

336
00:20:15,696 --> 00:20:20,046
So I often joke that after three hours
of a workshop when people are done and

337
00:20:20,056 --> 00:20:22,856
they're just like, "Wow, I can't believe
how much we did," I'm like, "Yeah, you

338
00:20:22,856 --> 00:20:27,526
know, this is basically a communication
workshop, but if I sold it as that,

339
00:20:27,886 --> 00:20:29,516
no one would pay me $1,000," right?

340
00:20:29,546 --> 00:20:33,106
They… If you put the word AI in
there, then it's suddenly valuable.

341
00:20:33,356 --> 00:20:39,346
And so tend to think that they're good
at prompting, but in reality, the way

342
00:20:39,346 --> 00:20:42,996
that, you know, the workshop starts
is you bring an idea, and then I say,

343
00:20:43,016 --> 00:20:45,866
"All right. Go write down your idea
in a Google Doc." Nothing high-tech.

344
00:20:45,876 --> 00:20:49,336
I'm just like, "Go write down your
idea." And then I ask them like, "Okay.

345
00:20:49,376 --> 00:20:51,346
Do you think someone would be
able to build this from scratch

346
00:20:51,366 --> 00:20:52,086
based on what you wrote?"

347
00:20:52,086 --> 00:20:55,046
And they go, "Of course not." And I
go, "Okay, so how would you write it

348
00:20:55,046 --> 00:20:59,436
so they can build it from scratch?"
And two sentences goes to a paragraph.

349
00:20:59,516 --> 00:21:00,726
A paragraph turns into two.

350
00:21:00,726 --> 00:21:05,896
Two turns into three pages when they
realize everything that is missing, and

351
00:21:05,896 --> 00:21:10,226
I show them techniques and tools for
how to, like, structure their writing

352
00:21:10,226 --> 00:21:11,626
but also structure their thinking.

353
00:21:12,016 --> 00:21:15,266
And then when they're done with, like,
their third or fourth revamp, they

354
00:21:15,266 --> 00:21:17,056
realize everything that was missing.

355
00:21:17,366 --> 00:21:20,726
So I think that's kind of my first
thought is everyone thinks they're good

356
00:21:20,726 --> 00:21:24,916
at prompting, and then the truth is the
better you are at prompting or context

357
00:21:24,916 --> 00:21:26,886
building, the better your results will be.

358
00:21:27,116 --> 00:21:30,166
So it's no surprise that a lot of
people are not getting good results

359
00:21:30,166 --> 00:21:34,046
with AI, is that they're missing this,
like, immediate fundamental skill

360
00:21:34,266 --> 00:21:35,846
or that it's not fully developed.

361
00:21:35,856 --> 00:21:38,796
They're not missing it in the sense that,
like, it's something that they can't do.

362
00:21:39,046 --> 00:21:40,796
It's just something that
they haven't practiced at.

363
00:21:40,876 --> 00:21:43,916
And so I consider that exercise
almost like the way that

364
00:21:43,916 --> 00:21:45,376
you play scales at a piano

365
00:21:46,376 --> 00:21:49,616
Leo Dion (host): So I think this
is a good opportunity to deep dive

366
00:21:49,616 --> 00:21:52,456
into some terminology for novices.

367
00:21:52,816 --> 00:21:55,486
But I'm gonna try, I'm gonna try
my hand at it, and you're gonna

368
00:21:55,486 --> 00:21:57,066
tell me whether I'm right or wrong.

369
00:21:57,616 --> 00:21:59,986
Like, you know, Claude would
tell me I'm wrong if I try to

370
00:21:59,986 --> 00:22:01,216
explain something incorrectly.

371
00:22:01,706 --> 00:22:01,886
I…

372
00:22:01,956 --> 00:22:05,676
And I think I'm really glad that you've
mentioned this term a lot, context and

373
00:22:05,676 --> 00:22:10,726
the context window, 'cause I think that
is… when I talk about vibe coding,

374
00:22:10,756 --> 00:22:16,016
I think that's the most important
piece when it comes to understanding

375
00:22:16,016 --> 00:22:20,286
how AI works is understanding how
to manage that context window.

376
00:22:20,876 --> 00:22:25,146
My, so my understanding is context
window is everything, essentially.

377
00:22:25,526 --> 00:22:27,606
It's everything you feed into the LLM.

378
00:22:29,096 --> 00:22:32,796
It's the system prompt that the
software developers who built

379
00:22:33,326 --> 00:22:35,936
Claude Code or Codex feed into it.

380
00:22:36,426 --> 00:22:39,306
You have your prompt,
which is the thing you say.

381
00:22:39,526 --> 00:22:41,746
You have the whole entire conversation.

382
00:22:42,056 --> 00:22:46,206
You have any files it's read in,
whether that's your CLAUDE.md,

383
00:22:46,206 --> 00:22:47,966
your constitution, your skills.

384
00:22:48,486 --> 00:22:52,396
And it, so basically every time
you say something, you're sending

385
00:22:52,446 --> 00:22:53,846
that context over and over.

386
00:22:53,846 --> 00:22:59,326
And the challenge is that people, I
think, misunderstand is that it is finite.

387
00:22:59,456 --> 00:23:05,666
It's limited to a certain part, and
managing that and figuring out what

388
00:23:05,666 --> 00:23:10,256
is the important parts that Claude or
Codex needs to know about, and what

389
00:23:10,256 --> 00:23:14,226
are the parts that can be left out
as you move through the conversation,

390
00:23:14,236 --> 00:23:18,436
is a big piece of understanding
how to use these tools properly.

391
00:23:18,496 --> 00:23:20,226
Is that… Did I say that right?

392
00:23:21,354 --> 00:23:24,494
Joe Fabisevich (guest): I feel like
that is almost 100% correct, and the

393
00:23:24,494 --> 00:23:29,229
only thing I'll caveat is Codex has
gotten ridiculously good at, like GPT

394
00:23:29,229 --> 00:23:33,634
5.5 specifically, has gotten so good at
managing the context window on its own.

395
00:23:33,894 --> 00:23:36,834
I have chats that have been going
on for a month with zero problems

396
00:23:36,834 --> 00:23:41,044
whatsoever, and I don't know what
the secret sauce that they've done.

397
00:23:41,044 --> 00:23:44,714
They have like a new memory
architecture that I think is paying

398
00:23:44,714 --> 00:23:48,924
huge dividends, but I genuinely don't
think about context windows anymore.

399
00:23:48,984 --> 00:23:50,034
It is not infinite.

400
00:23:50,054 --> 00:23:54,554
I know that it's not infinite, but
whatever they're doing to pull in more

401
00:23:54,714 --> 00:23:58,254
information that's been referenced in
the past that's valuable, it works.

402
00:23:58,294 --> 00:24:02,354
It is just like it feels like magic
that I can have a conversation

403
00:24:02,864 --> 00:24:04,364
that goes on for over a month

404
00:24:04,970 --> 00:24:07,890
Leo Dion (host): I think, so like one
of the things you may not notice when

405
00:24:07,890 --> 00:24:12,710
you're using these tools, at least with
Claude Code, is the idea of compacting.

406
00:24:13,440 --> 00:24:18,680
I did use Fable last week for that one,
those few hours that I was able to have

407
00:24:18,710 --> 00:24:22,730
access to it, and I I think I'd seen
that as well, where like it's not so

408
00:24:22,730 --> 00:24:27,080
much that the context window's gotten
bigger, it just knows how to modify

409
00:24:27,100 --> 00:24:29,130
the context window as you go through.

410
00:24:29,430 --> 00:24:33,100
We talked about this in the previous
episode of where like now with foundation

411
00:24:33,110 --> 00:24:36,550
models you can kind of do the same
thing, where you can modify the context

412
00:24:36,550 --> 00:24:38,890
window as your conversation goes through.

413
00:24:39,340 --> 00:24:41,080
So that's, yeah, that's a big part of it.

414
00:24:41,080 --> 00:24:45,840
I do think, so managing context window,
like I said, like you said, I do think

415
00:24:45,840 --> 00:24:49,970
it's gonna become less and less of an
issue, but I do think having a good

416
00:24:49,970 --> 00:24:55,460
pro- like you still have control over
your prompt and your like, I don't know,

417
00:24:55,510 --> 00:24:57,320
your co- your what do they call it?

418
00:24:57,350 --> 00:24:59,020
Like, the files you have.

419
00:24:59,020 --> 00:25:02,350
The markdown files you have, your
documentation, things like that I think

420
00:25:02,360 --> 00:25:04,170
are still gonna be important in the future

421
00:25:04,588 --> 00:25:05,048
Joe Fabisevich (guest): Totally.

422
00:25:05,128 --> 00:25:08,398
And also I'll probably… I should
caveat with one thing, which is like

423
00:25:10,238 --> 00:25:13,728
if you're building your own agent, how
you manage your context window is going

424
00:25:13,728 --> 00:25:15,278
to be actually still very important.

425
00:25:15,288 --> 00:25:20,068
So I don't wanna like shoo it away and say
like, "Everything's been solved." But so

426
00:25:20,068 --> 00:25:24,608
like if you're building a tool for an end
user, how you manage the context window

427
00:25:24,608 --> 00:25:29,248
is going to be on you, how you compact is
going to be on you, how you you know, how

428
00:25:29,248 --> 00:25:33,728
you store and replay memory is going to
be something that you have to do yourself.

429
00:25:34,068 --> 00:25:38,618
And so it's just that Codex and
GPT-5.5 have gotten really good

430
00:25:38,618 --> 00:25:41,098
at this, but that doesn't actually
mean that the problem has gone away.

431
00:25:41,098 --> 00:25:44,818
It just means that the tool that
I primarily use is very good

432
00:25:44,858 --> 00:25:49,478
at figuring out how to compact
and then seek back in memory

433
00:25:50,392 --> 00:25:53,092
Leo Dion (host): So you mentioned
building your own agent.

434
00:25:53,422 --> 00:25:57,302
This is where you're getting into stuff
I'm still a baby when it comes to.

435
00:25:57,312 --> 00:26:00,022
So what does that exactly
mean, building your own agent?

436
00:26:01,614 --> 00:26:03,884
Joe Fabisevich (guest): Yeah, an agent
is actually just like a… It's a really

437
00:26:03,884 --> 00:26:10,094
fancy term for a prompt plus tools plus…
So I always say it's like your prompts

438
00:26:10,324 --> 00:26:15,544
plus tools plus context running in a loop
until it achieves a success criteria.

439
00:26:15,864 --> 00:26:21,224
So imagine you have like a for loop
that instead of like being like run

440
00:26:21,224 --> 00:26:24,164
through this from zero to 10 or you
know, it's a while loop actually.

441
00:26:24,424 --> 00:26:27,004
It's a while loop where you set a
success criteria and you say like,

442
00:26:27,494 --> 00:26:32,884
"I want you to keep doing this until
you've achieved a goal." Right?

443
00:26:32,884 --> 00:26:35,684
So, and whatever is in the loop
is the this that you're doing.

444
00:26:36,144 --> 00:26:37,904
And then the loop, you
give it a bunch of tools.

445
00:26:37,924 --> 00:26:43,314
You could say like, "Here's a GitHub
search tool, here's web search, here's,

446
00:26:43,580 --> 00:26:45,290
Leo Dion (host): Here's a
Bash script, here's some

447
00:26:45,354 --> 00:26:46,224
Joe Fabisevich (guest):
here's a bash script.

448
00:26:46,934 --> 00:26:47,344
Yeah.

449
00:26:47,364 --> 00:26:49,914
Here's how to like y- work with
files," all that kind of stuff.

450
00:26:50,174 --> 00:26:54,734
You could provide it all those tools
and maybe the task is, I want you

451
00:26:54,734 --> 00:26:59,284
to read my file system and rename
all the files based on the way

452
00:26:59,284 --> 00:27:01,094
that I like to name files, right?

453
00:27:01,124 --> 00:27:01,424
I don't know.

454
00:27:01,484 --> 00:27:03,864
I'm saying this because I'm looking
at my computer as I'm talking to you.

455
00:27:03,974 --> 00:27:04,324
But

456
00:27:04,412 --> 00:27:06,672
Leo Dion (host): you… no, I'll tell
you an example because I'm working

457
00:27:06,692 --> 00:27:08,302
on this, I think, accidentally.

458
00:27:08,802 --> 00:27:10,992
So I've a library called SyntaxKit.

459
00:27:11,512 --> 00:27:13,292
It uses a result builder.

460
00:27:13,712 --> 00:27:19,362
It basically uses, makes building Swift
code easier, whether you're creating a

461
00:27:19,362 --> 00:27:21,862
macro or templates or things like that.

462
00:27:22,372 --> 00:27:24,942
And I hate Swift Syntax.

463
00:27:25,002 --> 00:27:26,722
I find it really difficult to use.

464
00:27:26,802 --> 00:27:30,902
So basically, it's a result builder,
uses a result builder, which is, you

465
00:27:30,902 --> 00:27:33,582
know, if you've built a SwiftUI view,
you know what I'm talking about.

466
00:27:34,052 --> 00:27:39,132
And then it takes that, makes it into…
It writes, it runs Swift Syntax code to

467
00:27:39,132 --> 00:27:41,462
then build the Swift code accordingly.

468
00:27:41,982 --> 00:27:45,842
So what I'm doing right now is I'm trying
to create, like, basically a bunch of,

469
00:27:46,422 --> 00:27:51,682
like, Swift code, my result builder
code, and I'm gonna create command

470
00:27:51,682 --> 00:27:56,742
line that's gonna go through, well, not
yet quite a loop, but eventually what

471
00:27:56,742 --> 00:28:00,842
I'd like to do is have a loop where
it constantly tries to, like, fix the

472
00:28:00,852 --> 00:28:05,992
SyntaxKit code so that whatever the
result builder expects the Swift code

473
00:28:06,032 --> 00:28:07,782
to look like, it will look like it.

474
00:28:08,202 --> 00:28:11,352
I mean, that s- sounds like
an agent kind of, right?

475
00:28:12,422 --> 00:28:12,782
Joe Fabisevich (guest): Yeah.

476
00:28:12,782 --> 00:28:14,342
So let's use your example actually.

477
00:28:14,362 --> 00:28:18,982
So in this example, let's say you
had this like loop, let's keep

478
00:28:18,982 --> 00:28:20,372
fixing this until it compiles.

479
00:28:20,722 --> 00:28:23,652
The until it compiles is
your success criteria, right?

480
00:28:23,726 --> 00:28:26,386
Leo Dion (host): Or until the Swift
code looks like the Swift code I expect

481
00:28:27,604 --> 00:28:27,914
Joe Fabisevich (guest): Sure.

482
00:28:27,964 --> 00:28:28,254
Yeah.

483
00:28:28,314 --> 00:28:32,684
The, that's… I think, for example,
actually, until it compiles I agree

484
00:28:32,684 --> 00:28:35,074
that it's not a perfect goal, so
we're gonna get into that in a second.

485
00:28:35,354 --> 00:28:38,734
But until it compiles is a
verifiable goal, so that's why it's

486
00:28:38,734 --> 00:28:41,254
inherently better than it looks
ident- like it looks like this.

487
00:28:41,534 --> 00:28:44,574
But if you could say it looks
identical to this, meaning that

488
00:28:44,574 --> 00:28:45,674
you provide a whole bunch of…

489
00:28:46,104 --> 00:28:49,364
Yeah, if you provide a whole bunch
of training data, you could say,

490
00:28:49,704 --> 00:28:54,234
"Make it look identical to this."
Now, that's your success criteria.

491
00:28:54,244 --> 00:28:55,334
It's going to run on a loop.

492
00:28:55,724 --> 00:28:59,764
The context is the goal that you're trying
to accomplish, but the tools actually

493
00:28:59,774 --> 00:29:03,224
can be… We're gonna, like, step away
from AI for a second, but it's like

494
00:29:03,584 --> 00:29:08,444
you have your compiler, you have your
syntax trees, you have, like, the AST.

495
00:29:08,634 --> 00:29:12,794
You can basically codify all these things
as tools and say, "This is what you have

496
00:29:12,794 --> 00:29:18,134
at your disposal to solve the problem."
So honestly, like, you know, a human

497
00:29:18,134 --> 00:29:19,594
at work is the same metaphor, right?

498
00:29:19,604 --> 00:29:26,334
It's like you are a person who has a job
to do, and then you are given a tool.

499
00:29:26,334 --> 00:29:29,404
You can be given the computer, you
can be given, like, a whole bunch

500
00:29:29,404 --> 00:29:32,174
of Google Docs and do whatever,
and someone says, "Go figure this

501
00:29:32,174 --> 00:29:33,664
out," and then you're an agent.

502
00:29:33,884 --> 00:29:34,424
That's it.

503
00:29:34,494 --> 00:29:37,824
That's all it is an agent is a fancy way
of saying, "I want to accomplish something

504
00:29:37,914 --> 00:29:41,354
and I want to keep doing it until I
accomplish it correctly." And so it's

505
00:29:41,354 --> 00:29:45,924
really important to define what correct
means, because a bad agent will say,

506
00:29:45,984 --> 00:29:49,024
"I've solved the problem," when it hasn't,
because it hasn't understood the problem

507
00:29:49,773 --> 00:29:49,813
Leo Dion (host): Right, right.

508
00:29:49,813 --> 00:29:53,593
Um, so w- When you talk agent, the other
thing that I am still a baby-- Well,

509
00:29:53,593 --> 00:29:57,773
I'm not a baby, I'm a teenager, 'cause
I'm starting to do, make these skills.

510
00:29:58,173 --> 00:30:03,633
So a skill is basically like a combination
of a prompt and a set of tools, right?

511
00:30:03,663 --> 00:30:09,013
Whereas agent runs in a loop, it may
run a skill in a loop more or less.

512
00:30:09,043 --> 00:30:10,723
Did I say that right?

513
00:30:12,001 --> 00:30:13,641
Joe Fabisevich (guest): They don't
necessarily have to run skills in a

514
00:30:13,641 --> 00:30:17,421
loop, but so I have a little skills
handbook, and the way that I define a

515
00:30:17,421 --> 00:30:19,811
skill is a prompt paired with resources.

516
00:30:19,931 --> 00:30:23,681
In practice, you use skills as a
domain expert, instructions for

517
00:30:23,681 --> 00:30:27,551
execute- executable commands, or a
combination of both, instructions

518
00:30:27,551 --> 00:30:29,081
for a domain expert to execute.

519
00:30:29,781 --> 00:30:34,241
So a skill could just be like, "Here's a
bunch of info that is valuable," right?

520
00:30:34,281 --> 00:30:38,381
"That you should know." And it could be,
I don't know, a code formatting skill.

521
00:30:38,381 --> 00:30:42,341
It says like, "This is how we like to
format our code." Now, a skill could

522
00:30:42,351 --> 00:30:44,701
also be the code formatter, right?

523
00:30:44,701 --> 00:30:48,001
It's like whenever you have to
format code, go run this command.

524
00:30:48,471 --> 00:30:50,311
Or ideally, the skill is both.

525
00:30:50,321 --> 00:30:53,891
It says, "This is Leo's code style,
and any time that we're writing

526
00:30:53,891 --> 00:30:57,081
code, we should reformat it to
make sure that it does this." You

527
00:30:57,081 --> 00:30:59,831
could build a skill around, for
example, your syntax kit, right?

528
00:31:00,211 --> 00:31:05,361
So every open source repo I publish now
has skills for integrating that library.

529
00:31:05,381 --> 00:31:10,471
So Boutique has a Boutique best practices
skill, and what this does is it saves

530
00:31:10,471 --> 00:31:13,711
me a whole bunch of support requests
when people say, "I added Boutique

531
00:31:13,731 --> 00:31:15,941
and I did it completely wrong," right?

532
00:31:15,961 --> 00:31:18,321
And I'm like, "Well, you should have
read the docs." Instead of reading the

533
00:31:18,321 --> 00:31:23,561
docs, I'm like, "Literally just go ask
Claude to integrate it for you into your

534
00:31:23,581 --> 00:31:26,851
app, and it will read the skill, and
the skill will teach you the best way to

535
00:31:26,851 --> 00:31:33,221
do it." So it's domain expertise about
the subject, in this case, databases.

536
00:31:33,771 --> 00:31:39,581
And yeah, like my new logging library will
have a skill that's just like how to best

537
00:31:39,621 --> 00:31:44,691
integrate Broadcast into your app, and
you don't even have to write the code.

538
00:31:44,691 --> 00:31:48,221
You just go, you know, dollar
sign or slash depending on Claude

539
00:31:48,221 --> 00:31:55,141
Coders u- or Codex, dollar sign,
Broadcast best practices, add a

540
00:31:55,141 --> 00:31:56,711
whole bunch of logs, and that's it.

541
00:31:56,721 --> 00:31:57,371
It will do it for you.

542
00:31:58,969 --> 00:32:02,809
Leo Dion (host): What would you recommend
for, like, a software developer today?

543
00:32:02,949 --> 00:32:06,569
Like, depending on their, I guess,
whether they're a beginner to a senior,

544
00:32:06,569 --> 00:32:11,139
like, what do they need to do to get
hired, I guess, or to be successful?

545
00:32:12,751 --> 00:32:15,271
Joe Fabisevich (guest): I think to get
hired is different than to be successful,

546
00:32:15,471 --> 00:32:16,371
Leo Dion (host): Yeah, I agree.

547
00:32:17,151 --> 00:32:18,181
Joe Fabisevich (guest): like,
what people should learn.

548
00:32:19,051 --> 00:32:22,151
I think the truth is that, one, you
should get good at using AI because it's

549
00:32:22,211 --> 00:32:25,831
undeniable that AI will be a part of how
software developers write code, right?

550
00:32:25,831 --> 00:32:29,441
So I don't wanna, like, get high
on my own supply, but I think it's

551
00:32:29,441 --> 00:32:31,331
really important to get good at AI.

552
00:32:31,441 --> 00:32:34,591
And by that I mean is you should
be seeing successful results

553
00:32:34,591 --> 00:32:37,501
from when you're using AI, and if
not, there's something to learn.

554
00:32:37,511 --> 00:32:41,501
And again, I'm always… I think
my philosophy in life is always you

555
00:32:41,501 --> 00:32:43,001
should be learning all the time.

556
00:32:43,101 --> 00:32:49,941
So if you continually compound learning,
you can't guarantee career success, but

557
00:32:49,961 --> 00:32:53,741
you can maximize the odds of it happening.

558
00:32:54,001 --> 00:32:57,991
So whatever domain you wanna get
good at, I would continue to learn.

559
00:32:58,001 --> 00:33:03,061
So I do think that engineering is
probably going to bifurcate and have,

560
00:33:03,061 --> 00:33:08,361
like, a barbell curve where people at
the highest end, you know, you could

561
00:33:08,361 --> 00:33:14,121
be, like… If you want to be a SwiftUI
developer or a Swift developer, you

562
00:33:14,121 --> 00:33:16,811
should aim to be in the, like, top 10
percentile, because there's still going

563
00:33:16,811 --> 00:33:21,551
to be demand for people who know Swift
so well that when AI is generating code

564
00:33:21,551 --> 00:33:23,861
that doesn't make sense, that they're the
ones who are like, "Ah, you know what?

565
00:33:23,861 --> 00:33:28,621
This is the way that we should be doing
it instead." But AI is… I don't know.

566
00:33:28,631 --> 00:33:32,691
There's, like, the saying, like, AI is
an av- is average because of the training

567
00:33:32,691 --> 00:33:36,321
data, and I think that's a bit of a
misnomer because you get to control the

568
00:33:36,321 --> 00:33:39,691
training data and curate it, so it's like
you could actually skew things higher.

569
00:33:40,031 --> 00:33:44,301
But you can just assume that AI, if
you are in the bottom 50 percentile of

570
00:33:44,301 --> 00:33:48,151
some subject matter and do not intend
to get better at it, you would be best

571
00:33:48,171 --> 00:33:52,861
off using AI to raise the floor of
your potential in that subject matter.

572
00:33:53,401 --> 00:33:57,631
And so you're probably gonna wanna
develop, like, some sort of expertise,

573
00:33:57,681 --> 00:33:59,211
regardless of what that expertise is.

574
00:33:59,381 --> 00:34:02,961
Whatever you're looking to get
hired for, you better be, you know,

575
00:34:03,001 --> 00:34:04,521
striving to be the best at it.

576
00:34:04,541 --> 00:34:09,811
Because otherwise, AI is going to
be better than you at it very soon.

577
00:34:10,401 --> 00:34:13,221
So if you're looking to optimize
career success, I'd say that.

578
00:34:14,209 --> 00:34:16,509
Leo Dion (host): How, like,
how will that affect as far

579
00:34:16,509 --> 00:34:18,249
as, like, recent grads though?

580
00:34:18,279 --> 00:34:20,519
Like, who, like, just got out of college.

581
00:34:20,519 --> 00:34:27,509
Like, what are they supposed to do to
get hired or be successful in either way?

582
00:34:28,587 --> 00:34:29,707
Joe Fabisevich (guest): Yeah,
I mean, I think this is the

583
00:34:29,947 --> 00:34:31,227
thorniest of all questions, right?

584
00:34:31,227 --> 00:34:31,867
Is like I

585
00:34:31,943 --> 00:34:32,373
Leo Dion (host): I know

586
00:34:32,377 --> 00:34:33,797
Joe Fabisevich (guest): think
there's a good answer here.

587
00:34:33,907 --> 00:34:37,007
I would be lying if I was like,
"Here's a recipe or formula." I do

588
00:34:37,007 --> 00:34:40,307
see a lot of people who are from
non-traditional backgrounds adopting

589
00:34:40,317 --> 00:34:44,757
AI, but also learning skills that are
adjacent, like product and marketing

590
00:34:44,757 --> 00:34:46,807
and public speaking and development.

591
00:34:46,817 --> 00:34:50,317
Like, you know, there is room for
people who are going to be generalists.

592
00:34:52,127 --> 00:34:54,377
So I've always said that the
right career… Like the right

593
00:34:54,377 --> 00:34:58,207
shape for a person is T-shaped
for career path success.

594
00:34:58,457 --> 00:35:00,647
And I don't know if you know what the
term T-shaped engineer is, but it's

595
00:35:00,683 --> 00:35:01,253
Leo Dion (host): No, I do.

596
00:35:01,283 --> 00:35:01,573
Yeah

597
00:35:02,557 --> 00:35:04,147
Joe Fabisevich (guest): Yeah,
well, I'll say it for anyone

598
00:35:04,157 --> 00:35:05,047
who doesn't, but it's like

599
00:35:05,383 --> 00:35:05,743
Leo Dion (host): Please

600
00:35:06,317 --> 00:35:08,757
Joe Fabisevich (guest): you want
to be an expert at something and

601
00:35:08,757 --> 00:35:10,117
really good at everything else.

602
00:35:10,167 --> 00:35:13,097
And the reason for that is not because…
Like you want to be able to lean on

603
00:35:13,097 --> 00:35:17,427
your expertise, but you also want to
be able to see where everyone else is

604
00:35:17,427 --> 00:35:22,197
coming from and achieve more success
by being able to be working together.

605
00:35:22,207 --> 00:35:26,867
So if you imagine this ideal team in
a world of AI, it's like a designer,

606
00:35:26,897 --> 00:35:31,907
a developer, a marketer, a product
person, and, you know, they're small,

607
00:35:31,907 --> 00:35:35,217
they're scrappy, but with AI they're
able to do so much more than they

608
00:35:35,217 --> 00:35:37,557
could have two, three years ago, right?

609
00:35:38,077 --> 00:35:41,637
And so what's important there is
that they're all really dang good

610
00:35:41,637 --> 00:35:44,397
at what they do, but they can
all collaborate well together.

611
00:35:44,417 --> 00:35:45,587
They are T-shaped.

612
00:35:45,977 --> 00:35:51,717
And I think the formula doesn't change
in a world of AI for a new grad either.

613
00:35:52,237 --> 00:35:56,207
The problem that we have is that a lot of
what you develop in the kind of generalist

614
00:35:56,207 --> 00:36:01,187
skills comes from experience, and so
that's why it's much harder for new grads.

615
00:36:01,577 --> 00:36:05,607
But when I say there's a barbell curve,
I mean, yes, you want to be the top

616
00:36:05,617 --> 00:36:12,067
10% at some specific skill, or you
want to be on the lower end, which is

617
00:36:12,067 --> 00:36:13,637
like, I kind of think the worst part.

618
00:36:13,687 --> 00:36:18,687
The worst place in your career to be
is six years into a profession where

619
00:36:18,687 --> 00:36:23,987
you were getting paid $150,000 and you
were okay at it, but the demand is so

620
00:36:23,987 --> 00:36:28,567
high for expertise in that subject that
people are willing to take the second,

621
00:36:28,567 --> 00:36:30,227
third or fourth best option as well.

622
00:36:30,577 --> 00:36:35,597
And so if you are new, you're possibly
in a better position to get jobs than

623
00:36:35,597 --> 00:36:41,187
someone who is, you know, in that
mid-level state and mid-level, but that

624
00:36:41,187 --> 00:36:44,997
their plan was to basically be mid-level
for the rest of their career as opposed

625
00:36:44,997 --> 00:36:49,397
to striving to become like senior or
super senior staff, that kind of thing.

626
00:36:49,737 --> 00:36:55,157
And I think basically the best advice
I can give is like continue to show

627
00:36:55,157 --> 00:36:58,997
that you have growth potential, because
if you combine growth potential with

628
00:36:58,997 --> 00:37:04,107
AI, then you have someone who can
deliver results as they're learning.

629
00:37:04,347 --> 00:37:08,287
I often come back to the fact that
like I learned how to code by making

630
00:37:08,347 --> 00:37:11,987
stuff, but also like using tools
like Stack Overflow and Googling.

631
00:37:12,337 --> 00:37:16,817
And then there's a whole generation of
developers shortly after me who learned

632
00:37:16,827 --> 00:37:17,997
everything they know from YouTube.

633
00:37:17,997 --> 00:37:22,067
And it's not a way that I learn, but it
doesn't mean that it's like… It doesn't

634
00:37:22,067 --> 00:37:24,187
disqualify them as having learned, right?

635
00:37:24,187 --> 00:37:27,595
They still can like do deep
dives and under- Understand

636
00:37:27,595 --> 00:37:29,415
things that I understand.

637
00:37:29,695 --> 00:37:35,195
So just because our tools for learning
change doesn't mean that learning stops,

638
00:37:35,255 --> 00:37:39,805
and I think it's really important to
internalize because there's gonna be a

639
00:37:39,805 --> 00:37:41,745
lot of people who are AI native learners.

640
00:37:41,745 --> 00:37:45,775
They learned with AI, and yes, the tools
are not always perfect, but they can

641
00:37:45,775 --> 00:37:50,475
accelerate their learning in a way that
we may not even be able to foresee yet.

642
00:37:50,485 --> 00:37:53,595
So I'm not here giving up on new
grads or anything like that, even

643
00:37:53,595 --> 00:37:58,195
though I don't have, like, a clean
answer other than to say exercise and

644
00:37:58,195 --> 00:38:00,025
show your potential early and often

645
00:38:00,813 --> 00:38:03,743
Leo Dion (host): And I like, you had
mentioned this earlier too, like I

646
00:38:03,743 --> 00:38:05,843
like the idea of learning via AI.

647
00:38:06,643 --> 00:38:10,683
Like look at what AI does, look
at what Claude Code does, and

648
00:38:10,703 --> 00:38:13,323
understand why it does what it does.

649
00:38:13,653 --> 00:38:16,873
Like if you want personal growth, that's
a big part of it, 'cause I've been in

650
00:38:16,873 --> 00:38:20,473
the same boat as well, and this idea
of like, I don't know if you've heard

651
00:38:20,473 --> 00:38:25,343
of like the centaur developer where
it's like the developer attached to the

652
00:38:25,353 --> 00:38:27,833
AI, like working together in unison.

653
00:38:27,853 --> 00:38:32,203
And I mean, I don't see that
going away anytime soon.

654
00:38:32,223 --> 00:38:35,473
I don't know about you, if we're gonna
get to a point where AI's just gonna

655
00:38:35,473 --> 00:38:38,033
be so good where you don't need that.

656
00:38:38,633 --> 00:38:39,233
But I don't…

657
00:38:39,251 --> 00:38:39,771
Joe Fabisevich (guest): on the day.

658
00:38:40,003 --> 00:38:40,693
Leo Dion (host): Yeah, right.

659
00:38:41,213 --> 00:38:44,833
But what do you, I guess what do
you recommend then like companies do

660
00:38:44,843 --> 00:38:46,543
when it comes to the hiring process?

661
00:38:46,573 --> 00:38:48,303
Because that's a mess.

662
00:38:48,363 --> 00:38:52,013
Like hiring is just awful right now.

663
00:38:52,283 --> 00:38:54,633
People don't know what, companies
don't know what they're doing in

664
00:38:54,633 --> 00:38:58,643
a lot of ways and are sort of just
still like, "Do you have a bachelor's

665
00:38:58,643 --> 00:39:01,893
degree and a resume?" When that l-
like doesn't really tell you anything.

666
00:39:01,913 --> 00:39:05,703
Or we're posting jobs that
aren't even, don't even exist.

667
00:39:05,713 --> 00:39:07,483
You know, it's a mess to say the least.

668
00:39:07,903 --> 00:39:12,243
Like w- what would you recommend your
ideal company do when it comes to hiring?

669
00:39:13,423 --> 00:39:14,683
Joe Fabisevich (guest): Yeah,
so I mean, actually I wanna take

670
00:39:14,683 --> 00:39:17,893
a step back to learning for, to
make one kind of, one more point.

671
00:39:17,903 --> 00:39:21,993
So you mentioned the idea that
you should pause and understand

672
00:39:21,993 --> 00:39:23,043
what Claude Code is doing.

673
00:39:23,813 --> 00:39:26,173
I think the higher level
concept there is friction.

674
00:39:26,173 --> 00:39:27,543
Friction is so important to learning.

675
00:39:27,783 --> 00:39:34,523
You can't just, like, you can't, I
don't know, open a funnel and then have

676
00:39:34,603 --> 00:39:38,193
learning happen by shoving more and
more knowledge through that funnel.

677
00:39:38,243 --> 00:39:39,353
It just doesn't work like that.

678
00:39:39,663 --> 00:39:42,193
You have to actually sit with an idea.

679
00:39:42,243 --> 00:39:46,903
I really like this like one Ezra Klein
quote about how reading for him is

680
00:39:46,903 --> 00:39:50,373
not just taking in the information,
it's… He takes his time with

681
00:39:50,373 --> 00:39:54,853
reading because sitting with the
book for a week or a month is what

682
00:39:54,983 --> 00:39:56,803
actually makes the learning happen.

683
00:39:56,893 --> 00:40:02,853
And so I think that to tie it back into
like employers, I still think it's gonna

684
00:40:02,853 --> 00:40:07,303
be really valuable if you wanna have a
durable firm to invest in your employees.

685
00:40:07,333 --> 00:40:10,293
And I think this is that long-term
versus short-term thinking that

686
00:40:10,293 --> 00:40:11,273
we were talking about earlier.

687
00:40:11,733 --> 00:40:17,573
I can't… I'm never gonna build a
company that is, you know, optimized

688
00:40:17,583 --> 00:40:19,213
for making money, and that's going to…

689
00:40:19,233 --> 00:40:24,313
That's like kind of a flaw in
my own place in the economy.

690
00:40:24,893 --> 00:40:28,913
But I found that long-term investment
has led to durable success in my

691
00:40:28,913 --> 00:40:33,343
career, and I think that has lur-
led to durable companies as well

692
00:40:33,693 --> 00:40:38,423
that outlast any particular fad and
that stand the test of time, right?

693
00:40:38,443 --> 00:40:43,473
Like, a company like Nintendo or
Microsoft or whatever, they're not…

694
00:40:44,023 --> 00:40:48,613
They're thinking at, from first principles
about like, what do we want users or

695
00:40:48,613 --> 00:40:52,893
people to be feeling when they interact
with our company and our products?

696
00:40:53,093 --> 00:40:56,003
They're not thinking about, how
do I make the most bang for my

697
00:40:56,003 --> 00:40:57,133
buck in the next three months?

698
00:40:57,393 --> 00:40:58,643
Arguably, sometimes they do.

699
00:40:58,773 --> 00:41:04,009
But if you like go back to the early
Microsoft days, it's like- We're building

700
00:41:04,009 --> 00:41:07,799
an operating system because, and that
because is the thing that matters.

701
00:41:07,999 --> 00:41:12,269
It's not just we found a random, like,
loophole in the market and we're gonna

702
00:41:12,269 --> 00:41:16,679
exploit it like a D2C company for
the next 30 days, and then figure out

703
00:41:16,679 --> 00:41:18,369
what the next exploitable loophole is.

704
00:41:18,799 --> 00:41:21,999
So if I was building a company,
I would focus on how to build

705
00:41:21,999 --> 00:41:25,039
a company that can last years
through any changes that happen.

706
00:41:25,479 --> 00:41:29,239
Now, in terms of hiring, I still
think hiring has… I mean,

707
00:41:29,249 --> 00:41:30,439
hiring's always been broken, right?

708
00:41:30,469 --> 00:41:32,929
Like, especially in our career.

709
00:41:33,289 --> 00:41:35,699
Part of the problem is that
we test for things that we do

710
00:41:35,699 --> 00:41:38,379
not actually value in the job.

711
00:41:38,389 --> 00:41:41,009
And so it's really important to
take a step back and say, "What

712
00:41:41,009 --> 00:41:46,009
do you value in the job?" And one
example of that would be, I…

713
00:41:46,219 --> 00:41:48,619
I'll link to this, but I designed
an interview process a while back.

714
00:41:48,659 --> 00:41:52,329
I actually designed an interview
process last year for a company, and

715
00:41:52,519 --> 00:41:53,549
it was, like, a screening process.

716
00:41:53,549 --> 00:41:54,219
It went really well.

717
00:41:54,539 --> 00:41:58,389
And then I also designed, like,
a full-day loop kind of thing.

718
00:41:58,459 --> 00:42:02,669
And I think that what you wanna do is
if your company is building product,

719
00:42:02,959 --> 00:42:08,189
you want an engineer to meet with a
product person, and then they discuss

720
00:42:08,209 --> 00:42:09,259
the product they're going to build.

721
00:42:09,259 --> 00:42:09,929
Like, imagine a day.

722
00:42:09,929 --> 00:42:14,309
You come in and you discuss the product,
and then you work with a backend

723
00:42:14,309 --> 00:42:17,679
developer to build the API, and then
you work with a client developer.

724
00:42:17,709 --> 00:42:20,579
If let's say you're an iOS developer
you then have to build out your,

725
00:42:21,049 --> 00:42:24,239
the actual API that you've designed
with the backend developer.

726
00:42:24,529 --> 00:42:27,549
And it's, like, a real collaborative
process because that's what we're

727
00:42:27,559 --> 00:42:31,379
imagining work to be, is this process
of collaboration between humans.

728
00:42:31,919 --> 00:42:36,559
Now, if you're more on the AI
side, then you're gonna wanna

729
00:42:36,559 --> 00:42:38,059
start testing those skills, right?

730
00:42:38,069 --> 00:42:42,819
Like, I interviewed with an AI company,
and there was no AI even though they were

731
00:42:42,819 --> 00:42:44,289
like, "We want you for your AI skills."

732
00:42:44,299 --> 00:42:47,279
And I'm like, that's what I would call
a misalignment between the interview

733
00:42:47,279 --> 00:42:48,989
process and the desired outcome.

734
00:42:49,619 --> 00:42:52,939
And of course, they're looking to
test my ability to write Swift and

735
00:42:52,939 --> 00:42:56,719
SwiftUI and, like, or TypeScript
or whatever I'm interviewing for.

736
00:42:57,059 --> 00:43:00,869
But they also just left
off a huge part of this.

737
00:43:00,869 --> 00:43:04,119
So it's really important for companies to
just step back and say, like, "If we're

738
00:43:04,119 --> 00:43:05,769
hiring, then what are we hiring for?

739
00:43:05,769 --> 00:43:08,369
And if we're hiring for that,
then how do we test for that

740
00:43:08,489 --> 00:43:10,029
and work from that perspective?"

741
00:43:10,661 --> 00:43:10,951
Leo Dion (host): Yeah.

742
00:43:11,311 --> 00:43:12,601
Yep, I like that a lot.

743
00:43:13,391 --> 00:43:17,641
It's just pain in the butt to get
a company to actually do that,

744
00:43:18,025 --> 00:43:19,755
Joe Fabisevich (guest): yeah,
I can't fix the economy, Leo,

745
00:43:19,805 --> 00:43:20,595
Leo Dion (host): yeah, I know.

746
00:43:20,755 --> 00:43:21,705
Not in this episode.

747
00:43:21,795 --> 00:43:22,585
Maybe in the next one.

748
00:43:23,155 --> 00:43:23,905
All right, so let's…

749
00:43:23,935 --> 00:43:24,615
Joe Fabisevich (guest):
back for a fourth time

750
00:43:27,045 --> 00:43:28,565
Leo Dion (host): Joe fixes the economy.

751
00:43:28,685 --> 00:43:31,785
So let's talk about, like,
your actual daily setup.

752
00:43:31,785 --> 00:43:33,905
So I'm a Claude Code person.

753
00:43:33,935 --> 00:43:36,225
I just pretty much do only Claude Code.

754
00:43:36,905 --> 00:43:38,265
I used to do Cursor.

755
00:43:38,345 --> 00:43:40,515
I don't really do use Cursor anymore.

756
00:43:40,705 --> 00:43:45,775
And now I… If I do need, like, to
do s- like, I guess I use VS Code when

757
00:43:45,775 --> 00:43:50,005
I'm doing something that involves a
dev container because then it's just

758
00:43:50,005 --> 00:43:54,455
easier to talk to Claude through
that so I can fix a Linux issue.

759
00:43:54,985 --> 00:43:58,505
But yeah, what, like, I guess,
what's your setup right now?

760
00:43:59,875 --> 00:44:01,855
Joe Fabisevich (guest): Yeah, so I want to
preface this with I think that the tools

761
00:44:01,855 --> 00:44:04,895
are good enough that you can basically do
anything you want with any tool you want.

762
00:44:04,935 --> 00:44:08,015
And so in a lot of ways it's kind
of like how some people use Vim,

763
00:44:08,015 --> 00:44:09,375
some people use what's it called?

764
00:44:09,455 --> 00:44:11,195
Emacs, and some people use, like, Xcode.

765
00:44:11,665 --> 00:44:14,145
Yeah, like, it doesn't matter.

766
00:44:14,265 --> 00:44:14,575
You know?

767
00:44:14,575 --> 00:44:15,425
Like, it's totally fine.

768
00:44:15,435 --> 00:44:17,305
Everyone wants to hear, like,
what's your setup, what's this.

769
00:44:17,665 --> 00:44:21,595
At the end of the day, I use the Codex
for Mac app, and I have my own workflow

770
00:44:21,595 --> 00:44:25,805
in it, and it is, like, super effective
for me, and I am very happy with it.

771
00:44:26,075 --> 00:44:28,935
And if someone doesn't like Codex
or they don't like OpenAI and they

772
00:44:28,935 --> 00:44:31,725
don't want to use any of their
products, you can accomplish all of

773
00:44:31,725 --> 00:44:32,895
this with any tool that you want.

774
00:44:32,905 --> 00:44:35,875
So, like, when I teach my workshops,
I ask people what they want to use,

775
00:44:36,165 --> 00:44:40,045
and then we do it in that tool, and
they are always happy nonetheless

776
00:44:41,071 --> 00:44:44,821
Leo Dion (host): So one of the things
I wanna talk about, and I've seen this

777
00:44:44,831 --> 00:44:49,861
problem a lot you talked about this
briefly with your skills about formatting,

778
00:44:50,421 --> 00:44:54,731
is like figuring out a way to do stuff.

779
00:44:54,901 --> 00:44:58,871
I guess, fig- like a lot of… And Claude
is pretty good at knowing when to do it.

780
00:44:59,061 --> 00:45:04,881
It's like don't do find and
replace using a AI tool.

781
00:45:04,891 --> 00:45:10,361
Have the AI tool write a command
or write a script to do it,

782
00:45:10,361 --> 00:45:12,071
and then automate it that way.

783
00:45:12,501 --> 00:45:14,551
Because otherwise you're
gonna go through tokens.

784
00:45:14,671 --> 00:45:17,451
And I call this automate
by AI and not with AI.

785
00:45:18,121 --> 00:45:23,621
I guess, I… My question i- my
question is, like, what other tips

786
00:45:23,661 --> 00:45:30,191
might you have to help, like, use
these tools optimally and not just fly

787
00:45:30,211 --> 00:45:32,351
through tokens when you don't need to?

788
00:45:33,655 --> 00:45:34,675
Joe Fabisevich (guest): Yeah,
actually I think you hit on the

789
00:45:34,715 --> 00:45:38,195
most important one, which is have it
write code that is reusable, right?

790
00:45:38,235 --> 00:45:42,225
You want… if you're building a
process for your business, you want

791
00:45:42,225 --> 00:45:46,005
it to be deterministic, and you
want it to be not using tokens.

792
00:45:46,005 --> 00:45:47,885
So I have AI write scripts all the time.

793
00:45:48,195 --> 00:45:53,245
There's a coding agent or a harness
called Pi, P-I, and its whole

794
00:45:53,245 --> 00:45:55,235
premise is it only has four tools.

795
00:45:55,235 --> 00:45:58,375
It's like read, write, bash, and grep.

796
00:45:58,855 --> 00:46:02,495
And if it wants to solve a problem, it
writes code that solves the problem.

797
00:46:02,725 --> 00:46:06,645
And so that's where bash comes in,
and all you have to do is write code.

798
00:46:06,965 --> 00:46:11,515
I think that this is part of the reason
that Pi is consistently among the top,

799
00:46:11,515 --> 00:46:13,445
if not the top, like coding agent.

800
00:46:13,485 --> 00:46:16,475
You know, it combines with the,
with a model like GPT-5 or whatever.

801
00:46:16,805 --> 00:46:20,735
But The thing that coding agents
are best at is writing code.

802
00:46:21,035 --> 00:46:24,815
And so we should be optimizing for
that, and we should be optimizing

803
00:46:25,125 --> 00:46:27,065
business processes towards that as well.

804
00:46:27,205 --> 00:46:33,795
There are also like, I don't 100% do
this myself, but I would say the models

805
00:46:33,835 --> 00:46:38,655
that are open and available and served
cheaply are getting so good that they

806
00:46:38,655 --> 00:46:42,215
are capable of doing many of the things
that, like, you would previously need

807
00:46:42,265 --> 00:46:46,795
a GPT or a, like, Claude model to do.

808
00:46:47,185 --> 00:46:51,815
And so if you're thinking about this as
like how do I save money then… And but

809
00:46:51,815 --> 00:46:57,085
I still want AI involved, not for coding
specifically, but for you build an agent.

810
00:46:57,105 --> 00:47:02,415
You can build an agent with like a
GLM 5.2 or like a Kimi model, and

811
00:47:02,415 --> 00:47:05,465
they're really cheap because the
open source models are dramatically

812
00:47:05,475 --> 00:47:10,745
cheaper to serve than, you know, like
the giant monster, GPT 5.5 model.

813
00:47:10,985 --> 00:47:15,295
And so at some point, we're probably
gonna have to start thinking about,

814
00:47:15,385 --> 00:47:18,205
more about like it's not this duopoly.

815
00:47:18,225 --> 00:47:23,005
There's a landscape of AI tools
available at our disposal, and some

816
00:47:23,005 --> 00:47:25,065
people, you know, they already do this.

817
00:47:25,065 --> 00:47:28,585
They say like, "I like to use Manus.
I like to use Devin. I like to use

818
00:47:28,615 --> 00:47:33,595
Codex." But when… You know, we
really have to think of this as like

819
00:47:34,075 --> 00:47:36,405
intelligence available to us for a price.

820
00:47:37,145 --> 00:47:40,455
For a price is really important because
we're gonna have to make trade-offs

821
00:47:40,455 --> 00:47:44,375
and say, "I don't actually need the
most frontier intelligence right now. I

822
00:47:44,375 --> 00:47:48,645
would actually get away with using…"
The same way that we choose between,

823
00:47:48,655 --> 00:47:51,825
like sometimes I can use Claude Haiku
instead of Claude Sonnet or Claude Opus.

824
00:47:52,155 --> 00:47:55,955
I do the mental math and say like,
"Yeah, this ch- this task is easy."

825
00:47:56,325 --> 00:48:01,885
And it's like a little… I think the
a- the AI industry is like obsessed

826
00:48:01,885 --> 00:48:06,625
with GPUs, but a lot of agentic work
is actually CPU bound, not GPU bound.

827
00:48:07,125 --> 00:48:15,485
And so we're gonna… Yeah, I mean,
the actual AI generation is, like

828
00:48:15,505 --> 00:48:17,815
the token generation is GPU bound.

829
00:48:18,545 --> 00:48:24,615
But coming up with like a task or
scheduling it or doing anything

830
00:48:24,615 --> 00:48:28,505
local oftentimes is CPU friendly.

831
00:48:29,015 --> 00:48:33,955
And so we are gonna start rethinking
our relationship with like GPUs, and

832
00:48:33,955 --> 00:48:38,565
that will drop the price of an agentic
task dramatically because not…

833
00:48:39,105 --> 00:48:42,155
You know, GPUs are not actually very
good at doing what CPUs do, and CPUs

834
00:48:42,185 --> 00:48:43,895
are not good at doing what GPUs do.

835
00:48:44,195 --> 00:48:47,825
So we're just gonna lean into each other's
strengths, and doing things on a CPU

836
00:48:47,825 --> 00:48:52,315
is so much cheaper that we're probably
gonna invent processes to do things

837
00:48:52,315 --> 00:48:53,735
on CPUs that could be done on GPUs.

838
00:48:53,765 --> 00:48:59,235
But even NVIDIA is now making CPUs,
so I feel confident that we will

839
00:48:59,685 --> 00:49:05,735
like, we will not have to worry as
much about, "Oh, this agent is doing

840
00:49:05,745 --> 00:49:07,555
X, Y, and Z, and it costs so much

841
00:49:08,591 --> 00:49:12,191
Leo Dion (host): What I wanna talk more
about local models and things like that.

842
00:49:12,191 --> 00:49:14,761
But before, there was a couple
things I wanna get into.

843
00:49:15,331 --> 00:49:21,891
We touched on, so like documentation
is really important, and one thing

844
00:49:21,941 --> 00:49:23,391
I started looking at SpecKit.

845
00:49:23,491 --> 00:49:26,611
I don't like any of these
like task management tools.

846
00:49:26,681 --> 00:49:31,461
I find them all to be a lot of
like cruft that isn't needed.

847
00:49:31,471 --> 00:49:35,641
But one thing I did pick up from
SpecKit was this idea of a constitution.

848
00:49:36,381 --> 00:49:39,171
You know, saying like, "Hey,
don't do this, don't do that," and

849
00:49:39,551 --> 00:49:41,321
seems to follow those rules okay.

850
00:49:41,321 --> 00:49:44,731
It hasn't, you know, done anything crazy.

851
00:49:44,781 --> 00:49:50,851
What are some documentation or like file
must-haves when you have your project?

852
00:49:52,649 --> 00:49:54,149
Joe Fabisevich (guest): I'm not
actually really sure how to answer that.

853
00:49:54,149 --> 00:49:57,059
I would say that, like, having a good
CLAUDE.md is probably the most important

854
00:49:57,059 --> 00:50:01,049
thing you can do, and so it's like using
the CLAUDE.md to define the process.

855
00:50:01,599 --> 00:50:07,549
I also think SpecKit and tools like
that, the… I would say that they are

856
00:50:08,769 --> 00:50:13,259
the best way to take someone who does
not know how to write code and give them

857
00:50:13,279 --> 00:50:15,899
a leg up in building software, right?

858
00:50:16,119 --> 00:50:19,999
So they may not inherently be
designed for software developers.

859
00:50:20,039 --> 00:50:23,439
I think they're actually designed
for builders more widely.

860
00:50:23,799 --> 00:50:28,219
And so you can substitute any of
this kind of stuff with 15 years

861
00:50:28,219 --> 00:50:29,449
of software development experience.

862
00:50:29,689 --> 00:50:33,079
The problem is that most people don't have
15 years software development experience.

863
00:50:33,399 --> 00:50:37,619
And so I don't have any, like… You
know, at first I feel like working with

864
00:50:37,619 --> 00:50:42,739
AI was tips and tricks of figuring out
how to get the AI to do the thing that

865
00:50:42,739 --> 00:50:46,369
you want it to do, and then we formalize
the process in things like SpecKit.

866
00:50:46,929 --> 00:50:51,329
But ultimately, it comes down to, like,
being able to communicate the desired

867
00:50:51,329 --> 00:50:55,139
task, being able to understand the
edge cases, being able to, like, figure

868
00:50:55,139 --> 00:50:58,949
out where to take shortcuts, where not
to take shortcuts, all these things

869
00:50:58,949 --> 00:51:03,709
that it's like you can't substitute
that with a doc or a skill per se.

870
00:51:03,909 --> 00:51:09,039
But for example, like, I oftentimes am…
I ran into a performance issue yesterday.

871
00:51:09,039 --> 00:51:12,819
Someone emailed me and
were like, "This is so bad.

872
00:51:12,819 --> 00:51:15,269
I don't understand how it's so bad." And I
was like, "I don't understand either." And

873
00:51:15,279 --> 00:51:16,789
I was like, "Send me a spin dump," right?

874
00:51:17,169 --> 00:51:20,139
And so they sent me a spin dump,
and I gave that spin dump to Codex

875
00:51:20,139 --> 00:51:21,229
and said, "You figure it out."

876
00:51:21,679 --> 00:51:27,349
And I gave it a Swift performance
audit tool and was like, "You figure

877
00:51:27,349 --> 00:51:30,089
it out with this skill," and it did.

878
00:51:30,279 --> 00:51:32,969
I was like… It was like, "Oh, your
timeline view is redrawing, like, a crazy

879
00:51:32,969 --> 00:51:36,669
amount in this one specific edge case."
And because I gave it a bunch of logs, it

880
00:51:36,669 --> 00:51:38,909
was able to figure out, like, you were…

881
00:51:39,529 --> 00:51:41,549
You caught this edge case here.

882
00:51:41,689 --> 00:51:42,669
Here's how it happened.

883
00:51:42,669 --> 00:51:45,139
Here's how it manifested, and then
here's what I would do to fix it.

884
00:51:45,139 --> 00:51:50,289
And I was like, "That's really dang
handy." And so the reason I wasn't,

885
00:51:50,399 --> 00:51:53,119
like, inclined to solve that on my
own is I'm not actually really good

886
00:51:53,139 --> 00:51:59,289
at performance profiling, but I am
good at knowing how to guide a tool

887
00:51:59,309 --> 00:52:00,839
to do good performance profiling.

888
00:52:00,869 --> 00:52:05,019
And so that's an example where even as a
software developer, I was making up for

889
00:52:05,129 --> 00:52:07,489
a lack of skill of mine with a skill.

890
00:52:08,327 --> 00:52:09,807
Leo Dion (host): Yeah,
that makes a lot of sense.

891
00:52:10,247 --> 00:52:14,057
I, like, I guess for me it's just
important to, like, what's important

892
00:52:14,057 --> 00:52:18,407
about documentation or you said agents,
I assume that's the same thing as

893
00:52:18,407 --> 00:52:20,057
like the Claude MD basically, but

894
00:52:20,381 --> 00:52:20,931
Joe Fabisevich (guest): CLAUDE.md, yeah

895
00:52:21,107 --> 00:52:21,427
Leo Dion (host): Yeah.

896
00:52:21,507 --> 00:52:27,097
But I think it's important to have some
set of like understanding, like for

897
00:52:27,097 --> 00:52:33,517
the context so that way a tool like
Codex or Claude will understand exactly

898
00:52:33,527 --> 00:52:35,807
like, "Okay, this is what I need.

899
00:52:35,817 --> 00:52:38,427
This is the context I need
before I get started every time."

900
00:52:39,585 --> 00:52:40,805
Joe Fabisevich (guest): Yeah,
like if there's, if you don't

901
00:52:40,805 --> 00:52:42,755
know what you're doing, then
you're not gonna have a good time.

902
00:52:42,835 --> 00:52:44,205
It's just the rule in life in general,

903
00:52:44,427 --> 00:52:44,857
Leo Dion (host): Yes

904
00:52:45,115 --> 00:52:47,025
Joe Fabisevich (guest): And I
think that's still true here.

905
00:52:47,045 --> 00:52:50,835
It's just the floor has been
dramatically raised for what you

906
00:52:50,835 --> 00:52:52,755
can do when you don't know anything.

907
00:52:52,835 --> 00:52:56,145
But it hasn't raised the ceiling on
people who don't know what they're doing.

908
00:52:56,285 --> 00:52:58,715
It's raised the ceiling for people
who do know what they're doing

909
00:52:58,935 --> 00:53:04,185
Leo Dion (host): so lastly, I
wanted to talk about WWDC came out a

910
00:53:04,185 --> 00:53:06,898
couple weeks Finally we're in 2024.

911
00:53:07,558 --> 00:53:08,328
We have AI,

912
00:53:11,128 --> 00:53:12,043
artif- Apple Intelligence.

913
00:53:12,043 --> 00:53:14,755
Just I guess before we get into
it, what's kind of your g- general

914
00:53:14,755 --> 00:53:17,615
impressions of WWDC this year?

915
00:53:18,619 --> 00:53:20,819
Joe Fabisevich (guest): Yeah, I
mean, clearly it was very AI focused.

916
00:53:20,819 --> 00:53:22,929
There was almost nothing else
besides performance improvements,

917
00:53:22,929 --> 00:53:25,319
which as a user of iOS, I love.

918
00:53:25,769 --> 00:53:29,859
I'd say I think Apple caught up
to about where we were six months

919
00:53:29,859 --> 00:53:33,229
ago, which is gonna be good enough
for a lot of people and a lot of,

920
00:53:33,269 --> 00:53:35,169
like, developers myself included.

921
00:53:35,169 --> 00:53:39,049
There are… I think if you'd
asked me before WWDC, will I be

922
00:53:39,049 --> 00:53:43,269
using any Apple AI features in my
apps, the answer would've been no.

923
00:53:43,579 --> 00:53:47,029
But I've watched a couple
videos, like WWDCvideos.

924
00:53:48,619 --> 00:53:50,689
I think they've done a lot of
really good work, for example, like

925
00:53:50,689 --> 00:53:54,719
the new evaluations framework for
generating evals for your prompts and

926
00:53:54,785 --> 00:53:57,345
Leo Dion (host): That's the thing
I'm really excited about is e-

927
00:53:57,585 --> 00:53:59,075
evaluations and playing around with that

928
00:54:00,835 --> 00:54:02,915
Joe Fabisevich (guest): Yeah, I was
actually very impressed, where I was like,

929
00:54:03,715 --> 00:54:11,315
"Okay, I think this will do two things."
One is it will your AI, what you build

930
00:54:11,315 --> 00:54:14,325
with AI, more serious in an iOS app.

931
00:54:14,995 --> 00:54:18,315
But it's also going to pull forward
a bunch of iOS developers who have

932
00:54:18,315 --> 00:54:21,065
been like, "I don't really get
this. This doesn't make sense."

933
00:54:21,095 --> 00:54:23,615
They've done a really good job of
presenting the case for why this stuff

934
00:54:23,615 --> 00:54:27,705
matters, where I feel like I've kind
of ventured into the world of the web

935
00:54:27,735 --> 00:54:31,865
to be like… Or like TypeScript or
all these non-iOS environments, to be

936
00:54:31,865 --> 00:54:35,305
like, "No, this clearly makes sense. We
should… As an engineer, I should be

937
00:54:35,305 --> 00:54:36,795
focusing my time and energy on this."

938
00:54:37,225 --> 00:54:42,365
Now as an, you know, as an iOS developer,
I get to say, "Okay, here's, like, how

939
00:54:42,375 --> 00:54:47,095
we do this in an iOS world as well."
And so I really like what they've done.

940
00:54:47,445 --> 00:54:52,705
And I do think that oftentimes the most
important things have to be built up

941
00:54:52,705 --> 00:54:56,725
from the foundation as opposed to, like,
gluing on a new API layer and saying,

942
00:54:56,735 --> 00:54:58,335
like, "We have Apple Intelligence."

943
00:54:58,375 --> 00:55:01,465
I, you know, I worked with App
Intents for the last two years,

944
00:55:01,475 --> 00:55:03,375
and it just felt incomplete.

945
00:55:03,415 --> 00:55:05,615
Now it feels like the story
is much more complete.

946
00:55:06,085 --> 00:55:09,525
I still have to… Like, I basically
have to revisit all the work that

947
00:55:09,525 --> 00:55:15,585
I did for iOS 17 or 18, I forget
which one, and see if that works now.

948
00:55:15,965 --> 00:55:20,145
But I had basically written
off Apple, and I do think that

949
00:55:20,145 --> 00:55:21,085
they've done a good job of…

950
00:55:21,385 --> 00:55:24,265
You know, I've used the new Siri a
couple times and been like, "Yeah,

951
00:55:24,365 --> 00:55:28,875
this'll substitute for, you know,
a Raycast query or a quick open of

952
00:55:28,885 --> 00:55:32,325
ChatGPT to Google." So, like, basically
the equivalent of Google something.

953
00:55:32,615 --> 00:55:35,615
And I think that's what most
consumers are gonna want.

954
00:55:35,665 --> 00:55:36,875
They're not building agents on their phone

955
00:55:37,983 --> 00:55:39,583
Leo Dion (host): Yeah, like
what betas have you installed?

956
00:55:40,583 --> 00:55:42,383
Joe Fabisevich (guest): I
have iOS 27 and macOS 27

957
00:55:43,351 --> 00:55:43,611
Leo Dion (host): Okay.

958
00:55:44,091 --> 00:55:46,481
On your main phone or main Mac or what?

959
00:55:47,385 --> 00:55:48,175
Joe Fabisevich (guest): Yes, I'm stupid

960
00:55:48,641 --> 00:55:48,941
Leo Dion (host): Okay.

961
00:55:48,991 --> 00:55:49,581
That's okay.

962
00:55:49,861 --> 00:55:50,471
I forgive you.

963
00:55:51,001 --> 00:55:52,481
What kind of phone do you have?

964
00:55:53,783 --> 00:55:56,193
Joe Fabisevich (guest): I have a
17, not the Pro, just the regular

965
00:55:56,365 --> 00:55:56,665
Leo Dion (host): Okay.

966
00:55:56,665 --> 00:55:59,594
'Cause I was curious what the difference
in the, like, we talked about this on

967
00:55:59,594 --> 00:56:04,034
the last episode, what's the difference
between, like, the advanced models on

968
00:56:04,034 --> 00:56:07,414
the Pro, 17 Pro as opposed to the 17?

969
00:56:07,414 --> 00:56:11,224
And it doesn't seem clear to me exactly

970
00:56:12,286 --> 00:56:13,246
Joe Fabisevich (guest): Yeah,
I can't really tell you.

971
00:56:13,246 --> 00:56:15,106
I think that there's two kinds of models.

972
00:56:15,106 --> 00:56:19,326
There's one that's like a three billion
parameter model, and for context,

973
00:56:19,326 --> 00:56:24,826
three billion is I don't know, probably
GPT 3.5 to GPT-4 intelligence, right?

974
00:56:25,346 --> 00:56:28,526
And there's a new model that they've
trained that I have not had any

975
00:56:28,536 --> 00:56:31,926
hands-on experience with, but it's
a 20 billion parameter model that's

976
00:56:32,286 --> 00:56:35,996
I believe a mixture of experts,
which means that only certain…

977
00:56:36,436 --> 00:56:40,506
If you imagine, like, you're in a
room with a bunch of experts, and

978
00:56:40,506 --> 00:56:41,956
there's a person who knows science,
and there's a person who knows

979
00:56:41,966 --> 00:56:44,156
math, and there's a person who knows
English, and there's a person who,

980
00:56:44,156 --> 00:56:46,486
like, is is a great programmer.

981
00:56:47,006 --> 00:56:52,136
And you have a question, and you ask,
like, "When exactly did, like, the

982
00:56:52,136 --> 00:56:56,696
US surren- or wh- when exactly did
Germany surrender in World War II?"

983
00:56:57,056 --> 00:56:57,436
Right?

984
00:56:58,076 --> 00:57:01,816
You basically wanna go to the history
expert and say, "Okay, you answer this

985
00:57:01,816 --> 00:57:06,866
question." And so a little layer above
that sits and says, "Who would be the

986
00:57:06,866 --> 00:57:10,986
best to answer this question?" And
evokes two experts at a time, gets their

987
00:57:10,986 --> 00:57:14,956
answer, synthesizes the best result, and
then gives you that answer in return.

988
00:57:15,536 --> 00:57:21,256
So when you have 20 billion parameters,
only one, two, three, four of them

989
00:57:21,256 --> 00:57:25,096
may be active at a time, and that's
those multiple experts that you're

990
00:57:25,096 --> 00:57:27,296
consulting with is the active models.

991
00:57:27,446 --> 00:57:34,486
And so, the 20 billion model, parameter
model is going to be that kind of tool, is

992
00:57:34,486 --> 00:57:39,146
gonna be smarter, but it will not actually
invoke all 20 billion parameters which

993
00:57:39,146 --> 00:57:43,696
would require some number of gigabytes
of RAM, like multiple gigabytes of RAM.

994
00:57:44,136 --> 00:57:47,676
It will instead just consult with
a couple of experts, and then

995
00:57:47,686 --> 00:57:49,456
you still get a better answer

996
00:57:50,592 --> 00:57:51,212
Leo Dion (host): Okay.

997
00:57:51,262 --> 00:57:53,522
That's super helpful 'cause
that makes a lot of sense.

998
00:57:54,052 --> 00:57:57,512
Do you have any thoughts on
private cloud compute as opposed

999
00:57:57,512 --> 00:58:00,152
to, like, the on-device models?

1000
00:58:00,152 --> 00:58:03,702
And what… I guess that goes back to
what I was gonna ask you is, like, what's

1001
00:58:03,702 --> 00:58:10,402
the future of on-device models over,
like, using big data center models?

1002
00:58:12,038 --> 00:58:13,628
Joe Fabisevich (guest): I think
anyone who tells you they have an

1003
00:58:13,638 --> 00:58:15,608
answer is just guessing, right?

1004
00:58:15,608 --> 00:58:19,838
Like, I think it's probably gonna be the
same way with like sometimes you use an

1005
00:58:19,848 --> 00:58:24,148
app that requires the server, and it's
like you can't have Uber without servers.

1006
00:58:24,858 --> 00:58:29,468
And you can't… But also, like, does
your calculator need to go to a server?

1007
00:58:29,578 --> 00:58:29,988
No.

1008
00:58:30,488 --> 00:58:35,458
And so we're all gonna find the right
balance for these tools in the right way.

1009
00:58:35,818 --> 00:58:39,708
So private cloud computing is kind
of like a nice mix where you're

1010
00:58:39,738 --> 00:58:44,758
given privacy promises, and the truth
is they're not fully validatable.

1011
00:58:44,818 --> 00:58:51,148
They are almost completely verifiable
all the way, but at some point you do

1012
00:58:51,208 --> 00:58:53,068
end up having to trust Apple still.

1013
00:58:53,658 --> 00:58:58,018
And so if privacy really matters to
you, that's never gonna be a substitute.

1014
00:58:58,108 --> 00:59:01,348
But for example, like, I don't know,
I have a medical question that I don't

1015
00:59:01,348 --> 00:59:06,838
want ChatGPT to know about, but if I
trust Apple's promises, I may be more

1016
00:59:06,838 --> 00:59:08,238
inclined to use private cloud compute.

1017
00:59:08,618 --> 00:59:11,838
But at the co- like the trade-off
is I know it's not as good of

1018
00:59:11,838 --> 00:59:15,168
a model, so asking it a medical
question may lead to worse medical

1019
00:59:15,168 --> 00:59:17,608
results, and so I'm constantly
doing this mental math in my head.

1020
00:59:17,988 --> 00:59:21,038
I think that most people are not
going to think about this, period.

1021
00:59:21,288 --> 00:59:26,908
And so the reason I say anyone who has,
like, a strong prediction about this, I, I

1022
00:59:26,908 --> 00:59:31,558
don't inherent- I think they're guessing,
is because no one can get in the mind of,

1023
00:59:31,558 --> 00:59:37,148
like, 6 billion consumers or 8 billion
consumers and make that singular decision.

1024
00:59:37,238 --> 00:59:40,268
Or, like, singul- you can place your
bets, but I really don't think that it's

1025
00:59:40,268 --> 00:59:44,418
going to… You're gonna have to play
it by ear as the technology develops.

1026
00:59:44,736 --> 00:59:46,756
Leo Dion (host): I mean, do
you think that there's a future

1027
00:59:46,756 --> 00:59:48,886
for on-device models, period?

1028
00:59:48,896 --> 00:59:50,346
I mean, I would say so.

1029
00:59:50,406 --> 00:59:50,746
Yeah.

1030
00:59:50,746 --> 00:59:51,096
Okay.

1031
00:59:51,716 --> 00:59:52,086
Yeah

1032
00:59:53,466 --> 00:59:55,306
Joe Fabisevich (guest): But what will
they do and what will we ask them

1033
00:59:55,360 --> 00:59:56,510
Leo Dion (host): Yeah, obviously.

1034
00:59:56,556 --> 00:59:57,166
Joe Fabisevich (guest): absolutely no idea

1035
00:59:57,560 --> 00:59:58,910
Leo Dion (host): Yeah,
that makes total sense.

1036
00:59:59,640 --> 01:00:03,850
So were there any APIs that came
out a few weeks ago that you're

1037
01:00:03,850 --> 01:00:07,360
like, "Oh, this is exciting. I'm
excited to try this out in Plinky"?

1038
01:00:08,872 --> 01:00:09,252
Joe Fabisevich (guest): Yeah.

1039
01:00:09,252 --> 01:00:12,732
So this is a little bit of, like,
a different take on AI per se, but

1040
01:00:13,142 --> 01:00:16,742
Plinky is obviously an app that saves
your links for later, but I also view

1041
01:00:16,742 --> 01:00:18,052
it as, like, your home for links.

1042
01:00:18,092 --> 01:00:21,512
And so if you are in your home, you get
to design your home however you want.

1043
01:00:21,742 --> 01:00:24,512
And so what I'm trying to do
is build, like, a completely

1044
01:00:24,512 --> 01:00:28,422
customizable link-saving
experience that works for anyone.

1045
01:00:28,722 --> 01:00:32,772
And one of those things is I wanna…
Right now when you, like, long press or

1046
01:00:32,772 --> 01:00:35,872
swipe on a link, you can set a reminder,
you can add it to a folder, you can

1047
01:00:36,152 --> 01:00:40,142
share it, you can do all these things
with the link and you… What… How

1048
01:00:40,142 --> 01:00:41,712
it works for you is totally up to you.

1049
01:00:42,092 --> 01:00:45,482
But going beyond that,
I wanna build a new…

1050
01:00:45,502 --> 01:00:50,352
I'm building a new feature called Pro
Actions which lets you take a link and

1051
01:00:50,352 --> 01:00:55,852
you can run a shortcut or you can make an
API call with the contents of that link.

1052
01:00:55,892 --> 01:00:59,762
Or now because of Private Cloud
Compute, I was already planning on

1053
01:00:59,762 --> 01:01:04,412
building this out using like, you
know, the GPT 5.5 and Claude and

1054
01:01:04,422 --> 01:01:05,932
all these models that are available.

1055
01:01:06,392 --> 01:01:12,412
But I think I'm actually gonna build,
like, to start with Private Cloud Compute

1056
01:01:12,552 --> 01:01:17,162
and, you know, all the features that are
built into the iOS device, your ability

1057
01:01:17,162 --> 01:01:22,332
to interact with AI as one of these Pro
Actions when you long press on a link or

1058
01:01:22,332 --> 01:01:24,552
swipe a link or do anything with a link.

1059
01:01:24,552 --> 01:01:28,682
Like, if you're in reader mode,
you can say take this article and

1060
01:01:28,692 --> 01:01:30,662
obviously you can say something
like summarize it, right?

1061
01:01:30,762 --> 01:01:32,942
Super, super common thing that people do.

1062
01:01:33,252 --> 01:01:35,892
But I wanna go many steps beyond that.

1063
01:01:35,942 --> 01:01:40,452
So imagine you are reading an article and
you finish it and it's like a, I don't

1064
01:01:40,452 --> 01:01:46,842
know, daring fireball piece on the newest
iPhone and John is very poetic and prosaic

1065
01:01:46,852 --> 01:01:48,512
about how this iPhone changes everything.

1066
01:01:49,092 --> 01:01:53,512
And then you can just click one button
and generate like a counterpoint.

1067
01:01:53,572 --> 01:01:58,242
Like take this person, write it from
another person's perspective, and then

1068
01:01:58,322 --> 01:02:04,382
write an article on demand that argues
against the thesis of this article

1069
01:02:04,622 --> 01:02:08,562
at the same grade level with the same
writing style to make sure that I can

1070
01:02:08,562 --> 01:02:10,422
actually get a fully rounded picture.

1071
01:02:10,722 --> 01:02:15,992
That's one example of where AI actually
goes above and beyond and is able…

1072
01:02:16,042 --> 01:02:19,712
Like we're tapping into the generative
nature of AI to say, what experiences can

1073
01:02:19,712 --> 01:02:21,272
we build that we couldn't build before?

1074
01:02:21,792 --> 01:02:24,442
This was already kind of in my plans,
but it's a little bit accelerated knowing

1075
01:02:24,442 --> 01:02:29,122
that I can get to a working version
quicker because I have Private Cloud

1076
01:02:29,122 --> 01:02:32,422
Compute without having to build out my
own stack, which has been like a work

1077
01:02:32,422 --> 01:02:33,592
in progress for the last few months.

1078
01:02:34,520 --> 01:02:35,950
Leo Dion (host): Yeah,
that sounds amazing.

1079
01:02:36,000 --> 01:02:37,720
I'm excited to, to see that.

1080
01:02:38,430 --> 01:02:44,440
What, as far as your, I don't know
if you've had time, but the updates

1081
01:02:44,490 --> 01:02:48,560
to Xcode and some of the stuff that's
been added, like skills and things like

1082
01:02:48,560 --> 01:02:52,600
that, have you had time to even look
at any of that stuff, and have you been

1083
01:02:52,600 --> 01:02:54,940
like at all, "Oh, that's really cool.

1084
01:02:54,940 --> 01:02:57,130
I'm gonna add that to my workflow"?

1085
01:02:58,622 --> 01:02:58,932
Joe Fabisevich (guest): Yeah.

1086
01:02:59,012 --> 01:03:01,522
I've only looked and I
haven't played with it at all.

1087
01:03:01,522 --> 01:03:05,282
It's been, you know, WWDCwas
10 days ago, nine days ago, and

1088
01:03:05,282 --> 01:03:06,502
there's been so much to take in.

1089
01:03:07,552 --> 01:03:07,912
Yeah.

1090
01:03:07,922 --> 01:03:11,822
But I would say Apple seems to have
done a decent job, and I think that,

1091
01:03:11,822 --> 01:03:16,452
like, it's not a great job by any means
because they're still playing catch up

1092
01:03:16,872 --> 01:03:21,602
and they're still moving slower than
everyone else is in a lot of ways.

1093
01:03:21,602 --> 01:03:25,972
And so for example, they added a plan
mode, and the plan mode looks okay,

1094
01:03:26,282 --> 01:03:29,642
and I think it's like, it's gonna do
the job, but it's not gonna give a

1095
01:03:29,642 --> 01:03:34,472
person who has never done AI the best
impression of what is p- already possible.

1096
01:03:34,912 --> 01:03:40,102
And so I don't think I will be using
Xcode any more than I use it today,

1097
01:03:40,102 --> 01:03:44,022
which is to say that, like, I drop
into Xcode to, like, make small changes

1098
01:03:44,022 --> 01:03:47,092
and review some diffs where it's like
I've just gotten so used to looking

1099
01:03:47,092 --> 01:03:50,732
at code in Xcode that when I want to
look at Swift code, I think Xcode.

1100
01:03:51,372 --> 01:03:56,222
But I, I think that they've done a
good job of, you know, like I said,

1101
01:03:56,222 --> 01:03:58,922
catching up to where we were six
months ago on the development front.

1102
01:03:59,352 --> 01:04:01,532
I think it's probably closer
to, like, nine months ago.

1103
01:04:01,582 --> 01:04:04,272
I'm just kind of, like, spitballing
here, but it's a little more

1104
01:04:04,272 --> 01:04:08,092
behind than the consumer
experience of AI that they shipped

1105
01:04:09,050 --> 01:04:13,100
Leo Dion (host): I, and I also feel like,
you know, I would say we have a decent

1106
01:04:13,110 --> 01:04:15,520
expertise at AI-assisted development.

1107
01:04:15,540 --> 01:04:19,630
We're kind of stuck in our ways, and
we like the way we do things now.

1108
01:04:20,200 --> 01:04:23,280
It's gonna be really hard for
someone like me to want to go

1109
01:04:23,280 --> 01:04:25,120
and, like, use these Xcode tools.

1110
01:04:25,430 --> 01:04:32,070
And, like, I tried the beta out
probably a few days after it came out

1111
01:04:32,100 --> 01:04:36,810
for Xcode, and, like, there was some
things that did okay, and then there

1112
01:04:36,810 --> 01:04:39,620
was a thing where it just didn't work.

1113
01:04:39,700 --> 01:04:44,280
Gave me a failure error message, and I
was like, "Okay, gonna … Like, I'm done.

1114
01:04:44,320 --> 01:04:44,970
I've tried it.

1115
01:04:45,500 --> 01:04:50,030
Back to going to Fable to write my
SwiftUI app based on my new…" Like,

1116
01:04:50,070 --> 01:04:53,850
'cause there was a, what is it, like,
the reordering and swipe action stuff

1117
01:04:53,850 --> 01:04:57,380
I wanted to play around with, and Fable
did perfectly fine with it, and it

1118
01:04:57,380 --> 01:05:00,360
was able to read the new documentation
and understand how to use it.

1119
01:05:00,360 --> 01:05:04,720
And like, okay, guess I'm not going
back to Xcode for doing AI stuff.

1120
01:05:04,720 --> 01:05:06,110
It's just not worth it to me.

1121
01:05:06,650 --> 01:05:07,840
So, yeah.

1122
01:05:07,950 --> 01:05:08,560
I mean, that's kind of

1123
01:05:08,698 --> 01:05:09,048
Joe Fabisevich (guest): the opposite

1124
01:05:10,170 --> 01:05:10,830
Leo Dion (host): Okay, go ahead.

1125
01:05:10,958 --> 01:05:11,468
Joe Fabisevich (guest): true, right?

1126
01:05:11,968 --> 01:05:16,648
I think the opposite is also true,
where, like, if you have not been AI

1127
01:05:16,648 --> 01:05:22,388
pilled and you are first getting into
this, then the Xcode version is going

1128
01:05:22,388 --> 01:05:25,918
to be a decent gateway into like, "Oh,
okay, I'm using more AI features."

1129
01:05:26,188 --> 01:05:30,688
I've even seen that with the Xcode 26
features, where it's like people who

1130
01:05:30,908 --> 01:05:35,328
are long, like lifelong iOS developers
are like, "Okay, I tried this.

1131
01:05:35,328 --> 01:05:38,158
I could see where it's
potential." And I think over

1132
01:05:38,158 --> 01:05:40,068
Xcode 26, it's a pretty big jump.

1133
01:05:40,298 --> 01:05:44,878
And so they will kind of have a little
bit of like the feeling, the AGI moment,

1134
01:05:44,918 --> 01:05:49,068
where they say, "Okay, like I can see
how this is progressing." And it may

1135
01:05:49,068 --> 01:05:55,418
not be as fast as you and I want using
Codex or Fable or whatever, but it may…

1136
01:05:55,708 --> 01:06:00,088
You know, we don't… It's
not like this like curved, you

1137
01:06:00,088 --> 01:06:01,658
know, angled line like this.

1138
01:06:01,868 --> 01:06:05,188
I feel like AI is often like we hit
a plateau, and we're like, "Wow." And

1139
01:06:05,188 --> 01:06:08,008
then we just get used to being on that
plateau, and then two months later,

1140
01:06:08,008 --> 01:06:10,898
a new thing comes out, and we reach a
new plateau, and we keep going from…

1141
01:06:11,158 --> 01:06:14,968
Like, we're stably on plateaus for a
couple months, and we get used to that.

1142
01:06:15,008 --> 01:06:18,568
And it actually looks like a
curve when you project it out

1143
01:06:18,568 --> 01:06:19,718
over a long period of time.

1144
01:06:19,778 --> 01:06:21,018
It looks like it's like
rocket ship growth.

1145
01:06:21,688 --> 01:06:24,788
But when you're actually
living it, you're on a plateau

1146
01:06:26,020 --> 01:06:26,360
Leo Dion (host): Yeah.

1147
01:06:26,370 --> 01:06:27,090
I know what you mean.

1148
01:06:27,560 --> 01:06:32,960
Well, and, but I think, like, with people
themselves, it's very much like, like I

1149
01:06:32,960 --> 01:06:36,490
said, like, I started off with ChatGPT
and then Cursor and then Claude Code.

1150
01:06:36,510 --> 01:06:41,160
It's sort of like getting on a bike and,
like, being able to have the confidence to

1151
01:06:41,160 --> 01:06:43,800
go faster with tools that can work faster.

1152
01:06:44,250 --> 01:06:48,670
And, like, you personally, your journey
is, like, more, like, accelerating and

1153
01:06:48,680 --> 01:06:53,370
being able to understand the, like,
jagged parts of AI and how to optimize it

1154
01:06:54,370 --> 01:06:57,370
Joe Fabisevich (guest): There's some
element to that I think that maybe as

1155
01:06:57,370 --> 01:07:00,900
the technology progresses, things get
easier and the curve won't feel as jagged.

1156
01:07:01,090 --> 01:07:05,270
It's gonna feel jagged at the frontier
always probably, or I'm guessing it

1157
01:07:05,270 --> 01:07:07,210
will always feel jagged at the frontier.

1158
01:07:07,650 --> 01:07:10,240
But the way that most people
use technology is not that

1159
01:07:10,240 --> 01:07:13,740
they're, like, doing the most
bleeding edge thing all the time.

1160
01:07:14,340 --> 01:07:18,810
If this thing is really out there curing
cancer, you and I are not going to be

1161
01:07:18,820 --> 01:07:23,990
using the bleeding edge model to write
our SwiftUI app, if that makes sense.

1162
01:07:24,300 --> 01:07:24,740
And so

1163
01:07:24,802 --> 01:07:25,112
Leo Dion (host): total sense

1164
01:07:25,500 --> 01:07:28,820
Joe Fabisevich (guest): Our experience
of it is going to get less and less

1165
01:07:29,500 --> 01:07:30,910
I mean, you use the term jagged.

1166
01:07:31,140 --> 01:07:32,770
It's gonna get less and less chaotic.

1167
01:07:33,080 --> 01:07:36,050
But it will still be always chaotic.

1168
01:07:36,060 --> 01:07:38,990
It's just gonna be like how
much of that chaos is felt by

1169
01:07:39,020 --> 01:07:43,160
us directly, which is also the
argument for, like, cheaper models.

1170
01:07:43,820 --> 01:07:50,160
Models will get cheaper because we will
not be using Fable to build an HTML page.

1171
01:07:50,160 --> 01:07:54,980
We will be like, "Opus is good enough,"
and… Or Fable Haiku is as good

1172
01:07:54,980 --> 01:07:58,970
as Opus 4.8 Claude 5 Haiku, right?

1173
01:07:59,210 --> 01:08:01,700
And so we're just gonna be like,
"Well, I don't need that, and

1174
01:08:01,700 --> 01:08:04,500
I'm not gonna spend those tokens
or that money on those tokens."

1175
01:08:04,868 --> 01:08:04,998
Leo Dion (host): Yep.

1176
01:08:05,828 --> 01:08:06,858
Yeah, exactly.

1177
01:08:07,378 --> 01:08:11,308
So speaking of spending tokens, I don't
know if you saw, I built a app month bar

1178
01:08:11,338 --> 01:08:16,648
specifically so I can track not my tokens,
but how much is left in the month, because

1179
01:08:16,648 --> 01:08:18,938
I was using Copilot at my worky work.

1180
01:08:19,578 --> 01:08:23,918
But yeah, I'm trying to figure out how
I can actually track people's like,

1181
01:08:24,818 --> 01:08:28,728
not tokens so much as like their usage
remaining in a time period and be able

1182
01:08:28,728 --> 01:08:30,828
to like show that in the menu bar.

1183
01:08:30,838 --> 01:08:34,098
I haven't gotten that far yet, but yeah,
that's like gonna be a big thing is

1184
01:08:34,098 --> 01:08:38,658
being able to like keep an eye on that
and when you can- when your what is it?

1185
01:08:38,778 --> 01:08:41,758
Your period of time
gets reset, so to speak

1186
01:08:42,982 --> 01:08:43,422
Joe Fabisevich (guest): Yeah.

1187
01:08:43,452 --> 01:08:46,912
And I mean, I know it's built into
Codex, so I check that all the time

1188
01:08:47,022 --> 01:08:48,542
rather than like having a menu bar app.

1189
01:08:48,562 --> 01:08:51,482
I did have Peter
Steinberger had a Codex bar.

1190
01:08:51,622 --> 01:08:53,182
It is an app that does that.

1191
01:08:53,562 --> 01:08:56,552
And I was using it, and then I
was like, honestly, it's just not

1192
01:08:56,552 --> 01:08:58,392
even that big of a problem for me.

1193
01:08:58,642 --> 01:09:02,012
But as the economics of AI
change, I suspect it will become

1194
01:09:02,122 --> 01:09:05,282
a problem either more or at times

1195
01:09:05,764 --> 01:09:07,794
Leo Dion (host): Do you think, like,
do you think there's gonna be a bubble

1196
01:09:07,794 --> 01:09:09,994
where, like, these prices are gonna go up?

1197
01:09:10,994 --> 01:09:14,164
Joe Fabisevich (guest): I think that it's
too hard to predict the actual… Like,

1198
01:09:14,174 --> 01:09:18,494
what I can tell you is there's more demand
than there is supply, and when there's

1199
01:09:18,504 --> 01:09:20,254
more demand than supply, prices go up.

1200
01:09:20,694 --> 01:09:26,154
But how much we can allocate
towards supply is a question

1201
01:09:26,164 --> 01:09:28,244
of many factors, right?

1202
01:09:28,244 --> 01:09:29,954
It's like there's a political factor.

1203
01:09:29,994 --> 01:09:32,194
People don't want data
centers being built right now.

1204
01:09:32,194 --> 01:09:35,624
It takes a long time to build something
in this world, and so you can't just

1205
01:09:35,634 --> 01:09:40,774
say, "Okay, there's new demand." And
in the digital world, we're so used to

1206
01:09:41,214 --> 01:09:44,554
being able to satisfy demand by going
and buying a new server, and that's

1207
01:09:44,554 --> 01:09:46,014
another million users we can serve.

1208
01:09:46,466 --> 01:09:46,676
Leo Dion (host): Great

1209
01:09:46,744 --> 01:09:48,634
Joe Fabisevich (guest): If you have
to build a data center, it's going

1210
01:09:48,634 --> 01:09:52,404
to take 18 months at the minimum, and
probably closer to multiple years.

1211
01:09:52,704 --> 01:09:55,834
And so I can't predict the price
because I can't predict what will

1212
01:09:55,834 --> 01:09:59,934
happen, but I feel very confident
that demand will continue to rise.

1213
01:10:00,304 --> 01:10:05,394
So depending on how supply is
constrained or opened up the price

1214
01:10:05,394 --> 01:10:06,524
will follow depending on that.

1215
01:10:06,804 --> 01:10:11,704
But also, in technology, we have
spent the last 40 years not really

1216
01:10:11,704 --> 01:10:13,474
facing, like, an optimization crunch.

1217
01:10:14,124 --> 01:10:17,094
We oftentimes are just like, we
can solve the problem with better

1218
01:10:17,104 --> 01:10:18,954
compute, more hardware, et cetera.

1219
01:10:19,414 --> 01:10:23,304
So one of the things that was kind
of, like, revolutionary about DeepSeek

1220
01:10:23,344 --> 01:10:27,274
was not its… The fact that it was
a better model, it was that it was

1221
01:10:27,274 --> 01:10:31,134
dramatically cheaper because they
invented a new mechanism for caching.

1222
01:10:31,644 --> 01:10:35,034
And so, like they did a… i'm pretty
sure it was DeepSeek that came up

1223
01:10:35,034 --> 01:10:37,124
with new techniques for KV caching.

1224
01:10:37,124 --> 01:10:39,924
But it made serving models cheaper, right?

1225
01:10:40,014 --> 01:10:47,104
And so we have been racing towards,
like, improving the frontier that

1226
01:10:47,144 --> 01:10:50,924
we don't even spend the time to
optimize what we have right now.

1227
01:10:51,324 --> 01:10:56,764
And so if there is some sort of crunch
that would drive token prices up- AI

1228
01:10:56,794 --> 01:11:01,754
labs may just reallocate their thought
on how to be more efficient and in

1229
01:11:01,754 --> 01:11:03,944
a way to keep token prices down.

1230
01:11:04,094 --> 01:11:08,364
So I am watching this just like
everyone else is in that regard.

1231
01:11:08,624 --> 01:11:13,814
I just think that we are not being
particularly clever or we're not being

1232
01:11:13,814 --> 01:11:17,674
particularly open-minded to the idea that
there are many ways that things may go.

1233
01:11:17,674 --> 01:11:22,114
Like, oftentimes when we're on… You
know, when we're like, we see some

1234
01:11:22,114 --> 01:11:26,284
news happen, we create an immediate
reaction to it, and create an immediate,

1235
01:11:26,284 --> 01:11:29,774
like, narrative around it, and then
one day later that's invalidated.

1236
01:11:30,304 --> 01:11:35,334
And so rather than trying to figure out
how to respond to the news of the day,

1237
01:11:35,584 --> 01:11:38,464
it's more important to develop, like,
the foundations and fundamentals to

1238
01:11:38,464 --> 01:11:40,974
figure out, like, okay, what does this
mean in the grand scheme of things?

1239
01:11:41,234 --> 01:11:44,054
So rather than making predictions,
I'm like, "Hey, there's like a million

1240
01:11:44,054 --> 01:11:48,214
ways this can go, and here's ways
that I foresee it w- will possibly

1241
01:11:48,214 --> 01:11:51,814
go." And I try and keep those in my
mind all at once rather than saying,

1242
01:11:51,814 --> 01:11:53,654
like, the price of AI will go up.

1243
01:11:53,764 --> 01:11:57,334
For example, the price of serving
an open source model is probably 80%

1244
01:11:57,334 --> 01:11:59,834
cheaper than serving a GPT quality model.

1245
01:12:00,194 --> 01:12:04,464
And so maybe we will just choose
to use open source models because

1246
01:12:04,464 --> 01:12:07,664
they're good enough at creating
our websites or creating our apps.

1247
01:12:08,214 --> 01:12:13,114
And so that would change and invert
the dynamics of the economics 100%.

1248
01:12:13,184 --> 01:12:16,334
It would just completely flip the
way that we currently think about

1249
01:12:16,334 --> 01:12:18,744
AI and what the price of a token is.

1250
01:12:18,814 --> 01:12:22,044
And so, yeah I think I've
said this like four times.

1251
01:12:22,044 --> 01:12:23,754
I'm like, I'm not in the
predictions business.

1252
01:12:23,754 --> 01:12:27,574
I'm just trying to, like, trying
to ride this wave with everyone

1253
01:12:27,574 --> 01:12:29,464
else and h- be one step ahead

1254
01:12:29,854 --> 01:12:31,624
Leo Dion (host): I think there's
a lot we don't know that we don't

1255
01:12:31,624 --> 01:12:33,244
know when it comes to this stuff.

1256
01:12:33,644 --> 01:12:37,724
So like anytime, like that's my thing
with like this whole like, oh, doomsday

1257
01:12:37,724 --> 01:12:43,524
type scenario with AI is like we, we
don't… I don't think we even understand

1258
01:12:43,524 --> 01:12:45,474
what intelligence actually means.

1259
01:12:45,984 --> 01:12:50,744
And I feel like over the next few decades
we'll be like, "Man, back in the old

1260
01:12:50,744 --> 01:12:54,114
days we thought intelligence was this
thing that you could do artificially."

1261
01:12:54,124 --> 01:12:58,394
And it's like, no, like LLMs are good
and these models are good at very

1262
01:12:58,394 --> 01:13:03,694
specific things that is a subset of
what I would qualify as intelligence.

1263
01:13:04,034 --> 01:13:08,354
But there are other pieces that humans
provide, like wisdom and experience,

1264
01:13:08,454 --> 01:13:13,994
that I think like we, w- I think
we'll value more in the future because

1265
01:13:14,394 --> 01:13:17,504
we've been able to take this thing
we call intelligence, and turn it

1266
01:13:17,504 --> 01:13:19,614
into software and make it a product.

1267
01:13:19,654 --> 01:13:23,944
But there's gonna be other pieces
that people will provide that I think

1268
01:13:23,974 --> 01:13:29,414
will be more important now that we're
able to put it in a machine, I guess.

1269
01:13:29,464 --> 01:13:30,544
Does that kind of make sense?

1270
01:13:30,604 --> 01:13:34,734
I don't know how to… But I
think there's a, just a lot we

1271
01:13:34,734 --> 01:13:35,904
don't know that we don't know.

1272
01:13:35,914 --> 01:13:40,064
Like I, I'm a big fan of history and like
with, when you look at like the Industrial

1273
01:13:40,064 --> 01:13:44,434
Revolution, we just kind of, we had no
idea what the future would look like.

1274
01:13:44,484 --> 01:13:47,744
And like, I think we kind of
think everything, it's a straight

1275
01:13:47,744 --> 01:13:50,454
line that goes up, but that's
not the way it works usually.

1276
01:13:51,666 --> 01:13:52,116
Joe Fabisevich (guest): Yeah.

1277
01:13:52,206 --> 01:13:53,036
There's two thoughts I have.

1278
01:13:53,036 --> 01:13:58,166
One is I often live by F. Scott
Fitzgerald's quote, which is, "The test

1279
01:13:58,166 --> 01:14:02,296
of a first-rate intelligence is the
ability to hold two opposed ideas in the

1280
01:14:02,296 --> 01:14:06,826
mind at the same time and still retain
the ability to function." It's like, so

1281
01:14:06,926 --> 01:14:09,706
we don't know what we don't know, and
there are things that we don't know how

1282
01:14:09,706 --> 01:14:14,546
to connect them, and so we have to just
accept that there are two things that

1283
01:14:14,596 --> 01:14:20,016
are very different and are both true,
and figure out how they can both be true.

1284
01:14:20,076 --> 01:14:23,716
That's just kind of what the act of living
is to figure out how these things are.

1285
01:14:24,066 --> 01:14:28,766
And to, like, provide an example,
you know, Noam Chomsky spent his

1286
01:14:28,766 --> 01:14:32,416
whole, like, life thinking about
linguistics and theory of mind and

1287
01:14:32,416 --> 01:14:36,766
connecting them, and now we have this,
like, really strong counter example.

1288
01:14:37,096 --> 01:14:42,666
You know, you can say it's like
language is thought, or that to

1289
01:14:42,666 --> 01:14:46,816
think is to use language, or language
is a universal human property.

1290
01:14:47,766 --> 01:14:52,476
And, you know, in the last three or
four years, we have this counter example

1291
01:14:52,476 --> 01:14:54,196
that proves that's just not true.

1292
01:14:54,296 --> 01:15:00,696
And it may be true in some ways, but
as an overarching thesis, it's not

1293
01:15:00,716 --> 01:15:04,056
inherently true that language is thought
or that language is inherently human.

1294
01:15:04,356 --> 01:15:07,996
Yes, it was… Like, you could say a
large language model is trained on human

1295
01:15:07,996 --> 01:15:13,056
language and human thought, but that
doesn't mean that the way that we embody

1296
01:15:13,056 --> 01:15:15,756
it will be uniquely human going forward.

1297
01:15:16,026 --> 01:15:19,236
And so that changes the way that
I think about this entirely,

1298
01:15:19,516 --> 01:15:21,866
and I just don't have answers.

1299
01:15:21,866 --> 01:15:24,026
I'm just trying to figure all
this stuff out as I go as well,

1300
01:15:24,046 --> 01:15:24,906
the same way that you said.

1301
01:15:25,196 --> 01:15:28,436
And being comfortable with being
uncomfortable, I think that being

1302
01:15:28,436 --> 01:15:32,656
comfortable with being uncomfortable
is going to be, like, a legitimately

1303
01:15:32,686 --> 01:15:38,696
necessary and useful skill in the future
of being a human being or the economy

1304
01:15:38,706 --> 01:15:41,496
or whatever you want to think about it.

1305
01:15:41,836 --> 01:15:45,176
It's just gonna be more and
more important by the day to be

1306
01:15:45,196 --> 01:15:46,526
okay with being uncomfortable.

1307
01:15:46,526 --> 01:15:50,236
And the truth is that, like, you and
I, we're, like, recording a podcast.

1308
01:15:50,236 --> 01:15:52,976
We're happy and healthy, and
we have a home and all that.

1309
01:15:53,256 --> 01:15:57,116
A lot of people dealing with the
discomforts despite, like, you

1310
01:15:57,116 --> 01:16:01,456
know, even in modern living, life is
not inherently comfortable, right?

1311
01:16:01,686 --> 01:16:04,506
Or, like, I think about my parents who
were, like, refugees from the Soviet

1312
01:16:04,506 --> 01:16:07,546
Union, and it's like they didn't know
what their next day would bring them.

1313
01:16:07,576 --> 01:16:11,196
And I think that we just have to be
a little more okay with the fact that

1314
01:16:11,206 --> 01:16:13,156
our lives will have elements of that.

1315
01:16:13,496 --> 01:16:17,336
And I think that's the part that actually
makes everyone really, like, anxious

1316
01:16:17,336 --> 01:16:22,766
about AI, is what happens when that
core assumption isn't baked in anymore?

1317
01:16:23,186 --> 01:16:27,466
And having grown up at home watching
that, I realize, you know, it's like it's

1318
01:16:27,466 --> 01:16:29,646
never been baked in, for me at least.

1319
01:16:29,736 --> 01:16:34,436
And but that's not to say, like,
"Oh, it's all bad." It's to say

1320
01:16:34,436 --> 01:16:37,746
that human flourishing and, like,
a lot of beauty still happens

1321
01:16:37,756 --> 01:16:39,526
even when there is discomfort.

1322
01:16:39,546 --> 01:16:42,256
And so we just gotta take
things one day at a time.

1323
01:16:42,656 --> 01:16:43,096
Leo Dion (host): Yeah.

1324
01:16:43,716 --> 01:16:49,356
The, sorry to be a downer, but I also
think, like, all the stuff, like the,

1325
01:16:50,186 --> 01:16:54,216
like, "Oh, the future is this, the future
is that," I think we, we go to b- worst

1326
01:16:54,216 --> 01:16:59,026
case scenarios when we're just not even
paying attention to, like, the issues that

1327
01:16:59,036 --> 01:17:05,196
AI already presents to us on a daily life,
to actual, like, real people nowadays.

1328
01:17:05,316 --> 01:17:09,666
Just, like, with people being more
isolated and people, like… Or, you

1329
01:17:09,666 --> 01:17:13,916
know, just the kind of the issues there
that have come up with AI or people not

1330
01:17:13,936 --> 01:17:18,686
understanding how to take what it says and
understand that what it actually means.

1331
01:17:19,526 --> 01:17:21,156
Oh, 100% like the internet.

1332
01:17:21,206 --> 01:17:21,586
Yeah.

1333
01:17:22,196 --> 01:17:22,506
Yeah

1334
01:17:22,694 --> 01:17:26,134
Joe Fabisevich (guest): I contend that we
have not caught up to the internet yet.

1335
01:17:26,144 --> 01:17:30,704
Like, we had this technological change
that was 25 to 30 years ago where it was

1336
01:17:30,704 --> 01:17:34,954
like, okay, suddenly things are connected,
and we still haven't grappled with how

1337
01:17:34,954 --> 01:17:38,184
to deal with that, and now we have this,
like, accelerant that we've thrown in.

1338
01:17:38,384 --> 01:17:42,764
And so we're moving on to the next problem
before we even solve the last one, because

1339
01:17:42,824 --> 01:17:44,684
that's the pace of technological change.

1340
01:17:44,894 --> 01:17:49,984
And so we're living a very compressed
version of that change when we haven't

1341
01:17:49,984 --> 01:17:53,554
even figured out how to deal with that
change, and are trying to come up with

1342
01:17:53,564 --> 01:17:58,884
answers on the fly that are going to
be resilient to whatever comes up.

1343
01:17:59,354 --> 01:18:01,964
And that's why I'm saying, like, I
think it's just more important to be

1344
01:18:02,474 --> 01:18:04,984
comfortable with being uncomfortable
and not having every answer.

1345
01:18:05,774 --> 01:18:05,994
Leo Dion (host): Yep.

1346
01:18:06,774 --> 01:18:11,654
Yeah, I mean, to me it's like, it reminds
me of like when I, you know, in history

1347
01:18:11,654 --> 01:18:15,374
when the printing press came out, and
just any old schmo could come out with a

1348
01:18:15,374 --> 01:18:17,514
pamphlet with all sorts of crazy things.

1349
01:18:17,944 --> 01:18:20,754
Like, that's essentially what
social media is right now, and

1350
01:18:20,754 --> 01:18:25,214
being able to understand how to like
sift between truth and fake news.

1351
01:18:25,244 --> 01:18:28,564
I hate that term 'cause it's overused,
but essentially that's what it is, right?

1352
01:18:29,608 --> 01:18:29,968
Joe Fabisevich (guest): Yeah.

1353
01:18:30,218 --> 01:18:36,148
We, and also like, I mean, just to take
it into a biological realm, but like we

1354
01:18:36,158 --> 01:18:40,058
aren't designed biologically to be dealing
with this many inputs all the time.

1355
01:18:40,058 --> 01:18:43,768
So at least the printing press was
like, okay, it took this guy like a day

1356
01:18:43,768 --> 01:18:45,148
to put together all these pamphlets.

1357
01:18:45,488 --> 01:18:52,678
But now I see like 10 things a
week that previously would've been

1358
01:18:52,998 --> 01:18:56,328
unfathomable and would've lasted
in my mind for like a month, right?

1359
01:18:56,978 --> 01:18:59,568
And I'm not equipped to handle this.

1360
01:18:59,578 --> 01:19:02,508
I wrote an essay about this called
Artifacts, but the premise is that like

1361
01:19:03,508 --> 01:19:08,458
people talked about the moon landing for
decades, and then you remember the dress?

1362
01:19:09,288 --> 01:19:09,808
Leo Dion (host): Yes.

1363
01:19:09,898 --> 01:19:10,908
Yeah, what color is it?

1364
01:19:10,908 --> 01:19:11,218
Yeah

1365
01:19:12,066 --> 01:19:12,546
Joe Fabisevich (guest): Exactly.

1366
01:19:12,846 --> 01:19:14,536
And then do you remember the two llamas?

1367
01:19:15,188 --> 01:19:16,728
Leo Dion (host): No, I don't
remember the two llamas

1368
01:19:17,652 --> 01:19:18,642
Joe Fabisevich (guest):
Okay, that's my point.

1369
01:19:18,832 --> 01:19:21,832
So the two llamas was this, like,
viral moment that happened where these

1370
01:19:21,832 --> 01:19:24,972
two llamas escaped from a zoo, and it
happened on the same day as the dress.

1371
01:19:24,982 --> 01:19:27,652
Three hours later, the dress
happened, and everyone forgot

1372
01:19:27,652 --> 01:19:30,302
about the llamas, and that's kind
of what, like, life is like today.

1373
01:19:30,312 --> 01:19:31,852
It's like one crazy thing
happens after another.

1374
01:19:32,232 --> 01:19:35,942
We spent months talking… or
years, decades talking about

1375
01:19:36,162 --> 01:19:37,462
an event like the moon landing.

1376
01:19:37,862 --> 01:19:41,132
We talk about, you know, there
was this, like, there was this

1377
01:19:41,412 --> 01:19:44,902
invasion by the Mongols where they
destroyed, I forget which library.

1378
01:19:44,932 --> 01:19:47,422
I have it in my essay, but it's,
like, this beautiful library that

1379
01:19:47,422 --> 01:19:48,582
the Persian Empire had built.

1380
01:19:48,912 --> 01:19:52,592
And there was a line that said,
like, the Tigris ran black

1381
01:19:52,592 --> 01:19:54,482
with ink as if it was blood.

1382
01:19:54,852 --> 01:19:57,332
And they just, like, took it so
seriously that it was literally a

1383
01:19:57,332 --> 01:20:03,972
century's worth of thought was put
into what happened and caused that.

1384
01:20:04,322 --> 01:20:07,162
We don't spend… We don't have
things that last centuries.

1385
01:20:07,162 --> 01:20:10,052
We don't even have things that last
more than a few minutes these days.

1386
01:20:10,312 --> 01:20:15,932
And so we're just not designed to
deal with all the stimulus, and

1387
01:20:16,612 --> 01:20:18,792
I'm not trying to say like, "Oh,
ban social media," or whatever.

1388
01:20:19,092 --> 01:20:22,592
I'm just saying that, like, we're
trying to say, like, how do we fix AI

1389
01:20:22,742 --> 01:20:26,262
when we haven't even figured out how
to fix the problems of the internet?

1390
01:20:26,462 --> 01:20:28,032
Leo Dion (host): Yeah, I agree completely.

1391
01:20:28,032 --> 01:20:28,422
Yeah.

1392
01:20:28,882 --> 01:20:30,192
Yeah, totally.

1393
01:20:31,132 --> 01:20:35,262
All right, before we leave, let's talk
about, you have a new library you've been

1394
01:20:35,262 --> 01:20:41,882
working on that's supposed to help with
logging and iOS development and stuff.

1395
01:20:41,892 --> 01:20:42,902
You wanna talk about it?

1396
01:20:43,998 --> 01:20:44,528
Joe Fabisevich (guest): Yeah.

1397
01:20:44,628 --> 01:20:48,488
So I'm sure that you've had this
experience where you're building an app

1398
01:20:48,538 --> 01:20:53,018
and it uses… Like, Claude goes in and
starts a bunch of logs so it can figure

1399
01:20:53,018 --> 01:20:54,728
out what the heck is happening, right?

1400
01:20:56,338 --> 01:21:00,568
So what I realized was what if your
whole app did that all the time?

1401
01:21:00,788 --> 01:21:04,538
You just had logs that documented
every action, every piece of state,

1402
01:21:04,588 --> 01:21:06,188
every metric, every diagnostic.

1403
01:21:06,408 --> 01:21:08,608
And this is actually kind of how, like,
a lot of enterprise apps are built.

1404
01:21:08,608 --> 01:21:11,578
They have just, like, unending amounts
of logs that you can then trawl through.

1405
01:21:11,988 --> 01:21:15,498
And the reason that I came up with this,
though, was because I would basically

1406
01:21:15,708 --> 01:21:19,608
have, like, Claude or Codex build a
thing, and then I would have it, like,

1407
01:21:19,618 --> 01:21:23,748
log out a whole bunch of stuff, and then
it would be able to read through its

1408
01:21:23,748 --> 01:21:27,548
own logs and determine what happened
to understand the cause and effect.

1409
01:21:27,988 --> 01:21:31,958
So if you can determine the cause and
effect from start to finish, what you

1410
01:21:31,958 --> 01:21:37,018
can do is hand… Like, these tools are
so good at taking 100, 1,000, 10,000

1411
01:21:37,028 --> 01:21:42,268
logs and parsing through them to figure
out what happened that then they can

1412
01:21:42,268 --> 01:21:45,608
start beginning to solve race conditions
and bugs that last, like, a year.

1413
01:21:45,608 --> 01:21:49,628
So I have a bug in Plinky that I finally
solved after a year once I built a

1414
01:21:49,628 --> 01:21:53,568
new library called Broadcast, which
is… Like, it's built upon this, like,

1415
01:21:53,568 --> 01:21:55,198
internal library that I had for Plinky.

1416
01:21:55,508 --> 01:22:00,438
But the idea is if you were logging
every state and action that a user takes,

1417
01:22:00,908 --> 01:22:04,468
that even when you are not watching
them use the app, you can figure out

1418
01:22:04,498 --> 01:22:05,958
what happened from start to finish.

1419
01:22:06,218 --> 01:22:09,798
So I mentioned that I gave
Codex a spin dump and logs.

1420
01:22:10,258 --> 01:22:14,888
I basically gave it… I asked the user to
upload logs, and it happens automatically

1421
01:22:14,888 --> 01:22:16,128
as part of my support process.

1422
01:22:16,618 --> 01:22:21,908
And it correlated the events that
they took to the actions of the spin

1423
01:22:21,908 --> 01:22:27,748
dump to figure out where things were
broken, and it's because lo- like,

1424
01:22:28,548 --> 01:22:31,168
agents can't see anything, right?

1425
01:22:31,168 --> 01:22:34,138
You can hand them screenshots, you
can give them tests, you can give

1426
01:22:34,138 --> 01:22:38,248
them, you know, after the fact
figure out what happened, but they

1427
01:22:38,248 --> 01:22:39,888
can't see things as it's happening.

1428
01:22:39,918 --> 01:22:43,448
A screenshot is just a moment
in time, but logs are forever.

1429
01:22:43,458 --> 01:22:45,308
Logs document what you did.

1430
01:22:45,568 --> 01:22:49,718
And so Broadcast is a logging
library that makes it…

1431
01:22:49,868 --> 01:22:52,448
is optimized for human readability.

1432
01:22:52,458 --> 01:22:55,368
It's like, you know, really clean,
nice, easy APIs that don't disrupt your

1433
01:22:55,368 --> 01:23:01,838
code as much, but also optimized for
agents to debug your hardest of problems

1434
01:23:02,734 --> 01:23:04,254
Leo Dion (host): This looks amazing.

1435
01:23:04,264 --> 01:23:08,034
We were talking offline about
how I'm working on my app At

1436
01:23:08,034 --> 01:23:11,984
Least which is a Apple Watch
app for, like, a passive timer.

1437
01:23:12,634 --> 01:23:17,024
And I wanted an iPhone component
because you really can't make an app

1438
01:23:17,094 --> 01:23:20,224
and have anybody see it in the App
Store unless it's an iPhone app too.

1439
01:23:20,814 --> 01:23:23,714
And I was working on watch connectivity
stuff so you can start the timer

1440
01:23:23,714 --> 01:23:26,694
on the phone, and Claude was great.

1441
01:23:26,794 --> 01:23:31,024
It wrote a bunch of logs, so that way I
can pay attention to what's going on the

1442
01:23:31,024 --> 01:23:32,664
watch and the phone at the same time.

1443
01:23:33,094 --> 01:23:38,084
This looks like exactly what I need
for At Least, because I wanna be

1444
01:23:38,084 --> 01:23:42,464
able to know, hey, is… Like, if
I tap on the phone, why is it not

1445
01:23:42,764 --> 01:23:44,144
starting the timer on the watch?

1446
01:23:44,144 --> 01:23:46,654
And I want all those log
statements and things like that.

1447
01:23:46,654 --> 01:23:49,914
So I am really excited to try this out.

1448
01:23:50,024 --> 01:23:53,824
We'll put a link in the show
notes to this, but this is amazing

1449
01:23:55,202 --> 01:23:57,552
Joe Fabisevich (guest): One thing
I forgot to mention actually is I--

1450
01:23:57,602 --> 01:23:59,322
it started as a composable logger.

1451
01:23:59,502 --> 01:24:03,222
So your watch and phone example is
perfect, where your watch can have a

1452
01:24:03,232 --> 01:24:07,162
logger, your phone can have a logger,
and you can actually merge those two logs

1453
01:24:07,162 --> 01:24:12,292
together to say, "When I took this action
on my phone, then it did this thing on

1454
01:24:12,292 --> 01:24:15,212
the watch, and when I did this thing on
the watch, it took an action on my phone."

1455
01:24:15,472 --> 01:24:21,112
And so you can hand one stream of logs
that combines a watch, a server your

1456
01:24:21,122 --> 01:24:24,962
phone, and I actually use this API.

1457
01:24:25,262 --> 01:24:26,432
I have an analytics logger.

1458
01:24:26,432 --> 01:24:27,572
I have a remote logger.

1459
01:24:27,612 --> 01:24:29,272
I have a multi-session logger.

1460
01:24:29,472 --> 01:24:32,632
I have all these different
loggers that I can then turn to.

1461
01:24:33,042 --> 01:24:37,172
And so analytics logs are formatted
differently, so that way I can read

1462
01:24:37,972 --> 01:24:39,512
my analytics in, like, a clear manner.

1463
01:24:39,512 --> 01:24:42,102
It's like tap this, did this, track that.

1464
01:24:42,592 --> 01:24:47,842
But if I'm trying to, like, optimize for
an AI, I have a token optimized format

1465
01:24:48,032 --> 01:24:52,492
that it exports to and says, "Great," like
minimize the usage of the context window.

1466
01:24:52,602 --> 01:24:56,522
So the idea is you can have many loggers
and you combine them or you can separate

1467
01:24:56,522 --> 01:25:00,352
them and filter them out as much as you
want to get the information that you need

1468
01:25:00,382 --> 01:25:06,362
for either for you, for support, for AI,
and for whatever, all with one simple API.

1469
01:25:06,362 --> 01:25:10,872
So you just call log.debug, log.error,
log.info, and it does all these

1470
01:25:10,872 --> 01:25:12,602
things in multiple places at once.

1471
01:25:13,568 --> 01:25:14,048
Leo Dion (host): Nice.

1472
01:25:14,158 --> 01:25:14,898
This is great.

1473
01:25:15,058 --> 01:25:21,118
And it's … Okay, so it ca- it can use
OS log, but otherwise you can't … Like,

1474
01:25:21,118 --> 01:25:25,208
what would you recommend somebody do
to get started using this with their

1475
01:25:25,218 --> 01:25:27,798
app as far as, like, a log destination?

1476
01:25:29,042 --> 01:25:30,962
Joe Fabisevich (guest): The simplest
destination is the session logger.

1477
01:25:31,002 --> 01:25:33,552
All it does, it writes to memory, right?

1478
01:25:33,552 --> 01:25:36,832
Like, it has a buffer that it's writing
to and says, "These are the logs for this

1479
01:25:36,832 --> 01:25:41,652
session." Now, what I found really helpful
is it ships with support for multiple

1480
01:25:41,652 --> 01:25:46,192
sessions at once using the open source
library I maintain called Boutique, which

1481
01:25:46,312 --> 01:25:52,482
it's basically like sometimes you have
bugs that creep up over multiple uses.

1482
01:25:52,712 --> 01:25:56,062
So, like, the user quit the app and
then this thing started happening.

1483
01:25:56,702 --> 01:26:00,592
If you have the multi-session logger,
you'll actually be able to track down bugs

1484
01:26:00,592 --> 01:26:05,382
that are much harder to figure out where
state is often a problem, or, like, state

1485
01:26:05,382 --> 01:26:07,642
restoration or state cancellation happens.

1486
01:26:07,650 --> 01:26:07,970
Leo Dion (host): Okay

1487
01:26:09,272 --> 01:26:11,992
Joe Fabisevich (guest): I use the
console logger so I can read through

1488
01:26:12,002 --> 01:26:13,572
all the console statements that I want.

1489
01:26:14,092 --> 01:26:20,262
But also combine it with a multi-session
logger so that way when a user reaches

1490
01:26:20,262 --> 01:26:24,762
out to support, it sends me their last
2,500 logs across multiple sessions, and

1491
01:26:24,762 --> 01:26:28,792
then I can hand those 2,500 logs to Codex
and Codex says, "Oh, here's the problem.

1492
01:26:28,812 --> 01:26:33,102
I figured out where it is in the code
and I fixed it for you autonomously."

1493
01:26:33,890 --> 01:26:34,330
Leo Dion (host): Awesome.

1494
01:26:34,760 --> 01:26:35,760
That is amazing.

1495
01:26:35,760 --> 01:26:36,810
Definitely check this out.

1496
01:26:37,330 --> 01:26:40,760
I see you have set up you have your
Claude plugin and your agents and

1497
01:26:40,760 --> 01:26:42,360
stuff, so this is, like, ready to go.

1498
01:26:43,070 --> 01:26:43,370
Yeah.

1499
01:26:43,490 --> 01:26:44,440
Fantastic.

1500
01:26:44,940 --> 01:26:47,130
Anything else you wanted to
mention before we close out?

1501
01:26:48,570 --> 01:26:49,870
Joe Fabisevich (guest): No, I mean,
the most important thing in my life

1502
01:26:49,870 --> 01:26:52,020
right now is just saying, "Let's
go Knicks," over and over again

1503
01:26:52,552 --> 01:26:52,852
Leo Dion (host): I know.

1504
01:26:52,982 --> 01:26:53,652
That was amazing.

1505
01:26:53,662 --> 01:26:58,562
That was my highlight of last week,
and that says a lot because we had WWDC

1506
01:26:59,082 --> 01:27:02,122
Nintendo Direct, so yeah, that was great.

1507
01:27:02,192 --> 01:27:03,572
Joe Fabisevich (guest): week
of my life was two months

1508
01:27:03,872 --> 01:27:04,152
Leo Dion (host): Yeah.

1509
01:27:05,132 --> 01:27:06,502
Thank you, Joe, for coming back on.

1510
01:27:06,502 --> 01:27:07,512
Where can people find you?

1511
01:27:08,804 --> 01:27:09,134
Joe Fabisevich (guest): Yeah.

1512
01:27:09,154 --> 01:27:10,224
So I'm on Bluesky.

1513
01:27:10,274 --> 01:27:11,904
I'm mergesort.me there.

1514
01:27:11,954 --> 01:27:12,694
I'm also on Mastodon.

1515
01:27:12,694 --> 01:27:15,784
I'm macaw.social/@mergesort.

1516
01:27:16,204 --> 01:27:18,094
Really, I write at
build.ms and fabisevich.ch.

1517
01:27:19,724 --> 01:27:24,524
That's mostly where I am, but yeah,
honestly, my name is very unique,

1518
01:27:24,524 --> 01:27:26,684
so feel free to just Google me
and you'll find whatever you need

1519
01:27:27,438 --> 01:27:27,948
Leo Dion (host): Awesome.

1520
01:27:28,078 --> 01:27:29,638
Thank you again for coming back on.

1521
01:27:29,798 --> 01:27:33,578
People can find me on
Mastodon at leogdion@c.im,

1522
01:27:33,578 --> 01:27:36,068
at leogdion everywhere else.

1523
01:27:36,638 --> 01:27:38,708
Brightdigit.com, that's my company.

1524
01:27:39,258 --> 01:27:44,888
I've posted on social media that if you
are interested and want to talk about

1525
01:27:44,918 --> 01:27:49,718
career, you wanna talk about automation
and Swift, or you wanna talk about how to

1526
01:27:49,738 --> 01:27:54,928
plug AI into your system for AI assisted
development, there's a link in the show

1527
01:27:54,928 --> 01:27:57,148
notes to free 30-minute consulting.

1528
01:27:57,148 --> 01:28:01,118
Let's just chat and see if there's a way
I can help you in any of those things.

1529
01:28:01,188 --> 01:28:04,718
And then if you're listening to
this on a podcast player, Overcast,

1530
01:28:04,738 --> 01:28:10,638
Apple, Spotify, please give me a
review, and if you're watching this

1531
01:28:10,638 --> 01:28:11,948
on YouTube, like and subscribe.

1532
01:28:11,998 --> 01:28:15,538
Consider joining our Patreon
patreon.com/brightdigit if you want

1533
01:28:15,698 --> 01:28:18,408
these episodes right away, like this one.

1534
01:28:18,908 --> 01:28:21,908
Thank you again, and I look
forward to talking to you later.

1535
01:28:21,958 --> 01:28:22,598
Bye, everybody