WEBVTT

NOTE Sentence-level transcript of https://www.youtube.com/watch?v=ZpK5PWX2YRM

NOTE One cue per sentence. Cue ids are the line anchors on /transcripts/ZpK5PWX2YRM.html. A cue ends where the next begins, or 2 s after its last word.

s1
00:00:01.309 --> 00:00:03.309
[music]

s2
00:00:12.440 --> 00:00:15.120
Two talks at the AI Engineer Europe.

s3
00:00:15.120 --> 00:00:17.840
One guy is saying, "Code is free."

s4
00:00:17.840 --> 00:00:19.840
and deleted his ID.

s5
00:00:19.840 --> 00:00:24.200
And the other one is saying, "Read every effing line of code."

s6
00:00:24.200 --> 00:00:34.240
So, should AI engineers still read code their agents output I named this the Desolation of Apollo Continuum.

s7
00:00:34.240 --> 00:00:37.480
And you guys probably have argued about this in Slack.

s8
00:00:37.480 --> 00:00:39.040
You probably talked about this in the hallway track.

s9
00:00:39.040 --> 00:00:40.600
So, let's talk about this here.

s10
00:00:40.600 --> 00:00:45.480
Because code got cheap, attention didn't.

s11
00:00:46.680 --> 00:00:52.320
As you may know, back in December 2025, something big changed.

s12
00:00:52.320 --> 00:00:56.440
AI engineering has changed forever, and it broke its own trendline.

s13
00:00:56.440 --> 00:01:05.440
Actually, Swyx, the organizer of AI Engineer, is collecting evidence to that single moment in time at the website called wtfhappened2025.com.

s14
00:01:05.640 --> 00:01:07.040
I recommend you go and check it out.

s15
00:01:07.040 --> 00:01:08.120
It's really, really funny.

s16
00:01:08.120 --> 00:01:13.080
Uh this is just one example from METR, the machine evaluation center.

s17
00:01:13.080 --> 00:01:21.680
And it shows that models, for the first time, started completing tasks that would take engineers over 16 hours to do.

s18
00:01:21.680 --> 00:01:26.080
And in fact, we've gone way up the curve, way up the trendline after that.

s19
00:01:26.080 --> 00:01:31.040
This is the backdrop to everything that AI engineering is experiencing.

s20
00:01:31.920 --> 00:01:34.760
Because we don't write code anymore.

s21
00:01:34.760 --> 00:01:36.280
Most of us, at least.

s22
00:01:36.280 --> 00:01:41.920
I want to see one Can you guys give me a raise of hands if you still handcraft and write code, most of your code?

s23
00:01:41.920 --> 00:01:42.880
Anybody here?

s24
00:01:42.880 --> 00:01:44.720
Most of your code is written by hand?

s25
00:01:44.720 --> 00:01:45.160
Amazing.

s26
00:01:45.160 --> 00:01:48.320
This is the Token Maxing track, after all.

s27
00:01:48.320 --> 00:01:52.000
I I think the one person here who still writes code is maybe a little shy of raising their hand.

s28
00:01:52.000 --> 00:01:52.880
That's okay.

s29
00:01:52.880 --> 00:01:53.920
Cuz we don't type code anymore.

s30
00:01:53.920 --> 00:01:55.600
We're not handcrafters.

s31
00:01:55.600 --> 00:01:56.760
We supervise.

s32
00:01:56.760 --> 00:01:59.400
I like to say we babysit agents.

s33
00:01:59.400 --> 00:02:03.120
And the greatest example for this obviously is Boris's journey.

s34
00:02:03.120 --> 00:02:06.880
You guys know Boris, the creator of Cloud Code, uh Anthropic.

s35
00:02:06.880 --> 00:02:10.679
100% of his code is written and authored by Cloud Code at this point.

s36
00:02:10.679 --> 00:02:14.240
And he didn't stop being engineer, he moved up the layer.

s37
00:02:14.240 --> 00:02:18.040
He still ships 20 to 30 PRs, maybe more.

s38
00:02:18.040 --> 00:02:21.600
And recently he talked about he deleted his ID.

s39
00:02:21.600 --> 00:02:22.800
I found it really funny.

s40
00:02:22.800 --> 00:02:26.000
Just just no reason to just hand type code anymore.

s41
00:02:26.000 --> 00:02:32.840
In fact, 80% of Anthropic's code is now AI written and this is this stat is at least a few months old.

s42
00:02:32.840 --> 00:02:36.240
It's likely more right now.

s43
00:02:36.240 --> 00:02:37.760
And he's not the only one.

s44
00:02:37.760 --> 00:02:41.040
Some of you have seen this chart from GitHub.

s45
00:02:41.040 --> 00:02:44.680
Some of you have maybe remembered this chart while GitHub was down for you.

s46
00:02:44.680 --> 00:02:51.320
Uh the reason is GitHub is on track to to to get 14 billion commits this year.

s47
00:02:51.320 --> 00:02:55.920
All of 2025 all of yes we yet last year was 1 billion.

s48
00:02:55.920 --> 00:02:58.240
They're 14x the number of commits.

s49
00:02:58.240 --> 00:03:03.280
They're seeing 14x the number of commits, which is insane and most of this is AI assisted.

s50
00:03:03.280 --> 00:03:05.760
And it's a lot of code.

s51
00:03:06.320 --> 00:03:12.519
And so the AI engineering has changed forever and I want to tell you about AI Engineer World's Fair.

s52
00:03:12.519 --> 00:03:15.400
I been to every single one and I'll tell you about this later.

s53
00:03:15.400 --> 00:03:22.680
And AI Engineer is a great place to get the zeitgeist of where our career is going and how is it changing, okay?

s54
00:03:22.680 --> 00:03:25.360
This one obviously is 3x bigger than last year.

s55
00:03:25.360 --> 00:03:26.360
This is just one of the rooms.

s56
00:03:26.360 --> 00:03:27.600
There's like a bunch of rooms.

s57
00:03:27.600 --> 00:03:29.480
7,000 people I think we clocked in.

s58
00:03:29.480 --> 00:03:31.160
36 tracks.

s59
00:03:31.160 --> 00:03:35.959
And if you want to know what happens in AI engineering, you kind of have to be here.

s60
00:03:35.959 --> 00:03:37.920
So this would be a little bit of meta talk.

s61
00:03:37.920 --> 00:03:42.560
So one of the guys at AI Engineer EU talked about code as ship.

s62
00:03:42.560 --> 00:03:46.400
The other one talked about uh we should read every line of code.

s63
00:03:46.400 --> 00:03:48.160
Let's listen to them for just a second, okay?

s64
00:03:48.160 --> 00:03:50.560
This is Ryan LePopulaire from OpenAI.

s65
00:03:50.560 --> 00:03:54.280
I I think he's he made it here, but uh this is Ryan LeFebvre from Open AI.

s66
00:03:54.280 --> 00:03:58.520
The models at this point are good enough where they're isomorphic they can write code.

s67
00:03:58.520 --> 00:03:59.400
Can you guys

s68
00:03:59.400 --> 00:04:00.000
Do you mind?

s69
00:04:00.000 --> 00:04:06.680
These tools are high quality that solve real user problems in real code bases.

s70
00:04:07.400 --> 00:04:08.880
Code is free.

s71
00:04:08.880 --> 00:04:17.320
It's free for you to produce, free to refactor, and it is not a thing to get hung up on anymore.

s72
00:04:17.320 --> 00:04:20.880
Humans no longer need to concern themselves with implementation.

s73
00:04:20.880 --> 00:04:25.520
The important thing is not the code, but the prompt and the guardrails that got you there.

s74
00:04:25.520 --> 00:04:28.320
You can just simply say, "Do not produce slop."

s75
00:04:28.320 --> 00:04:29.120
Don't accept slop.

s76
00:04:29.120 --> 00:04:30.800
You won't get slop in your code base.

s77
00:04:30.800 --> 00:04:41.760
But, to do that requires taking short-term velocity hits in order to back up or double-click into a task to figure out what it is the agents are struggling with.

s78
00:04:41.919 --> 00:04:43.560
So, this is Ryan LeFebvre, okay?

s79
00:04:43.560 --> 00:04:50.040
He came up on stage at the engineer, and he opened with like, "Hey, I'm a talking billionaire, and I want you to be as well."

s80
00:04:50.040 --> 00:04:54.360
In fact, the talking billionaire lounge that's in front of the leadership track that you guys see, that's because of him.

s81
00:04:54.360 --> 00:04:55.880
He came up with this concept.

s82
00:04:55.880 --> 00:04:57.640
And he got the golden card and everything.

s83
00:04:57.640 --> 00:05:02.560
On the other side, the same conference, the other side, Mario Zechner, creator of Pi.

s84
00:05:02.560 --> 00:05:04.440
Slow the down.

s85
00:05:04.440 --> 00:05:06.880
Everything's broken.

s86
00:05:07.600 --> 00:05:11.160
And then there's people that say, "Our product's been 100% built by agents."

s87
00:05:11.160 --> 00:05:12.720
Yes, we know it sucks now.

s88
00:05:12.720 --> 00:05:14.720
Congratulations.

s89
00:05:15.713 --> 00:05:17.713
[applause]

s90
00:05:21.240 --> 00:05:28.200
Agents are actually compounding booboos, which is my word for errors, with zero learning and no bottlenecks and delayed pain.

s91
00:05:28.200 --> 00:05:30.320
The delayed pain is for you.

s92
00:05:30.320 --> 00:05:32.080
Those are my most beloved people.

s93
00:05:32.080 --> 00:05:34.200
I don't even read the code anymore.

s94
00:05:34.200 --> 00:05:35.520
Congratulations.

s95
00:05:35.520 --> 00:05:39.120
Something is broken, and your users are screaming, so who you going to call?

s96
00:05:39.120 --> 00:05:41.800
Not yourself, because you haven't read the code.

s97
00:05:41.800 --> 00:05:44.360
Non-critical code, sure, write slop ahead.

s98
00:05:44.360 --> 00:05:46.640
Critical code, read every line.

s99
00:05:46.640 --> 00:05:52.560
So, two folks, same conference, day after day, talking about the one anxiety that we all feel.

s100
00:05:52.560 --> 00:05:54.920
Should we all still be reading code in 2026?

s101
00:05:54.920 --> 00:06:00.800
By the way, these two folks are the number six and number seven most watched YouTube videos from AI engineer from all time.

s102
00:06:00.800 --> 00:06:04.160
So, they're obviously representing something that we're feeling we're talking about.

s103
00:06:04.160 --> 00:06:08.520
And this is being the leadership track, something that you folks that report to you are talking about, okay?

s104
00:06:08.520 --> 00:06:09.880
Should they still be reading code?

s105
00:06:09.880 --> 00:06:11.520
And what's the what's the level quality?

s106
00:06:11.520 --> 00:06:13.720
So, they name the same anxiety from both ends.

s107
00:06:13.720 --> 00:06:17.440
Uh at this point, I probably should introduce Uh hey, I'm Alex Volkov.

s108
00:06:17.440 --> 00:06:19.040
I'm the host of Thursday AI podcast.

s109
00:06:19.040 --> 00:06:20.080
It's a podcast and a newsletter.

s110
00:06:20.080 --> 00:06:22.320
We go live every week to talk about AI.

s111
00:06:22.320 --> 00:06:28.520
For the past 3 and 1/2 years, we've been tracking every change in the AI engineering, every release from every lab, every model.

s112
00:06:28.520 --> 00:06:32.280
And I'm also an AI evangelist with Weights & Biases and Core Weave.

s113
00:06:32.280 --> 00:06:34.760
Um what also should I tell you about myself?

s114
00:06:34.760 --> 00:06:38.440
That I've been covering AI engineers specifically since the first one in 2023.

s115
00:06:38.440 --> 00:06:40.080
And oh boy, has it changed.

s116
00:06:40.080 --> 00:06:44.880
And so, you can treat this as a dispatch from the front line because all of these people now are my friends.

s117
00:06:44.880 --> 00:06:49.840
And we constantly talk about this in the speakers' room in the hallway track.

s118
00:06:50.800 --> 00:06:52.560
I couldn't stop thinking about that tension.

s119
00:06:52.560 --> 00:06:57.240
I couldn't stop thinking about that kind of disparity between the two folks, okay?

s120
00:06:57.240 --> 00:07:01.960
And I put them both on a line, Zeshner from one end, Lapopolo on the other end.

s121
00:07:01.960 --> 00:07:03.360
I called it a continuum.

s122
00:07:03.360 --> 00:07:08.320
And I basically started asking people, "Hey, where are you on this line?

s123
00:07:08.320 --> 00:07:09.480
Where are you a Zeshner?

s124
00:07:09.480 --> 00:07:11.440
Do you still read every line of code?

s125
00:07:11.440 --> 00:07:12.440
Are you a Lapopolo?

s126
00:07:12.440 --> 00:07:17.080
Do you just YOLO and don't even look at code and think agents are good enough, etc.?"

s127
00:07:17.080 --> 00:07:20.680
And I I got the framing wrong.

s128
00:07:20.680 --> 00:07:22.320
But, I'll tell you about this in just a second, okay?

s129
00:07:22.320 --> 00:07:24.680
So, before this, I want you to be kind of honest with yourself.

s130
00:07:24.680 --> 00:07:33.480
And again, if you don't write code, or let me say this, if you don't babysit your own agents, but you you have reports that babysit agents for you, uh think about them when you answer this, okay?

s131
00:07:33.480 --> 00:07:34.960
And be honest.

s132
00:07:34.960 --> 00:07:39.040
On the ZL continuum, where are you?

s133
00:07:39.040 --> 00:07:48.640
And let's take um let's take by vote of hands, who here has committed code that they've never looked at before?

s134
00:07:48.640 --> 00:07:50.440
Amazing, love that.

s135
00:07:50.440 --> 00:07:54.000
Who here still reads every line of code or at least critical code?

s136
00:07:54.000 --> 00:07:55.360
I see one cowboy over there.

s137
00:07:55.360 --> 00:07:56.080
I love that, man.

s138
00:07:56.080 --> 00:07:57.240
I'm going to talk to you afterwards, okay?

s139
00:07:57.240 --> 00:07:59.920
I want to understand exactly why you do this.

s140
00:07:59.920 --> 00:08:03.400
Um, and so who's right?

s141
00:08:03.400 --> 00:08:04.160
Let's talk about who's right.

s142
00:08:04.160 --> 00:08:07.920
Let's talk about where we are right now, and we start with Ryan Lepopolo.

s143
00:08:07.920 --> 00:08:11.440
If you get to meet Ryan over here, he's very AGI built.

s144
00:08:11.440 --> 00:08:17.400
I think even within OpenAI, the AGI organization, Ryan is kind of like the more AGI built person.

s145
00:08:17.400 --> 00:08:23.840
If you have a chance to go downstairs and grab the AGI pills that 2X prescribed, I think Ryan had all of them.

s146
00:08:23.840 --> 00:08:28.320
He works at OpenAI, where he sits code is literally free.

s147
00:08:28.320 --> 00:08:29.800
So are tokens.

s148
00:08:29.800 --> 00:08:34.039
It's really We renamed Ryan, do you guys know the dash dash YOLO in Codex?

s149
00:08:34.039 --> 00:08:36.520
It's kind of like the skip dangerous permissions in cloud code.

s150
00:08:36.520 --> 00:08:41.640
So we renamed the Ryan Lepopolo YOLO Popolo.

s151
00:08:41.840 --> 00:08:43.560
He's okay with it, by the way, I asked him.

s152
00:08:43.560 --> 00:08:48.080
So if we check his kind of side, the the folks like him against the data, they're actually right.

s153
00:08:48.080 --> 00:08:50.680
The optimists are right, at least about output.

s154
00:08:50.680 --> 00:08:56.200
This is from Faros AI, I think I'm not the only speaker at this conference who sites this essay.

s155
00:08:56.200 --> 00:08:57.640
It's sorry, this survey.

s156
00:08:57.640 --> 00:08:59.920
It's new from April 2026.

s157
00:08:59.920 --> 00:09:04.839
I think it's one of the best kind of evidence of where we're going that we can now site, okay?

s158
00:09:04.839 --> 00:09:07.320
22,000 engineers were surveyed about code.

s159
00:09:07.320 --> 00:09:10.000
They call this the acceleration whiplash.

s160
00:09:10.000 --> 00:09:18.480
And they're talking about my favorite stat on here, and you can read this yourself, 861% increase in code deletion per PR.

s161
00:09:18.480 --> 00:09:22.520
So us together with AI agents, we love deleting code.

s162
00:09:22.520 --> 00:09:29.440
Anthropic also said that they're shipping eight times more code per quarter than in 2025.

s163
00:09:29.880 --> 00:09:33.200
But is it all good code?

s164
00:09:33.200 --> 00:09:34.720
Okay, let's play a game.

s165
00:09:34.720 --> 00:09:38.480
If you know the answer, you let me have my moment on here on stage, okay?

s166
00:09:38.480 --> 00:09:41.200
But if you don't know the answer, let's guess.

s167
00:09:41.200 --> 00:09:44.280
Whose status page is this?

s168
00:09:45.360 --> 00:09:47.920
I think I I hear a few answers.

s169
00:09:47.920 --> 00:09:48.960
I think most of us guessed it.

s170
00:09:48.960 --> 00:09:50.600
This is Claude.

s171
00:09:50.600 --> 00:09:55.320
Uh in fact, as you can see on the on the right, it was down when I took the screenshot.

s172
00:09:55.320 --> 00:09:56.400
It was really funny.

s173
00:09:56.400 --> 00:10:04.360
Uh this Anthropic is the company that probably uses the most AI-generated code, and their status page looks like a Christmas tree.

s174
00:10:04.360 --> 00:10:06.000
Now, I'm not here to dunk on Anthropic.

s175
00:10:06.000 --> 00:10:13.560
Uh they did an incredible job back on stage uh talking about Claude and etc. Um this may be due to scale, this may be due to other factors.

s176
00:10:13.560 --> 00:10:15.052
Uh I'm not here to dunk on them, but

s177
00:10:15.052 --> 00:10:15.240
[snorts]

s178
00:10:15.240 --> 00:10:17.680
it just goes to show that they're not the only ones like this.

s179
00:10:17.680 --> 00:10:21.680
Obviously, GitHub is famously also suffers from a little bit of golf.

s180
00:10:21.680 --> 00:10:23.840
Output does not mean stability, okay?

s181
00:10:23.840 --> 00:10:27.040
So, maybe this is a good example of what?

s182
00:10:27.040 --> 00:10:34.360
Same essay, 31% increase in PRs merged with no review at all, human or agentic.

s183
00:10:34.360 --> 00:10:35.080
Don't do this.

s184
00:10:35.080 --> 00:10:35.840
I beg of you.

s185
00:10:35.840 --> 00:10:39.640
Don't It's it's We'll talk about how to fix this in a second.

s186
00:10:39.640 --> 00:10:44.360
So, when you ship this fast and this much, something gives, and usually it's quality.

s187
00:10:44.360 --> 00:10:46.240
So, maybe Mario's right, yeah?

s188
00:10:46.240 --> 00:10:49.200
Maybe the bill does come due in production.

s189
00:10:49.200 --> 00:10:54.680
Same study, 242% increase in incident per PR.

s190
00:10:54.680 --> 00:10:55.560
This kind of scary.

s191
00:10:55.560 --> 00:10:56.839
The second study is also scary.

s192
00:10:56.839 --> 00:11:02.000
Bugs per developer is up six times than 2025.

s193
00:11:02.080 --> 00:11:06.839
So, even Anthropic can see this.

s194
00:11:06.839 --> 00:11:14.000
I don't know if you guys read the RSI essay they posted, the recursive self-improvement, where they talk about, "Hey, what does the future hold?"

s195
00:11:14.000 --> 00:11:15.440
They outline two scenarios.

s196
00:11:15.440 --> 00:11:21.200
One of them says, "Maybe the acceleration will stop, and we're going to get used to this."

s197
00:11:21.200 --> 00:11:23.920
They all They say, "That's actually not likely to happen.

s198
00:11:23.920 --> 00:11:26.560
We just added this uh eventuality for clarity.

s199
00:11:26.560 --> 00:11:27.920
We don't think that's likely to happen.

s200
00:11:27.920 --> 00:11:41.200
What we think is going to happen is uh engineers and companies 10Xing to 100x-ing to 1,000x-ing their output and productivity, and then they say this, "We, as we began to push more code around the organization,

s201
00:11:41.200 --> 00:11:44.320
human code review has become a new bottleneck."

s202
00:11:44.320 --> 00:11:52.880
They're citing Amdahl's law that shows that if you have an explosion of productivity in one area, another area going to going to get blocked.

s203
00:11:52.880 --> 00:11:58.320
And nobody removes the human in these organizations.

s204
00:11:58.320 --> 00:12:01.800
In fact, careers in Anthropic and careers in OpenAI, they're still hiring humans.

s205
00:12:01.800 --> 00:12:08.440
So, nobody's removing the human, and they're both saying that human code review is still a concern.

s206
00:12:08.560 --> 00:12:09.840
And here's my mea culpa.

s207
00:12:09.840 --> 00:12:12.000
I promise you I'll tell you where I got it wrong the framing.

s208
00:12:12.000 --> 00:12:13.920
My mea culpa is the continuum is real.

s209
00:12:13.920 --> 00:12:17.320
The Zeal continuum is real, but it's not about the people.

s210
00:12:17.320 --> 00:12:20.000
It's about the tasks.

s211
00:12:20.000 --> 00:12:21.160
The continuum is real.

s212
00:12:21.160 --> 00:12:22.200
It's not about the people.

s213
00:12:22.200 --> 00:12:24.200
It's about the task.

s214
00:12:24.200 --> 00:12:34.920
Same engineer could be a Ryan LeCompte on Apollo on one piece of code, and has to be Mario Zechner and read every line of other pieces of code.

s215
00:12:34.920 --> 00:12:38.640
Different task just need different proof.

s216
00:12:39.400 --> 00:12:44.000
If we look at them closely, I obviously character- characterized them.

s217
00:12:44.000 --> 00:12:46.839
I practiced this word multiple times and I still got it wrong.

s218
00:12:46.839 --> 00:12:47.600
Characterized them.

s219
00:12:47.600 --> 00:12:51.120
Their character on both ends for the Zeal continuum.

s220
00:12:51.120 --> 00:12:56.240
But, if you look at them closely and what they're saying closely, they're actually not that different.

s221
00:12:56.240 --> 00:12:58.360
Ryan's mechanism is moving attention up the layer.

s222
00:12:58.360 --> 00:13:03.760
He's saying humans are unreliable at catching repeated mistakes of the same type.

s223
00:13:03.760 --> 00:13:05.880
Repeatedly catching the mistakes of the same type.

s224
00:13:05.880 --> 00:13:15.560
So, when they you do catch a mistake during the PR review, write the documentation, the linter, and the reviewer needs to remember this once, so the system will catch this type of bugs.

s225
00:13:15.560 --> 00:13:17.280
He's not saying don't inspect your code.

s226
00:13:17.280 --> 00:13:20.480
He's saying inspect the system, not every line.

s227
00:13:20.480 --> 00:13:23.960
Mario from the other end is saying route by task.

s228
00:13:23.960 --> 00:13:25.240
If it's not critical, let it rip.

s229
00:13:25.240 --> 00:13:27.560
He said it.

s230
00:13:27.720 --> 00:13:29.520
And if it's critical, you read every line.

s231
00:13:29.520 --> 00:13:32.280
How do you know what's critical?

s232
00:13:32.280 --> 00:13:33.400
Well, his answer is easy.

s233
00:13:33.400 --> 00:13:34.920
You read the effing code.

s234
00:13:34.920 --> 00:13:38.120
Uh my answer to add to this is also you ask your clankers.

s235
00:13:38.120 --> 00:13:44.120
They're great at looking at a large repository and saying you and telling you, "Hey, this line is actually critical.

s236
00:13:44.120 --> 00:13:45.320
You should look at this area.

s237
00:13:45.320 --> 00:13:47.320
This primitives over here are critical."

s238
00:13:47.320 --> 00:13:49.240
So, you ask your clanker.

s239
00:13:49.240 --> 00:13:53.400
So, they agree more than I can have given credit for.

s240
00:13:53.560 --> 00:14:00.840
And so, I think at the beginning of this, the wrong question is should I still be reading code in 2026?

s241
00:14:00.840 --> 00:14:06.800
I think the better question right now for all of us is what proof does this specific change need?

s242
00:14:06.800 --> 00:14:10.160
What proof does this specific change need?

s243
00:14:10.160 --> 00:14:12.480
And so, I took Mario on the left, obviously.

s244
00:14:12.480 --> 00:14:13.720
I took Ryan on the right.

s245
00:14:13.720 --> 00:14:19.640
And then I took a bunch of other uh great AI engineers, friends, some friends, uh many of them speakers at this conference.

s246
00:14:19.640 --> 00:14:22.400
I kind of distilled their advice down to a routing table.

s247
00:14:22.400 --> 00:14:29.120
And uh they told me, I think Swyx told me on Twitter, "Just give me one slide that I will that people need to take a screenshot of.

s248
00:14:29.120 --> 00:14:29.800
It's going to be this slide.

s249
00:14:29.800 --> 00:14:33.839
You don't have to read it with me, but in the end, you're welcome to take a picture of this."

s250
00:14:33.839 --> 00:14:35.960
This is the your Monday artifact.

s251
00:14:35.960 --> 00:14:38.480
Routing the change where the proof needs it.

s252
00:14:38.480 --> 00:14:41.240
Routing the change to the proof that it needs.

s253
00:14:41.240 --> 00:14:47.200
You read every line of authentication, money movement, permissions, and irreversible data.

s254
00:14:47.200 --> 00:14:50.600
You inspect the critical path yourself, and then obviously, you keep going.

s255
00:14:50.600 --> 00:14:52.480
Uh decomposing, I think is very important.

s256
00:14:52.480 --> 00:14:55.120
The more code is getting written, the more it's hard.

s257
00:14:55.120 --> 00:14:59.040
Your eyes are starting to glaze over a a very long pull request.

s258
00:14:59.040 --> 00:15:01.160
So, split it into atomic reviewable PRs.

s259
00:15:01.160 --> 00:15:02.280
You know who's good at it?

s260
00:15:02.280 --> 00:15:02.800
Agents.

s261
00:15:02.800 --> 00:15:04.240
They're great at decomposing code.

s262
00:15:04.240 --> 00:15:05.880
Ask them to do it.

s263
00:15:05.880 --> 00:15:07.520
You verify that doesn't go away.

s264
00:15:07.520 --> 00:15:12.400
This has been with us in the in engineering software engineering and the AI engineering doesn't go away.

s265
00:15:12.400 --> 00:15:14.040
Traces, evals, shadow mode.

s266
00:15:14.040 --> 00:15:15.480
Come talk to me after after this talk.

s267
00:15:15.480 --> 00:15:19.560
I don't have enough time, but shadow mode is a really cool one that I learned about preparing this talk.

s268
00:15:19.560 --> 00:15:21.760
And then, I think the most important one is separating.

s269
00:15:21.760 --> 00:15:26.160
Many people have the same agent that writes the code, also inspects the outputs and writes the tests.

s270
00:15:26.160 --> 00:15:27.840
Separating is very important.

s271
00:15:27.840 --> 00:15:33.440
If you don't separate, it's kind of like if I came up with an exam and then I took an exam and I scored myself on the exam.

s272
00:15:33.440 --> 00:15:35.960
It's not not really productive, right?

s273
00:15:35.960 --> 00:15:37.840
And then last one is engineer.

s274
00:15:37.840 --> 00:15:39.320
Rails observability rollback.

s275
00:15:39.320 --> 00:15:41.400
This is what Ryan Le Popo talks about.

s276
00:15:41.400 --> 00:15:49.400
Build the system that builds the system because read spends your attention once, engineer makes the system remember.

s277
00:15:50.000 --> 00:15:52.000
Right?

s278
00:15:52.120 --> 00:15:56.480
And you might sit in might be sitting there and saying, "Hey, did you hear the news, Alex?

s279
00:15:56.480 --> 00:15:57.280
Fable is back.

s280
00:15:57.280 --> 00:15:57.960
What about Fable?

s281
00:15:57.960 --> 00:15:58.680
What about Mythos?

s282
00:15:58.680 --> 00:16:02.520
Is this still relevant at this next scale of capability?"

s283
00:16:02.520 --> 00:16:05.840
Because when I coined the Zeal Continuum, it was only 82 days ago.

s284
00:16:05.840 --> 00:16:07.280
Mythos has just been announced.

s285
00:16:07.280 --> 00:16:08.640
We weren't sure like what was going on.

s286
00:16:08.640 --> 00:16:11.480
Only the people in Anthropic got access to it.

s287
00:16:11.480 --> 00:16:22.400
And Jared S. Chipar, that was on stage from Anthropic, he said about Mythos and and Fable, "We used to check if Claude is doing the work right?

s288
00:16:22.400 --> 00:16:28.640
And with Fable 5, I instead check if Claude is doing the right work."

s289
00:16:28.880 --> 00:16:29.720
Let it land for a second.

s290
00:16:29.720 --> 00:16:31.520
I don't know if you read the statement.

s291
00:16:31.520 --> 00:16:38.160
When I read the statement, I felt like little chills at the back of my neck about the next like level of capability, okay?

s292
00:16:38.160 --> 00:16:40.480
We used to check if Claude is doing the work right.

s293
00:16:40.480 --> 00:16:44.320
With Fable, we check if Claude is doing the right work.

s294
00:16:44.320 --> 00:16:50.280
And our favorite senpai who recently joined Anthropic and is getting unnecessary heat on Twitter said this.

s295
00:16:50.280 --> 00:16:56.320
Andrej Karpathy has said, "It's never felt so tempting to stop looking at code at all.

s296
00:16:56.320 --> 00:16:58.560
But don't do this in production."

s297
00:16:58.560 --> 00:17:01.120
Senpai is great for the sole reason.

s298
00:17:01.120 --> 00:17:04.079
Do you guys know the sentence, "This meeting could have been an email?"

s299
00:17:04.079 --> 00:17:07.839
So, this presentation could have been Andrej Karpathy's one sentence, okay?

s300
00:17:07.839 --> 00:17:11.079
He's naming the anxiety from both ends.

s301
00:17:11.079 --> 00:17:13.199
It's never been so tempting to stop looking at code.

s302
00:17:13.199 --> 00:17:16.959
Don't do this in production, even with Fable.

s303
00:17:16.959 --> 00:17:22.240
And so, if you guys noticed, I have a little thing here.

s304
00:17:22.240 --> 00:17:24.240
This.

s305
00:17:24.839 --> 00:17:28.040
Uh it's so wide you can see my little laser pointer.

s306
00:17:28.040 --> 00:17:30.960
Do you guys see the arrow, the capability drift arrow?

s307
00:17:30.960 --> 00:17:32.760
This thing?

s308
00:17:32.760 --> 00:17:37.640
When I wrote the continuum, I realized that it it's only a temporary place in time.

s309
00:17:37.640 --> 00:17:42.720
Capability increases move us towards the popular.

s310
00:17:42.720 --> 00:17:46.760
So, we're going to talk about capability increases as well because the review layer moves.

s311
00:17:46.760 --> 00:17:56.400
If yesterday we inspected the outputs and we read the code, and today we inspect the task direction and kind of like directed to the right proof, maybe tomorrow we're inspecting the loops.

s312
00:17:56.400 --> 00:17:58.680
Capability drift changes where proof belongs.

s313
00:17:58.680 --> 00:18:02.560
It doesn't remove the requirement of proof.

s314
00:18:02.720 --> 00:18:03.840
Talking about loops.

s315
00:18:03.840 --> 00:18:05.040
Is that the next primitive?

s316
00:18:05.040 --> 00:18:14.160
I think most of this conference, I think the zeitgeist for this one is going to be is talking factories and co-factories are real, and is loops is a real thing that I need to be doing at this point.

s317
00:18:14.160 --> 00:18:17.600
By raising of hands, who here heard of loops?

s318
00:18:17.600 --> 00:18:24.680
Keep your hands up, please, and take them down if you are not running loops right now and you haven't anything about they are.

s319
00:18:24.680 --> 00:18:36.600
There's a good perception There's good number of people here who heard about loops, and they started with both these folks, Peter Steinberger, creator of Open Cloud Open Claw, and now is Open AI, and Boris Cherny.

s320
00:18:36.600 --> 00:18:42.400
And pretty much within the span of 2 days, both of them start talking about loops that became kind of the zeitgeist.

s321
00:18:42.400 --> 00:18:47.840
And loops are moving us from prompting each turn to designing the system that writes the actual prompts.

s322
00:18:47.840 --> 00:18:53.240
By the way, do you guys know what's common between these guys and what's different between me and these guys?

s323
00:18:53.240 --> 00:18:55.280
Their tokens are free.

s324
00:18:55.280 --> 00:19:01.840
So, when they talk about loops and their tokens are free, they're not telling you, "Hey, you should be doing this right now specifically."

s325
00:19:01.840 --> 00:19:07.040
But because they work at bigger labs, you can treat them as kind of a lighthouse that's pointing where we're all going.

s326
00:19:07.040 --> 00:19:15.160
Kind of like Gretzky skate where the puck is going to to They're going to tell us what all of our enterprises are going to get get up on.

s327
00:19:15.160 --> 00:19:19.040
And if it's if it's loops, then let me at least give you a TLDR, okay?

s328
00:19:19.040 --> 00:19:22.800
Loops are basically fancy cron jobs that run on a schedule.

s329
00:19:22.800 --> 00:19:27.720
But what they do is they discover a task and kind of start writing a prompt for this task from the plan.

s330
00:19:27.720 --> 00:19:34.880
They write the plan, they execute, and most importantly for my talk here, they verify themselves, and if it doesn't work, they try again.

s331
00:19:34.880 --> 00:19:40.480
So, an agent that loops grades its own work against the goal with less human intervention.

s332
00:19:40.480 --> 00:19:45.040
But if the builder grades itself, you didn't remove the review, you hid it.

s333
00:19:45.040 --> 00:19:48.000
Okay, this this connects to my routing table.

s334
00:19:48.000 --> 00:19:49.880
This comes from Adi Olsmanyu recently at Google.

s335
00:19:49.880 --> 00:19:52.360
He's also at this conference, a great engineer.

s336
00:19:52.360 --> 00:20:05.240
Uh he said if I if I if I wasn't reviewing the code myself, or relied entirely on automated loops to fix my code, let's say a bug comes up in Jira and my loop picks it up and starts fixing this, my product quality would suffer.

s337
00:20:05.240 --> 00:20:09.480
I'd likely end up in a downward spiral, digging myself into a deeper hole.

s338
00:20:09.480 --> 00:20:15.600
So again, loops don't remove judgment, but they do raise the stakes on where you put it.

s339
00:20:15.600 --> 00:20:17.520
So what about the future, folks?

s340
00:20:17.520 --> 00:20:21.040
Nobody knows.

s341
00:20:21.040 --> 00:20:26.960
Anthropic did not know that Claude code is going to explode on them and this is going to be a billion-dollar product.

s342
00:20:26.960 --> 00:20:32.680
Uh nobody knew that coding agents and Harness are going to be the generalized agent, and now everybody's pursuing them.

s343
00:20:32.680 --> 00:20:39.920
Folks at OpenAI with Codex, Elon with with Grok code, uh Google with with anti-gravity.

s344
00:20:39.920 --> 00:20:47.280
Model capability is jumping at an insane pace, and what I implore and to tell you here is that flexibility is required.

s345
00:20:47.280 --> 00:20:49.920
You need to keep being able to keep up with with the trends.

s346
00:20:49.920 --> 00:20:51.560
This is why you stay engineer.

s347
00:20:51.560 --> 00:20:55.600
And by the way, I told some folks here about my podcast Thursday and news.

s348
00:20:55.600 --> 00:21:05.440
If you want to keep tracking where that line moves, feel free to scan this QR code, join our uh our, you know, newsletter, etc. Um and I'll leave you with this.

s349
00:21:05.440 --> 00:21:07.200
Because it's my time, I'll leave you with this.

s350
00:21:07.200 --> 00:21:11.480
Not every line in 2026 needs your eyes.

s351
00:21:11.480 --> 00:21:13.320
Every system still needs your judgment.

s352
00:21:13.320 --> 00:21:15.520
Thank you.

s353
00:21:15.954 --> 00:21:17.954
[applause]