WEBVTT

NOTE Sentence-level transcript of https://www.youtube.com/watch?v=1KOdiGgMtpY

NOTE One cue per sentence. Cue ids are the line anchors on /transcripts/1KOdiGgMtpY.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
00:00:12.560 --> 00:00:16.080
How is the conference for all of you so far?

s3
00:00:16.080 --> 00:00:16.720
Great.

s4
00:00:16.720 --> 00:00:18.280
Awesome.

s5
00:00:18.280 --> 00:00:25.560
Um well, I think this was the best most productive year for so many of us.

s6
00:00:25.560 --> 00:00:26.560
I'm Lena.

s7
00:00:26.560 --> 00:00:32.960
A few days ago, I solved a production incident on a trail near a waterfall.

s8
00:00:32.960 --> 00:00:37.960
My friend ran 18 agents while riding his bike.

s9
00:00:37.960 --> 00:00:40.800
We're literally drowning in abundance.

s10
00:00:40.800 --> 00:00:47.360
We have more output, more speed, more leverage than any of us have ever had.

s11
00:00:47.360 --> 00:00:54.520
So, why do we have this feeling like the ground underneath is moving too fast?

s12
00:00:54.520 --> 00:01:07.600
One of the engineers that I met at this conference said yesterday that it feels like the opportunity cost for not working 9:00 a.m. to 9:00 p.m. 6 days a week is too high right now.

s13
00:01:07.600 --> 00:01:09.200
So, we're all token maxing.

s14
00:01:09.200 --> 00:01:12.400
We're all working all the time.

s15
00:01:12.400 --> 00:01:17.920
But the same abundance that made you fast, it also made everyone else fast.

s16
00:01:17.920 --> 00:01:21.560
So, now everyone can build everything.

s17
00:01:21.560 --> 00:01:25.520
Your competitor can build your feature this afternoon, too.

s18
00:01:25.520 --> 00:01:31.040
So, the cost of the average just went to zero and so did its value.

s19
00:01:31.040 --> 00:01:37.360
A year ago, the superpower, as we were told, was to be good at using AI.

s20
00:01:37.360 --> 00:01:49.520
But models got so good and they got so easy and everybody now is a lot more skilled at using AI and everybody's pointing AI at the same goals.

s21
00:01:49.520 --> 00:01:55.360
Cuz AI gives everyone the same answer because everybody is asking the same question.

s22
00:01:55.360 --> 00:02:00.200
It It on data and data is a record of what has already happened.

s23
00:02:00.200 --> 00:02:04.680
So, when you point AI at tasks like, "Tell me what users want.

s24
00:02:04.680 --> 00:02:05.760
Make more money.

s25
00:02:05.760 --> 00:02:07.120
What should we build?

s26
00:02:07.120 --> 00:02:08.640
Make this viral."

s27
00:02:08.640 --> 00:02:11.440
It answers from the common knowledge.

s28
00:02:11.440 --> 00:02:17.480
Very competent, very confident, but also very identical to what it tells your competitor.

s29
00:02:17.480 --> 00:02:28.600
To see something that data doesn't show yet, we need to have a vision, a point of view, a read on where it's going, and then use all that automation to execute it.

s30
00:02:28.600 --> 00:02:32.320
AI is a really smart convergence machine.

s31
00:02:32.320 --> 00:02:37.040
So, if you leave it alone, it makes everything the same.

s32
00:02:37.560 --> 00:02:41.200
There is one decision, though, that AI can't and shouldn't make for you.

s33
00:02:41.200 --> 00:02:44.920
It is to decide what to point at.

s34
00:02:44.920 --> 00:02:55.920
So, the job, the new job for everyone of us is deciding what it makes, being the reason the right people choose your version over the identical-looking rest.

s35
00:02:55.920 --> 00:03:04.680
But also, I'm sure many of you uh walked around the Expo Hall at this conference, and there are so many amazing products, so many tools and vendors.

s36
00:03:04.680 --> 00:03:07.040
They're all solving important problems.

s37
00:03:07.040 --> 00:03:11.200
But what Why do they all sound the same?

s38
00:03:11.519 --> 00:03:15.160
So, when anyone can build anything, what makes me different?

s39
00:03:15.160 --> 00:03:16.920
What makes you different?

s40
00:03:16.920 --> 00:03:20.720
Why should anyone pick your version, your product?

s41
00:03:20.720 --> 00:03:24.280
Um I call this work uh signal layer.

s42
00:03:24.280 --> 00:03:28.640
And there are two There's two halves to solving it, to getting this right.

s43
00:03:28.640 --> 00:03:31.600
So, that's how we will walk through it.

s44
00:03:31.600 --> 00:03:37.120
The first half is knowing your signal, being able to define it very clearly.

s45
00:03:37.120 --> 00:03:40.920
What you're building and why it's yours and not the average.

s46
00:03:40.920 --> 00:03:45.320
So, that's the build side, the code, the product, the road map.

s47
00:03:45.320 --> 00:03:50.239
And the second half is emitting that signal without distortion.

s48
00:03:50.239 --> 00:04:00.440
So, making sure that your customers um making sure what your customers come to believe about you actually matches what you believe and what you've built.

s49
00:04:00.440 --> 00:04:05.200
That's the ship side the content and go to market engineering.

s50
00:04:05.400 --> 00:04:08.280
And I've had an unusual vantage point in this.

s51
00:04:08.280 --> 00:04:10.200
I've built products as an engineer.

s52
00:04:10.200 --> 00:04:12.400
I created my own as a founder.

s53
00:04:12.400 --> 00:04:15.280
I brought other people's products to market.

s54
00:04:15.280 --> 00:04:19.760
So three very different jobs with one identical challenge.

s55
00:04:19.760 --> 00:04:22.600
The signal doesn't always survive.

s56
00:04:22.600 --> 00:04:25.840
So let's start with the build side.

s57
00:04:25.840 --> 00:04:27.800
So what do we even work on?

s58
00:04:27.800 --> 00:04:30.360
Everything is implementable.

s59
00:04:30.960 --> 00:04:42.520
Two years ago the best autonomous coding agents you know solved on the fraction of the tasks on the standard software benchmark and now the best agents are in the high eighties.

s60
00:04:42.520 --> 00:04:49.520
So we nearly tripled the amount of writing and shipping barely moved a third.

s61
00:04:49.520 --> 00:04:59.640
The benchmark was measuring the part of software engineering that has a greater and shipping is where all the ungraded parts come back in.

s62
00:04:59.640 --> 00:05:03.560
So here is the rule underneath it.

s63
00:05:03.560 --> 00:05:09.320
Anything that you can measure you can train against as Sarah Guo puts it.

s64
00:05:09.320 --> 00:05:11.600
A compiler is a free grader.

s65
00:05:11.600 --> 00:05:14.080
A test suite is a free grader.

s66
00:05:14.080 --> 00:05:22.240
And the instant a task can grade itself you can grind a model against you know that grade until it wins.

s67
00:05:22.240 --> 00:05:28.880
Automation of code was first because it's the most checkable thing that we have.

s68
00:05:28.880 --> 00:05:40.400
So implementation is converging for free for everyone at the same time and the most buildable thing and the most valuable thing are almost never the same thing.

s69
00:05:40.400 --> 00:05:46.120
So the model will build whatever you point it at but it will tell you nothing about where to point.

s70
00:05:46.120 --> 00:05:49.520
Anything visible is replicatable.

s71
00:05:50.120 --> 00:05:54.680
So now, some people when they hear everything is implementable, they panic.

s72
00:05:54.680 --> 00:05:56.920
Um but we can flip the question.

s73
00:05:56.920 --> 00:05:59.320
The pointing is actually the job.

s74
00:05:59.320 --> 00:06:01.480
It has always been the job.

s75
00:06:01.480 --> 00:06:08.360
We just had so much implementation work in the way um that we never had to get good at it.

s76
00:06:08.360 --> 00:06:11.400
So, how do you decide where to point that?

s77
00:06:11.400 --> 00:06:14.840
Paul Graham shared some wisdom on this.

s78
00:06:14.840 --> 00:06:20.680
Where you find something that people genuinely want is by feeling the need yourself.

s79
00:06:20.720 --> 00:06:31.480
Build something you and your friends need because the market hasn't formed yet, surveys can't see it, and your own need is the only signal that isn't a crap signal.

s80
00:06:31.480 --> 00:06:41.640
And the best ideas may sound genuinely lame at first, like a guy uh strapped with a with a camera on his head live streaming his his life.

s81
00:06:41.640 --> 00:06:44.600
That sounds really ridiculous, but it became Twitch.

s82
00:06:44.600 --> 00:06:54.040
Um and the convergence machine doesn't really, you know, propose proactively these uh weird, specific, genuinely embarrassing ideas.

s83
00:06:54.040 --> 00:07:03.919
But even with the Twitch example, it worked, but there were a thousand other similar startups I start startup ideas that didn't.

s84
00:07:03.919 --> 00:07:09.560
So, the weird specific signal is necessary, but it is not sufficient.

s85
00:07:10.080 --> 00:07:17.880
It's also really tempting to say that we just need to have good judgment and good taste and call it safe.

s86
00:07:17.880 --> 00:07:26.480
But taste is really just preference under feedback, and preference under feedback is exactly what these systems can learn.

s87
00:07:26.480 --> 00:07:34.960
Anything you can demonstrate enough times uh with a better or worse signal attached, the machine can eventually imitate.

s88
00:07:34.960 --> 00:07:39.200
So, broad good taste is not really a differentiator.

s89
00:07:39.560 --> 00:07:44.720
What actually resists training is more narrow and more durable.

s90
00:07:44.720 --> 00:07:46.520
So, two things.

s91
00:07:46.520 --> 00:07:50.480
Taste and judgement about what hasn't happened yet.

s92
00:07:50.480 --> 00:07:53.280
Because there's no data for an event that hasn't occurred.

s93
00:07:53.280 --> 00:07:58.440
And then taste and judgement embedded in a relationship that the model can't observe.

s94
00:07:58.440 --> 00:08:05.680
What this customer in this situation with this history that you share actually needs.

s95
00:08:05.680 --> 00:08:12.320
The model has read everything ever written about your customer, but it has never actually met them.

s96
00:08:12.320 --> 00:08:21.760
So, if broad judgement is not safe, and the AI just handed everyone the ability to build anything, what's left to be good at?

s97
00:08:21.760 --> 00:08:32.200
Richard Hamming uh spent his career studying why some scientists did great work and others who were just as smart didn't.

s98
00:08:32.200 --> 00:08:37.080
He found that the great ones worked on important problems.

s99
00:08:37.520 --> 00:08:42.080
And the problem isn't important because it just sounds impressive.

s100
00:08:42.080 --> 00:08:45.480
It's important when you have a reasonable attack on it.

s101
00:08:45.480 --> 00:08:52.640
For example, time travel is consequential, he would say, but it's not important because nobody has an attack on it.

s102
00:08:52.640 --> 00:09:04.000
So, Hamming would tell you to keep 10 or 20 ideas um on important problems in the back of your mind so that when you finally have an attack, a new tool, a new angle

s103
00:09:04.000 --> 00:09:08.040
I think that only you noticed, then you go for it.

s104
00:09:08.040 --> 00:09:12.000
But in Hamming's world, the rarest thing was having an attack.

s105
00:09:12.000 --> 00:09:15.960
And AI just handed everyone an attack on everything.

s106
00:09:15.960 --> 00:09:19.920
So, the rarest thing is knowing which problem is actually worth attacking.

s107
00:09:19.920 --> 00:09:32.320
And that judgement comes from being a real person, close to a real domain, with your own battle scars, your weirdly specific experience, the thing that you care about more than is reasonable.

s108
00:09:32.320 --> 00:09:34.320
And you don't actually need to be first.

s109
00:09:34.320 --> 00:09:45.200
You just need to be genuinely close to a problem you actually understand where your insight is in the delta between what AI has been trained on and what should exist.

s110
00:09:45.200 --> 00:09:46.760
So, let's say you did it.

s111
00:09:46.760 --> 00:09:53.520
You found that sweet spot problem that the one that you had an honest attack on, that you built this thing.

s112
00:09:53.520 --> 00:09:54.520
It's genuinely good.

s113
00:09:54.520 --> 00:09:58.120
It's genuinely yours, not the average.

s114
00:09:58.120 --> 00:10:03.760
You can still lose uh because knowing your signal is only half the job.

s115
00:10:03.760 --> 00:10:09.880
The other half is getting it from your head into the head of a person that it was meant for.

s116
00:10:09.880 --> 00:10:12.200
And it's about reaching the right people.

s117
00:10:12.200 --> 00:10:15.720
And what do most of us do for that?

s118
00:10:15.720 --> 00:10:16.880
We make content.

s119
00:10:16.880 --> 00:10:22.760
So, let's talk about what uh AI convergence machine does to that.

s120
00:10:23.280 --> 00:10:26.240
What happened to the internet in the last 2 years?

s121
00:10:26.240 --> 00:10:30.760
Open any feed, everything has started to sound the same.

s122
00:10:30.760 --> 00:10:42.080
The same LinkedIn posts, the same, you know, three bullet points and a bold takeaway, and the same blog post that uh says nothing but actually looks very polished.

s123
00:10:42.120 --> 00:10:46.960
Um your readers can now pattern match AI in just half a second.

s124
00:10:46.960 --> 00:10:56.480
So, if a model could have written your post from a one-line prompt, your reader brain just skips it for the same reason.

s125
00:10:56.800 --> 00:10:58.720
So, AI has really learned the algorithm.

s126
00:10:58.720 --> 00:11:00.839
It has learned the format that performs.

s127
00:11:00.839 --> 00:11:07.680
It has learned what gets clicks, and everyone wants to hand the machine a paragraph and say, you know, "Make it viral.

s128
00:11:07.680 --> 00:11:09.000
Make me rich."

s129
00:11:09.000 --> 00:11:13.560
It fills every gap that you leave with sameness.

s130
00:11:13.560 --> 00:11:17.200
So, what do you put in and what do you let it fill in?

s131
00:11:17.200 --> 00:11:24.120
Cuz these there there are two different ways to use this thing, and they look very identical from the outside.

s132
00:11:24.120 --> 00:11:28.360
One is you give it an average prompt, and gives you the average output.

s133
00:11:28.360 --> 00:11:34.320
And you ship one more indistinguishable drop into an ocean of indistinguishable drops.

s134
00:11:34.320 --> 00:11:38.880
So, you've automated your own irrelevance very efficiently.

s135
00:11:38.880 --> 00:11:49.320
And two, you can bring in the part that it can't have, your specific point of view, the real story that you were actually in the room for, and then let the machine do the converging work,

s136
00:11:49.320 --> 00:11:57.640
the formatting, the drafting, the algorithm optimization, the cleanup around the core that it could have never generated.

s137
00:11:58.000 --> 00:12:00.960
The signal distorts on the way out.

s138
00:12:00.960 --> 00:12:10.600
So, you can have the signal perfectly clear for you and still watch it fall apart between your brain and your users' understanding of it.

s139
00:12:10.600 --> 00:12:14.080
And in my experience, it breaks in three places.

s140
00:12:14.080 --> 00:12:22.040
And there are fixes for each, but they're different depending on product, the type, and the size of the company.

s141
00:12:22.040 --> 00:12:26.920
One of them is source distortion, which is very common in startups.

s142
00:12:26.920 --> 00:12:37.400
Founders, actually, they usually know the signal so well that they always have this accidental gift of compressing it past legibility.

s143
00:12:37.400 --> 00:12:48.240
They often assume the context that the audience doesn't have, and the room hears something technically very cool, but they doesn't they don't really understand why it matters.

s144
00:12:48.240 --> 00:12:53.440
I helped this one YC company with uh this exact thing recently.

s145
00:12:53.440 --> 00:13:05.880
Absolutely brilliant founders, genuinely new product, but every pitch that they started um was you know starting with architecture, with the clever parts, with things that they were very proud of.

s146
00:13:05.880 --> 00:13:12.480
But it really landed as noise because the customer pain has been deleted from the whole story.

s147
00:13:12.480 --> 00:13:18.720
So, we rewrote the opening to include the thing that the user hated, and this product actually killed.

s148
00:13:18.720 --> 00:13:28.920
So, the same product, the same week, the next conversations converted into pilots, and then we turned that into repeatable GTM system.

s149
00:13:28.920 --> 00:13:35.400
Organization distortion is another type of distortion that almost every big company has.

s150
00:13:35.400 --> 00:13:46.520
As signal travels through layers of management, through legal, through sales, through every department, at every hand handoff, it gets rewound towards the average.

s151
00:13:46.520 --> 00:13:50.960
And this really doesn't come from incompetence, it comes from investment.

s152
00:13:50.960 --> 00:13:59.520
So, hand a founder and the person three layers down the same task and the same AI, and you get two different things.

s153
00:13:59.520 --> 00:14:09.200
Um the founder really sweats the unaverageable details because the outcome is really theirs and they're personally invested and affected by it.

s154
00:14:09.200 --> 00:14:18.360
And others just ship it to spec, they close Jira tickets, they were asked for, you know, something like compliance, not as much conviction.

s155
00:14:18.360 --> 00:14:27.960
So, a long delegation chain plus convergence machine is really a factory for automating the signal out of your own company.

s156
00:14:27.960 --> 00:14:35.640
So, the first instinct usually is to add more process, which adds layers, bureaucracy, and slows everything down.

s157
00:14:35.640 --> 00:14:37.240
And we don't want that.

s158
00:14:37.240 --> 00:14:48.320
Um to fix this, we have to help take the signal back and reattach it to the outcome like a founder and add the very thin signal layer to your go-to-market engineering,

s159
00:14:48.320 --> 00:14:55.760
where its only job is to validate and carry the original intent across the handoffs intact.

s160
00:14:55.760 --> 00:14:59.040
Machine distortion is another way you can lose signal.

s161
00:14:59.040 --> 00:15:11.520
You write one careful launch, your claim, your evidence, and your scope is very clear, but then of course AI remixes it um into a tweet, into a sales deck, into a partner one-pager.

s162
00:15:11.520 --> 00:15:23.600
For example, you might have had this one very narrow eval that scored 94% but it was repeated enough times that your customers actually heard it as a promise.

s163
00:15:23.840 --> 00:15:25.280
So, we see the same through line.

s164
00:15:25.280 --> 00:15:29.920
Your signal has to survive the trip undistorted.

s165
00:15:30.360 --> 00:15:33.360
And this is something you can engineer.

s166
00:15:33.360 --> 00:15:43.640
So, we need a thin signal layer, a small deliberate function whose job is to make sure that what your users take away is still the specific thing you meant.

s167
00:15:43.640 --> 00:15:45.880
Say you're building a monitoring tool.

s168
00:15:45.880 --> 00:15:50.880
There are 12 other tools in this category, but yours does something different.

s169
00:15:50.880 --> 00:15:54.480
It tells you what not to wake up for, for example.

s170
00:15:54.480 --> 00:16:00.640
It stays quiet on the noise, so when you get paged at night, you believe it.

s171
00:16:00.640 --> 00:16:05.200
So, that quiet, that trust earned by silence is your signal.

s172
00:16:05.200 --> 00:16:09.000
So, first, say it in one sentence with the limit built in.

s173
00:16:09.000 --> 00:16:14.760
Definitely don't say intelligent AI-native observability platform.

s174
00:16:14.760 --> 00:16:24.360
Say something like uh stays quiet on anything it can't tie to a real user impact and shows you everything it silenced so you can overrule it.

s175
00:16:24.360 --> 00:16:28.080
The promise and the scope are welded together here.

s176
00:16:28.080 --> 00:16:30.760
Then make sure that the limit can't be edited out.

s177
00:16:30.760 --> 00:16:34.640
So, in the product, every suppressed alert is visible.

s178
00:16:34.640 --> 00:16:43.600
In the launch, statements like 90% fewer pages uh live next to statements like every silence is visible and reversible.

s179
00:16:43.600 --> 00:16:55.520
So, when AI chops your launch into a tweet, it can keep the impressive number, but also remove the part that keeps the part that uh keeps the product honest.

s180
00:16:55.520 --> 00:16:58.880
And before you scale it, check what people actually heard.

s181
00:16:58.880 --> 00:17:07.160
So, give a readme to an SRE who has never seen your project and ask a person to describe the product back to you.

s182
00:17:07.160 --> 00:17:14.040
The gap between what they say and what you meant is the distortion that you were about to broadcast.

s183
00:17:14.040 --> 00:17:23.480
And it's a very lightweight signal layer, and a lot of it is buildable, So, you can automate more of the checking and the catching and the surveying than most people realize.

s184
00:17:23.480 --> 00:17:31.320
So, step back and ask what all of this, the building, the shipping, the undistorted signal, is actually for.

s185
00:17:31.320 --> 00:17:42.800
It's for one thing of getting a human or increasingly an agent to choose you and rely on you when they have an infinite identical-looking alternatives.

s186
00:17:42.800 --> 00:17:44.400
So, that's trust.

s187
00:17:44.400 --> 00:17:47.320
Trust is the one thing that's left with no greater.

s188
00:17:47.320 --> 00:17:50.520
There's no benchmark for it, no reward signal.

s189
00:17:50.520 --> 00:17:57.600
It can't be entirely automated because it's granted slowly through relationship with consent.

s190
00:17:57.600 --> 00:18:05.280
For example, doctors who open one particular tool every morning, they didn't have that habit trained into them.

s191
00:18:05.280 --> 00:18:08.040
And what happens if we get this wrong?

s192
00:18:08.040 --> 00:18:11.160
Getting your signal wrong is actually not neutral.

s193
00:18:11.160 --> 00:18:12.880
It's negative.

s194
00:18:12.880 --> 00:18:15.160
Producing averageness is not free.

s195
00:18:15.160 --> 00:18:27.920
You actually pay for it in tokens, in infra, in the salaried hours of good people, you know, with with customers that take a look at your product once, decide once, and never come back.

s196
00:18:27.920 --> 00:18:33.320
So, every generic post teaches them that your name isn't worth the click.

s197
00:18:33.320 --> 00:18:37.360
So, you spend real money to make yourself harder to choose.

s198
00:18:37.360 --> 00:18:39.440
So, back to the main question.

s199
00:18:39.440 --> 00:18:44.520
We got faster, but the speed is not where the value went.

s200
00:18:44.520 --> 00:18:50.800
Uh the value moved up to deciding what is worth building, what is worth saying, what deserves trust.

s201
00:18:50.800 --> 00:18:58.120
And where does the thing that you actually um that that you meant survives the trip to the people that it was for.

s202
00:18:58.120 --> 00:18:59.560
So, you don't need to be first.

s203
00:18:59.560 --> 00:19:08.160
You need a real problem and enough conviction to carry the signal clearly through to, you know, right people to find it.

s204
00:19:08.160 --> 00:19:12.120
So, when you can build anything, you should build trust.

s205
00:19:12.120 --> 00:19:21.160
Have the strongest conviction, define the signal yourself, protect it from distortion, and use AI aggressively for everything else.

s206
00:19:21.160 --> 00:19:21.840
Thank you.

s207
00:19:21.840 --> 00:19:24.800
Let's connect and happy to chat with you afterwards.

s208
00:19:24.800 --> 00:19:26.568
Thank you.

s209
00:19:26.568 --> 00:19:28.568
[applause]

s210
00:19:40.323 --> 00:19:42.323
[music]