WWC26 - The Signal Layer: What to Build When Anything Can Be Built

[music]

>> Welcome back.

And thank you for being here. Hope you

had a good break. Hope you met somebody.

Hope you inspired somebody. Hope you

started a million dollar startup as SAS

using lovable. I don't know what you

did. Hopefully did something good. So

next up we have a person that I met in

2015 when I worked at Microsoft and I

just started there as the as a dev lead.

And

she was one of the people that was it

was a machine learning language

conference thing. And I tried to be as

clever as the people around them and she

just blew me away.

She also got me into coding in F sharp

which I have yet to do. But it's an

interesting language or was an

interesting language I guess. So now she

works for Akamai and she's a senior

director developer of AI

and Akamai.

And she's going to talk to us about the

signal layer signal noise kind of thing.

What to build when anything can be

built. So think about what you build

rather than just like let the machine do

it for you. Without further ado Lena

Hall please.

>> Hi.

Hello.

Hi Berlin. How are you?

How is the conference going for you?

Good?

Awesome.

Well, I think this was the best the most

productive year so far for many of us.

I'm Lena. A few days ago I solved a

production incident on a trail near a

waterfall.

My friend ran 18 agents while riding his

bike.

We are literally drowning in abundance.

We have more output, more speed, more

leverage than any of us have ever had.

So, why do we have that feeling like the

ground underneath us is moving too fast?

Last week I was speaking at AI Engineer

conference in San Francisco and one of

the engineers that I met there said that

it feels like the opportunity cost for

not working 9:00 a.m. till 9:00 p.m. 6

days a week is just too high.

We're all token maxing. We're working

all the time.

How many of you have already maxed out

your fables limits?

Okay.

And how many of you have already set an

alarm for 7:00 p.m. today exactly for

for one very specific reason?

Because this is when chat GPT-56 is

going to come out.

Of course.

So, look at us. Um, the same abundance

that made you fast, it made everyone

else fast. Now, everyone can build

everything.

Your competitor can build your feature

this afternoon, too.

And the cost of average just went to

zero, but so did its value.

A year ago, the superpower was, as we

were told,

being good at using AI.

But the 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.

AI gives everybody the same answer

because everyone is asking it the same

question.

It runs on data and data is a record of

what already happened.

So, when you point AI at a task and tell

it

"Tell me what users want." or "Make more

money." or "What should we build?" or

"Make this viral." it answers from the

common knowledge.

Competently, confidently, but also

identically to what it tells your

competitor.

To see something that the 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 of that automation to

execute on it.

AI is a really smart convergence

machine. If you leave it alone,

it makes everything the same.

There is one decision though that AI

can't and shouldn't make for you.

It is to decide what to point at.

So, the job the new job for every one of

us is deciding what it makes and being

the reason the right people choose your

version over the identical-looking rest.

But also, I'm sure many of you walked

around the expo hall at this conference

or at any other conference.

There's so many amazing products, so

many tools and vendors. They're all

solving important problems, but why do

they all sound the same?

So, when anyone can build anything, what

makes me different? What makes you

different? Why should anyone pick your

version, your product?

I call this work the signal layer, and

there are two halves to getting this

right. So, that's how we'll walk through

it.

The first half is knowing your signal,

being able to define it very clearly.

What you're building and why it's yours

and it's not the average. That's the

build side. It's the code, the product,

and the road map.

And the second half is emitting that

signal without distortion. So, making

sure that

what your customers come to believe

about you actually matches what you

believe and what you have built. That's

the ship side, the content, the

go-to-market engineering.

And uh I think I've had an unusual

vantage point on this. I've built

products as an engineer. I've created my

own as a founder, and I brought other

people's products to market. These are

very different jobs with one identical

challenge. The signal doesn't always

survive the trip. So, let's start with

the build side.

So, what do we even work on?

Everything is implementable.

2 years ago, the best autonomous coding

agent only solved a fraction of the

tasks on the standard software

benchmark. And now, the best agents are

in the high 80s.

We nearly tripled the writing,

but shipping actually barely moved a

third. The benchmark was measuring the

part of software engineering that has a

grader.

Uh and shipping is where all the

ungraded parts come back in.

Here is the rule underneath it.

Anything that you can measure, you can

train against, as Sarah Guo puts this.

A compiler is a free grader.

A test suite is a free grader.

And the instant a task can grade itself,

you can grind a model against the grade

until it wins.

And automation of code was first because

it's really the most checkable thing

that we have.

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

the same thing.

The model will build whatever you point

it at, but it will tell you nothing

about where to point it.

Anything that's visible is replicable.

Now, some people, when they hear

everything is implementable, they panic.

But we can flip the question.

The pointing is the job. The pointing

has always been the job. We just had so

much implementation work in the way that

we never get had to get good at it.

So, how do you decide where to point at?

Paul Graham

shared some wisdom about this.

The way you find something people

genuinely want is by feeling the need

yourself.

Build something that you or your friends

need, because the market hasn't formed

yet. The surveys can't see it, and your

own need is the only early signal that

isn't a crap signal.

And the best ideas may sound genuinely

lame at first, like a guy strapped

camera to his head live-streaming his

life. That sounds ridiculous, but it

became Twitch. Um and the convergence

machine generally doesn't proactively

offer you these weird, specific,

slightly embarrassing ideas.

But even with the Twitch example, it

worked. But a thousand other similar

ideas

uh for startups, they didn't.

The weird, specific signal is necessary,

but it's not sufficient.

It's really tempting to say that we just

need to have good judgement or good

taste and call it safe.

But taste is really just preference

under feedback. And preference under

feedback is exactly what those systems

can learn.

Anything that you can demonstrate enough

times with a better or worse signal

attached, the machine can eventually

imitate. So, the broad definition of

good taste is not really a

differentiator.

What actually resists training is more

narrow and more durable. So, two things.

Taste and judgement about what hasn't

happened yet because there is no data

for an event that hasn't occurred.

And taste and judgement embedded in a

relationship that the model can't

observe directly. So, what this customer

in this situation

with this history that you share

actually needs.

The model has read everything about your

customer, but it has never actually met

them.

So, if broad judgement isn't safe and AI

just handed everyone the ability to

build anything, so what's left to be

good at?

Richard Hamming spent his career

studying why some scientists did great

work and others, just as smart, didn't.

He found that the great ones worked on

important problems.

And the problem isn't important because

it just sounds impressive.

It's important when you have a

reasonable attack on it. For example,

time travel

is consequential, but he would say it's

not important because nobody has an

attack.

And Hamming would tell you to keep 10 to

20

important problems live in the back of

your mind

so that when you finally have an attack,

a new tool, a new angle, a thing that

only you noticed, you go for it.

But in having's world, the rare thing

was having an attack.

And AI just gave

everyone an attack on everything. So,

the rare thing is knowing which problem

is actually worth attacking.

And that judgment comes from being a

real person, close to a real domain,

with your own battle scars, your weirdly

specific experience,

and the thing that you care about more

than is reasonable.

So, you don't need to be first. You do

need to be genuinely close to a problem

you actually understand, and where your

insight is in the delta between what AI

has been trained on and what should

exist.

So, let's say you did it. You found a

sweet spot problem, the one that you had

an honest attack on, and you build the

thing, and it's genuinely good, and it's

genuinely yours, and it's not average.

You can still lose because knowing your

signal is only half the job. The other

half is getting it from your head into

the head of the person it was meant for.

It's about reaching the right people.

And what do most of us do for that?

We make content. So, let's talk about

what AI convergence machine does to

that.

What happened to the internet in the

last 2 years?

Um you open any feed, everything has

started to sound the same.

The same LinkedIn posts, the same three

bullet points and a bold takeaway,

the same blog post that says nothing but

in a very polished way.

Um and your readers can now pattern

match AI in just half a second. So, if a

model could have written your post from

a one-line prompt, your reader's brain

just skips it for the same reason.

AI has really learned the algorithm. It

has learned the format for each platform

where it performs. It has learned what

gets clicks.

And everyone wants to hand the machine a

paragraph and say, "Make this viral.

Make me rich." Um it will fill every gap

that you leave with sameness, though.

So, what do you put in and what do you

let it fill in?

Because there are two completely

different ways that we can use this

thing.

And they look identical from the

outside. One is you give it an average

prompt and it gives you an average

output.

And you ship one more indistinguishable

drop into an ocean of indistinguishable

drops.

So, you've automated your own

irrelevance very efficiently.

And two is you bring the part that it

can't have. Your specific point of view,

the thing that you believe that the

training data doesn't. The real story

that you were actually in the room for.

And then you let the machine do the

converging work, the formatting, the

drafting, the algorithm optimization,

the cleanup around the the core that it

could have never generated.

Because the signal distorts on the way

out.

So, you can have the signal perfectly

clear to you and still watch it fall

apart between your brain and your user's

understanding of it.

And in my experience, it breaks in three

places.

And there are fixes for each, but

they're very different depending on the

company, the product, the type, and the

size of of those things.

One of them is source distortion, very

common in startups.

Founders usually know the signal of

their product of their company so well

that they always have this accidental

gift of compressing it

past legibility.

So, they often assume the context that

the audience doesn't have, and the room

hears something really cool, but doesn't

understand why it matters.

I helped this one YC company recently

with this.

Brilliant founders, genuinely great

product, but every pitch that they had

started with the architecture, with the

clever parts, the things that they were

genuinely proud of.

But it landed as noise to the customers

because the customer pain has been

deleted from the story.

So, we rewrote the opening to include

the thing that users hated the most, and

this product actually solved.

So, same product, same week, and the

next conversations turned into pilots,

and we then turned it into a repeatable

GTM system.

Organization distortion is another type

of distortion that almost every big

company has.

So, as signal travels through layers of

management, through legal, through

sales, through every department, at

every handoff, it gets re-rounded toward

uh the average.

It doesn't really come from

incompetence. It comes from the level of

investment.

And the founder and the person that's

three layers down the organizational

layer system, the same task and the same

AI, and you will get two completely

different uh outputs.

The founder sweats the unaverageable

details because the outcome is theirs.

They're personally invested and affected

by it.

But others ship to spec because they

were asked for compliance and not

conviction. And AI is great at clearing

tickets.

So, a long delegation chain plus a

convergence machine is really a factory

for automating the signal right out of

your own company.

The first instinct is just to add

process, um which adds layers, adds

bureaucracy, and slows everything down.

And we don't really want that. So, to

fix this, we need a we need to take the

signal back, reattach it to the outcome

like a founder,

and add a very thin signal layer to your

go-to-market engineering, where its only

job is to carry the original intent

across um all the layers and keep the

signal intact.

Machine distortion is another way that

you can lose signal.

You can write one very careful launch

announcement, for example. Your claim,

your evidence, your scope are all clear.

And then, of course, AI

uh remixed it into a tweet, into um a

sales deck, into a partner one-pager.

For example, you might have had one

narrow eval that scored 94%.

But it was repeated enough times that

your customers actually started hearing

it as a promise.

So, we see the same through line. Your

signal has to survive the trip

undistorted.

And this is something you can engineer.

So, we need a thin signal layer, a small

deliberate function whose job is to make

sure that your that what your users take

away is still the specific thing that

you meant.

So, we can make this more concrete.

Let's say you're building a monitoring

tool, and there are 12 other tools in

the same category. But, yours does

something different.

Um it tells you what not to wake up for.

It stays quiet on the noise. So, when it

does page you at night,

you actually believe it.

So, that quiet, that trust earned by

silence, that is your signal.

So, first, we need to say it in one

sentence with the limit of it built in.

Definitely not this, not intelligent AI

native observability platform.

Something more like uh stays quiet at

anything it can't tie to real user

impact, and shows you everything it

silenced, so you can override it. So,

the promise and the scope are welded

together.

And then, make sure that the limit can't

be edited out.

In the product, every suppressed alert

is visible. Um statements like 90% fewer

pages live next to statements like every

silence is visible and reversible. So,

when AI inevitably chops your launch

into a tweet, it can't keep the

impressive number and remove the part

that keeps your product honest.

And before you scale it, um check what

people actually heard. Give the read me

to an SRE who has never seen the

project, and ask um another person to

describe the product back to you. So,

the gap between what they say and what

you meant is exactly the distortion that

you were about to broadcast.

So, that's the lightweight signal layer,

and a lot of it is buildable. You can

automate more of the checking and

catching than most people realize.

But, if we step back and ask what all of

this, the building, the shipping, the

undistorted signal is actually for,

it's for one thing. It's for getting a

human, or increasingly an agent, to

choose you and rely on you when they

have infident identical looking

alternatives. And that's trust. And

trust is one thing left with no greater.

There's no benchmark for it. There is no

uh reward signal.

It cannot be entirely automated because

it's granted slowly over time. It's a

relationship with consent.

For example, doctors who open one

particular tool every morning,

they didn't have that habit trained into

them.

And what happens if we get this wrong?

Getting your signal wrong isn't actually

neutral. It's negative.

Producing averageness is is not free.

You actually pay for it

in tokens, in infrastructure, in the

hours of good people

with the customers that look at your

product, decide once, and never come

back. So, every generic post, every

average functionality teaches them that

your name is not worth the click. So,

you spend real money to make yourself

harder to choose.

So, back to the main question.

>> [clears throat]

>> We got faster.

But the speed is not where the value

went. The value moved up to deciding

what is worth building, what is worth

saying, what deserves trust. And whether

the thing that you

meant actually survives the trip to the

people that it was for.

And you don't need to be first. You just

need a real problem. You have enough

conviction to carry the signal clearly

enough so that the the right people can

find it.

So, when you can build anything, build

trust.

So, have the strongest conviction,

define the signal yourself, and protect

it from distortion at costs and use AI

aggressively to do anything else.

Thank you. Uh let's connect. I'll be

around after the talk. I'll be happy to

chat with you and make sure to stop by

Apple my cloud booth for some amazing

demos and swag.

Thank you.

>> [applause]

>> Thank you so much, Lena. Uh you talked

to me earlier and said you don't want to

do a Q&A although you got lots of time

left, but you want to do a more personal

Q&A. So, please after she's been

un-miked, uh

grab her here, not grab her here. Talk

to her here and actually get all your

questions answered about what to get as

a proper signal. Thanks very much, Lena.

This was always a great

>> Thank you.

>> And here's another person I met at the

same event.

Cool. So,