AI Engineer World's Fair 2026
The Signal Layer: What to Build When Anything Can Be Built — Lena Hall, Akamai
Loading slides…
Read the talk
The Signal Layer: What to Build When Anything Can Be Built
Lena Hall argues that cheap implementation shifts the hard work upstream to choosing a worthwhile problem—and downstream to preserving its specific meaning until customers understand and trust it.
From a talk by Lena Hall
At a glance
Ideas worth remembering
Repeatable graders make implementation easier to optimize, but they do not decide whether the underlying problem is worth solving.
Hall places differentiation in specific domain experience, emerging needs, and customer relationships the model cannot fully observe; unusual ideas alone do not guarantee success.
For AI-assisted communication, supply firsthand substance and point of view before delegating drafting, formatting, optimization, and cleanup.
Protect the signal against three different failures: missing customer context at the source, dilution across organizational handoffs, and machine repackaging that separates claims from their scope.
Attach limits to promises in both product behavior and messaging, then ask an unfamiliar reader to explain the product back before scaling distribution.
Generic output has a real cost: it spends compute, labor, and future attention. Hall’s ultimate objective is trust earned through a clear, faithfully delivered signal.
Abundance makes average work cheap
Lena Hall opens with two absurdly practical signs of abundance: she resolved a production incident from a trail near a waterfall, while a friend ran 18 agents from his bike. Yet the resulting leverage has not produced calm. One engineer described feeling that the opportunity cost of working less than 9 a.m. to 9 p.m., six days a week, was too high. More capacity has become pressure to consume still more capacity.
The competitive catch is that everyone received the leverage at once. If a competitor can reproduce a feature that afternoon, competent implementation stops being scarce. Hall sharpens this into a claim that the cost—and therefore the value—of average work approaches zero. Later she qualifies the economics: generic output still consumes tokens, infrastructure, salaries, and audience attention. “Zero” here describes lost differentiation, not zero production expense.
Hall calls AI a convergence machine. Models learn from records of what has already happened, so broad requests such as what users want, what to build, or how to make something viral tend to draw from common knowledge. The result may be competent and confident while resembling the answer a competitor receives. Automation can execute a point of view, but it does not supply the point of view merely because someone asked a generic strategic question.
The expo hall makes the consequence visible: many products address important problems, yet their descriptions sound alike. Hall names the missing work the signal layer, with two linked responsibilities. The build side defines what the product is and why it is specifically yours; the ship side ensures that customers come to believe the same thing the team believes it built. Across her work as an engineer, founder, and go-to-market operator, the recurring failure was that this signal did not always survive.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Graders optimize execution, not direction
Hall contrasts coding-agent benchmark progress with the smaller apparent improvement in shipping software. She cites leading benchmark results reaching the high 80s, but the recording does not name the benchmark or define the shipping measure behind the comparison. The useful distinction does not depend on those missing details: a benchmark grades a bounded portion of engineering, while shipping reintroduces product choices, integration, operations, and other work without a simple answer key.
A compiler or test suite supplies repeatable feedback. Once a task can grade itself, developers or training systems can repeatedly optimize against that result. Code was especially open to automation because so much of it is mechanically checkable. This explains why implementation capability can converge quickly without answering the product question: passing a test establishes that a chosen behavior was implemented, not that anyone needs it.
This changes the bottleneck. A model can build what it is pointed at, and visible implementation can be copied. The scarce responsibility is choosing the target. Hall argues that this was always part of the job; abundant implementation merely removes enough mechanical work to make weak problem selection impossible to ignore.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Specific experience beats generic taste
Hall begins problem selection with direct need: build something you or people close to you actually require, especially before a market is legible enough for surveys. Her example is the initially awkward idea of livestreaming life through a head-mounted camera, which she connects to Twitch. But she immediately limits the lesson. Many similarly unusual ideas failed, so weird specificity can reveal a signal without proving that the business will work.
The easy consolation is that human taste will remain defensible. Hall rejects that broad claim by defining taste as preference under feedback. If a system sees enough examples labeled better or worse, it can learn to imitate the preference. Aesthetic or editorial judgment does not become untrainable merely because its grader is a person rather than a compiler.
She locates harder-to-copy judgment in two narrower places:
- The future without examples: An event that has not happened has no direct historical record from which to learn.
- A relationship the model cannot observe: A particular customer’s need may depend on the present situation and a shared history, not merely on everything written about that customer.
Both advantages come from access to relevant experience. They are narrower than a permanent human monopoly on “good taste.”
Richard Hamming’s idea of an “attack” gives Hall a way to turn proximity into action. As she presents it, a consequential problem becomes practically important when someone has a reasonable way to work on it. Time travel would matter enormously, but consequence alone does not supply an attack. Hamming advised keeping 10 or 20 important problems in mind so that a new tool, angle, or observation could make one tractable.
AI changes the scarce half of that equation. If many people now have an attack on many problems, the valuable judgment is deciding which problem deserves one. Hall grounds that choice in closeness to a domain: accumulated mistakes, oddly specific experience, and care that exceeds what seems reasonable. Being first matters less than understanding the gap between what existing data describes and what should exist for people living with the problem.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Supply the substance, then automate the expression
Choosing and building the right product completes only half the job. Its meaning still has to move from the builder’s head into the intended customer’s understanding. Hall turns to content because that transfer increasingly passes through AI-assisted posts, launch materials, sales collateral, and other generated forms.
Her diagnosis of repetitive feeds follows the same convergence mechanism as repetitive products. Models have learned familiar formats that attract clicks, and a vague request for virality leaves the system to fill missing substance with those formats. Hall says readers quickly recognize the pattern and skip material that could have been generated from a one-line prompt. The precise claim that recognition takes half a second is presented rhetorically rather than as a measured study.
Two AI workflows can look identical from outside while producing different value:
- Average in, average out: A generic prompt yields another polished but interchangeable artifact. Hall describes this as efficiently automating your own irrelevance.
- Firsthand signal in, convergence work out: Supply the specific point of view or story you actually witnessed, then delegate drafting, formatting, optimization, and cleanup.
The dividing line is what the human contributes. AI can shape the expression around a core it was given; it cannot recover firsthand substance that never entered the process.
Even a strong core can distort during transmission. Hall identifies three failure points—at the source, across an organization, and during machine repackaging. They require different fixes because each loses meaning for a different reason.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Three ways the signal distorts
Source distortion begins when founders know the product so well that they compress its explanation past legibility. They assume context and lead with architecture or technical ingenuity, leaving listeners unable to tell why the work matters. Hall describes an unnamed YC company whose pitch had deleted the customer pain. Rewriting the opening around something users hated—and that the product removed—was followed by pilot conversions in later conversations that week. No counts or controlled comparison are provided, so the example demonstrates a plausible communication fix rather than a measured conversion effect.
Organization distortion appears as intent crosses management, legal, sales, and other departments. Hall does not blame incompetence. A founder may protect unusual details because the outcome is personally theirs, while someone three layers away reasonably optimizes for the assigned specification and closing a Jira ticket. Giving both people the same AI does not give them the same investment in the outcome. A long delegation chain can therefore amplify convergence and repeatedly round the work toward the mean.
Adding more process can worsen that problem by introducing more handoffs and delay. Hall proposes a thin go-to-market function whose narrow job is to reconnect work to the intended outcome and validate that the original meaning survives. The proposal is deliberately smaller than a new bureaucracy: protect the signal at handoffs rather than adding general oversight everywhere.
Machine distortion happens after a careful launch is remixed into tweets, sales decks, and partner one-pagers. Hall’s hypothetical example begins with a narrow evaluation scoring 94%. Repetition gradually turns the bounded result into a customer promise. The number remains while its conditions disappear, changing a measurement into an expectation the evidence did not support.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Weld the limit to the claim
Hall makes the remedy concrete with a hypothetical monitoring tool in a category containing 12 alternatives. Its signal is not generic “AI-native observability.” It knows what not to wake an operator for: it stays quiet on noise so that a nighttime page earns belief. The product’s distinguishing value is therefore trust created through selective silence.
The first design rule is to state the promise with its limit attached. Hall’s illustrative wording says the tool stays quiet on anything it cannot tie to real user impact and shows everything it silenced so an operator can overrule it. That sentence exposes both the decision rule and human control. The limitation is part of the value proposition, not a disclaimer appended afterward.
The product and launch materials must preserve the same coupling. Every suppressed alert remains visible. An illustrative claim of 90% fewer pages appears beside the condition that every silence is visible and reversible. Neither percentage in this example is reported product performance. Hall uses it to show why impressive numbers and their limiting conditions must travel together: automated shortening will otherwise retain the number and discard the part that keeps it honest.
Before scaling distribution, test the received meaning. Give the README to an SRE unfamiliar with the project and ask them to explain the product back. The difference between their account and the intended account is the distortion about to be amplified. Hall says much of this checking and surveying can be automated, but she does not specify an implementation, metric, or acceptance threshold; the essential step is still comparing intended meaning with observed understanding.
A lightweight signal-preservation loop
Read the diagram as text
- Specific product signal. The monitoring tool earns trust by staying quiet on noise.
- Promise with its limit. Tie silence to real user impact and preserve operator override.
- Visible, reversible behavior. Every suppressed alert remains inspectable and reversible.
- AI-assisted repackaging. Launch material may be shortened into other formats.
- Unfamiliar-reader check. An SRE explains the product back before distribution scales.
- Repair the distortion. Revise the message where received meaning differs from intent.
- Specific product signal → Promise with its limit: define.
- Promise with its limit → Visible, reversible behavior: encode in behavior.
- Visible, reversible behavior → AI-assisted repackaging: support each format.
- AI-assisted repackaging → Unfamiliar-reader check: test what survived.
- Unfamiliar-reader check → Repair the distortion: compare with intent.
- Repair the distortion → Promise with its limit: revise before scaling.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Trust has no complete grader
Hall finally asks what building and faithful shipping are for: getting a person—or increasingly an agent—to choose and rely on one product among many similar alternatives. She calls the outcome trust. Unlike a compiler result or narrow benchmark score, trust has no complete grader in her framing. It develops slowly through a relationship and consent, as with doctors who choose to open a particular tool every morning.
Failure is not neutral. Generic work consumes tokens, infrastructure, and skilled labor, then may train customers to ignore the producer. A customer can inspect a product once and never return; every interchangeable post can teach a reader that the company’s name is not worth another click. Cheap generation can therefore spend real money while making the product harder to choose.
Hall’s conclusion moves value away from raw speed and toward direction: deciding what deserves to be built, what deserves to be said, and what deserves trust. Builders do not have to be first, but they need a real problem and enough conviction to carry its specific meaning to the right people. Her operating rule is demanding but practical: define the signal yourself, protect it from distortion, and use AI aggressively for the surrounding execution.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Resources
From the talk
The official recording page includes the video, timestamped transcript, chapters, and a companion reading version.
- Lena Hall on XReference
Hall’s supplied public speaker profile and contact destination.
Further reading
- AI EngineerReference
Conference information and a catalog of AI engineering talks and workshops.
Related talks
- The engineer of the future is the person who is able to choose what is worth doing — Addy Osmani
Extends Hall’s argument that abundant agent execution increases the value of human judgment, production accountability, and deciding what should ship.
- Ending AI Slop
Offers a complementary mechanism for turning subjective, context-specific creative qualities into measurable signals while retaining expert judgment.
- Your AI Product Will Fail Unless You Can Explain It
Develops the customer-centered communication problem behind Hall’s source distortion: begin with concrete pain and make the resulting workflow change legible.
Full transcript
Loading transcript…