AI makes implementation fast and widely available, so choosing which problem deserves attention becomes the scarce work.
2
Taste that can be learned from feedback is reproducible; judgment about future events and specific customer relationships is harder for models to copy.
3
A signal layer keeps a product's distinctive claim intact from the builder's intent through the customer's understanding, while AI handles drafting, formatting, and other repeatable work.
Summary
Lena Hall argues that AI has made average software and content cheap to produce, because competitors can point similarly capable systems at the same questions and receive similar answers. The advantage therefore moves to deciding what to build and how to preserve the reason it should be chosen. She separates this into knowing the signal, then shipping it without distortion. Judgment survives when it concerns something that has not happened yet or depends on a relationship and context a model cannot observe. Hall describes source distortion in startups, organization distortion in larger companies, and machine distortion when a careful claim is remixed into shorter marketing assets. Her proposed signal layer carries the original intent through those handoffs. It includes stating a product's limit, keeping evidence of that limit visible, and testing what an unfamiliar reader understands. Trust is the outcome with no benchmark or reward signal, and generic output can cost money while making a product harder to choose.
Hall starts with abundance: she solved a production incident from a trail near a waterfall, while a friend ran 18 agents from a bicycle. The same speed reaches competitors, so a feature can be copied that afternoon. She says, "the cost of the average just went to zero and so did its value." Models answer questions such as what users want or what to build from common knowledge. Their answers are competent and confident, but they are also likely to resemble what a competitor receives. A model can execute a direction. It cannot decide what the direction should be.
A signal has to be defined and carried without distortion
Hall calls the work of being chosen the signal layer. It has two parts. The build side defines what is being made, why it exists, and why it is not the average version. The ship side carries that meaning through content and go-to-market work so customers understand the product as its builders intended. She has worked as an engineer, founder, and go-to-market operator, and says the same problem appeared in each role: the signal did not always survive the trip from the product to the customer.
Coding agents improved sharply on a standard software benchmark, while actual shipping moved much less. Hall explains that compilers and test suites provide free graders, so anything that can measure itself can train a model against its score. Coding was an early target because it is unusually checkable. This makes implementation converge for everyone, while the most buildable thing is rarely the most valuable thing. The model will build what it is pointed at, but it offers little help with choosing the point.
Useful judgment concerns the future and a real relationship
Hall rejects broad taste as a lasting advantage. She defines taste as preference under feedback, and says that systems can learn preferences when examples have better or worse signals attached. What resists training is narrower: judgment about events that have not happened yet, where no data exists, and judgment inside a relationship the model cannot observe. A model may have read everything written about a customer, but it has never met that customer. Specific experience and close contact with a problem can reveal the gap between what the model learned and what should exist.
The rare decision is which problem deserves an attack
Using Richard Hamming's account of important problems, Hall says a problem matters when someone has a reasonable attack on it. Time travel may be consequential, but it is not an important problem in this sense because nobody has an attack on it. Hamming advised keeping many ideas in mind until a new tool or angle creates an attack. AI has now given everyone an attack on nearly everything, so the scarce judgment is choosing which problem is worth attacking. Hall locates that judgment in a person's battle scars, domain closeness, and unusually specific experience.
Hall says internet content has started to converge around familiar formats, including polished posts with predictable bullet points and takeaways. Readers can recognize model-written patterns quickly. AI has learned what gets clicks and fills unspecified parts of a request with sameness. One use is to give an average prompt and ship another indistinguishable output. The other is to supply a specific point of view and a story from a situation the author actually experienced, then let the model handle formatting, drafting, and cleanup around that material.
The first distortion Hall names is source distortion, which she associates with startups. Founders know their signal so well that they omit context their audience needs. She describes a YC company whose founders opened every pitch with architecture and clever technical details. The product was new, but the presentation became noise because it removed the customer's pain. Hall helped put the problem the product had killed at the start of the story. The same product then converted later conversations into pilots and a repeatable go-to-market system.
In larger companies, signal is distorted as it moves through management, legal, sales, and other departments. Hall says this comes from incentives and investment rather than simple incompetence. A founder personally sweats the details because the result affects them directly. Someone several layers down may complete the specified Jira task with less conviction. A long delegation chain combined with a convergence machine can automate the distinctive parts out of a company. Hall proposes a thin signal layer that reconnects work to the original outcome and checks the intent across handoffs.
Hall uses a monitoring tool to show how to weld a product's claim to its scope. The distinctive behavior might be staying quiet about alerts that cannot be tied to real user impact, so a night-time page is more trustworthy. Instead of calling it an "intelligent AI-native observability platform," she suggests stating the behavior and showing what was silenced. Suppressed alerts should remain visible, and a claim such as "90% fewer pages" should sit beside the fact that every silence is visible and reversible. An unfamiliar SRE can then repeat the product's meaning back to the team before the claim spreads.
"The most buildable thing and the most valuable thing are almost never the same thing."05:28
Who should watch
You are building a product in a category where a competitor can reproduce the visible feature quickly.
Your team uses AI to write product, sales, or marketing material, and you need to keep the original claim from being diluted.
You are deciding what to build next and have access to a specific customer problem, domain experience, or insight that common data does not capture.
