For most of the last decade, the competitive edge in software was moving fast. Shipping a working product before the competition — that was the thing that mattered. AI has made that edge obsolete.
When speed stops being the moat
Autonomous coding agents went from solving a small fraction of standard benchmark tasks two years ago to clearing the high-eighties today. Competent, fast, AI-assisted implementation is now effectively free — and like anything that becomes free, its value has collapsed to match. Every team with access to the same tools gets the same output. AI is, at its core, a convergence machine: pointed at the same goals with the same prompts, it produces the same answers, because it trains on data from the past rather than on what should exist.
The implication is uncomfortable but worth sitting with: the cost of “average” dropped to zero, and so did its value.
The rarest thing: knowing which problem to aim at
The model will build whatever it is aimed at. It offers no guidance on where to aim.
The mathematician Richard Hamming observed that great scientists tended to work on important problems — not impressive ones, but ones where they had what he called a reasonable attack. For most of history, having a reasonable attack on a hard problem was itself rare. AI just handed everyone a reasonable attack on almost everything. The scarcest input to the whole process is now the judgment about which problem is worth attacking in the first place.
Paul Graham’s heuristic for startups reads differently in this light: build something you and your close circle genuinely need, before the market has formed and surveys can surface it. That felt, lived, specific need is the only signal AI cannot synthesize from public data — because it does not exist yet as data.
What actually resists AI training falls into two categories. First: judgment about what has not happened yet, for which no training data can exist. Second: judgment embedded in a relationship the model cannot observe — what this specific customer, in this specific context, with this shared history, actually needs right now. The model has read everything ever written about the customer. It has never met them.
(Notably, broad “good taste” is not a durable differentiator here. Taste is preference developed under feedback, and that is exactly what models get better at given enough demonstrations.)
The signal layer
Once a team has a genuine signal — a weird, specific insight grounded in something real rather than synthesized — the next problem is getting that signal to the people it is meant for without losing it along the way.
Three distortion modes tend to kill signal in practice.
Source distortion is common in early-stage teams. Founders compress the original signal past legibility — they lead with architecture or technical cleverness instead of the customer pain that made the problem worth solving. The fix is to rewrite the opening around what the user actually experiences, not the parts the team finds elegant.
Organization distortion happens at scale. As intent passes through management, legal, and sales layers, each handoff regresses it toward the mean. Add AI content generation to a long delegation chain and the result is an efficient factory for averaging out the company’s own original signal. A thin “signal layer” function — one role or process whose only job is to carry original intent intact across handoffs — is the structural answer.
Machine distortion is the newest failure mode. AI remixes one careful, scoped launch claim into tweets, decks, and one-pagers, silently dropping the limiting context along the way. A 94 percent accuracy figure that was narrowly scoped to one use case becomes a broad, unqualified promise by the time it reaches the audience. The fix is to weld the promise and its scope together in a single sentence that is structurally hard to edit out — and to check what the audience actually heard before scaling distribution.
Trust as the un-gradeable asset
There is no benchmark for trust. No reward signal. No automated grader. It accumulates slowly, through relationship, with consent — and generic output actively erodes it. Producing average content is not neutral: it costs real money in tokens, infrastructure, and salaries, while training an audience to ignore the name attached to it.
The two modes of AI-assisted content creation look identical from the outside but produce opposite outcomes. One starts with an average prompt, gets average output, and adds one more indistinguishable voice to the noise. The other brings something the model cannot have — a specific point of view, a story from inside the room — and lets the model handle everything else.
The practice worth building now is not “use AI faster.” It is: find the signal, keep it intact, and trust that specificity compounds.
Text summarized and optimized using Anthropic’s models and reviewed by a human.