What matters in AI world

When building gets cheap and models write most of the code, the scarce inputs move up the stack. This note tracks what the AI-era market actually prices: which human traits hold value when execution is commoditized, and why.

Taste is one of the highest-value traits

The anchor exchange (samwhoo asks, Geoffrey Huntley answers; both captured):

“How do you, you personally, define ‘taste’?” — Sam Rose

“Someone who gives a lot of shits and won’t sleep if the thing they’re using feels like crap. They just keep refining because they give a lot of shits.” — Geoffrey Huntley

Taste, operationalized: not an aesthetic sense but sustained, uncompensated care about quality of the thing itself. The mechanism is refusal to ship “fine” — the same trait Huntley elsewhere frames as the mentor-transmissible craft knowledge (“a lot of information and experience about software failures and how software should be designed, colloquially known as taste”) that the incoming generation will need taught (ghuntley.com/replaced, captured). Taste is the judgment layer that decides what good looks like when the generator can produce any amount of output.

Decode: why taste appreciates as models improve

• Generation is cheap, selection is not. When a prompt yields ten working implementations, the binding constraint is telling which one is right — vibecoding “needs taste (can you tell it’s good without opening it?)” (Vibecoding vs AI engineering).
• The supply is structurally limited. The “chronically online operator every AI startup wants” — taste + technical understanding + shipping — is scarce not as a skills problem but an incentives problem: the people who match “can usually generate more money, ownership, and creative control on their own than most startups are willing or able to offer” (adelwu_ wave, recap captured).
• It compounds with AI, not against it. Taste + agents = leverage; the tasteless + agents = faster production of mediocrity. The judge model over the worker model (Agent skills, AI engineering) exists precisely because raw generation without a quality filter is worthless.

The other scarce traits (roster to grow)

• Judgment about failure — which failure modes matter and what contains them; the AI-engineering counterpart to taste (Vibecoding vs AI engineering).
• Distribution — “building is largely solved, so everyone is focused on marketing” (adelwu_ wave recap) — being chronically online IS a distribution asset the market pays for.
• Onboarding-proof context — the war stories and “we tried that, it didn’t work” knowledge agents can’t accumulate because “it is on its first day, permanently” (Vectorize onboarding piece, captured); see Agent knowledge systems.
• Motion — execution bias beats raw intelligence when generation is cheap: “a retard in motion is better than a genius at rest… the problem is that nowadays there are too many retards at rest” (IterIntellectus, quoted by Daniel Lockyer, captured). In the agent era the gap widens: agents multiply whatever motion you already have, so an average intelligence shipping beats a genius still deliberating. The dark mirror of taste — motion without taste ships polished mediocrity, so the pairing (taste + motion) is the actual scarce combination.
• Focus / “Not Yet” — the governor on motion. When software generation is cheap “you can work on anything and everything feels worth the effort. That’s the trap” — WIP still kills progress, so the discipline is relentlessly protecting main priorities, a small daily budget for ad-hoc fixes, triage for the rest (swizec, captured; half his sprint’s work arrives through triage). Independently confirms the problem-first bullet: the constraint is triage capacity, not generation.
• Review as the bottleneck — “Nobody cares how fast or how much code you produce. What did you ship to production that users can see?” Review (code + product) is the highest-leverage use of senior talent; agentic review “lacks judgement” — swizec’s team rejects half of BugBot’s comments as irrelevant, and bots can’t tell a bet-the-company feature from a throwaway internal tool. The operational form of taste: selection over generation.
• Ownership of outcomes — swizec’s org runs on “you’re signing that thing with your phone number”: when AI writes 97% of the code, what’s employed is not the typing but the accountability — you ship it, you run it, you make it effective. There are no “translate english into code” developers left; “we have AI for that.”
• Problem-first inversion — agents “will happily generate features, roadmaps, and solutions on demand”; without a problem statement the AI “autocompletes a roadmap for you — potentially a very polished, very wrong one” (nurijanian, captured). Framing the problem is the human input the generator can’t supply itself; the constraint is “triage capacity, not idea generation.”
• A reservoir of big problems — the demand-side twin of problem-first. Karpathy’s capability funnel (captured, 9k❤/379k views): impact “scales with ambition, problem size, and horizon — a question with a paragraph answer barely stresses the system. You need a reservoir of big, difficult problems that you really care about.” The funnel’s tiers (~6B barely touched LLMs → ~1-2B casual free-tier → ~20M professional math&code seeing agents do weeks-of-work per prompt → ~5k internal frontier access, “swarms of thousands of agents… minting zero days, discovering new science”) is the market the other traits operate in: the ~0.2% sliver experiencing “the vertigo of the curve going vertical” is where what-matters is being repriced. Also the strongest grounding for this note’s opener — when capability is jagged and access-tiered, being in the professional tier with taste/focus/ownership is the compounding position.
• Quality as the market’s new subsidy — Collison’s economics, endorsed by Tobi Lütke (both captured, 1.7k❤/148k views): companies are “the original superintelligence” that profit from consumer bounded rationality — and personal agents end that. “Personal agents will, I think, act as a kind of structural subsidy for product quality”: agents hunt coupons, cancel forgotten subscriptions, research purchases like a 10-hour audit. His close: the market is “a massively parallelized, decentralized, and highly imperfect value attribution machine” and agents change “the nature of the backpropagation such that the rewards for better products increase… companies that make things that they’re proud of will feel that the universe is a little more partisan in their favor.” The macro version of taste appreciating: when every buyer has an agent, distribution’s edge shrinks and quality’s edge grows — the economy-side complement to this note’s micro claims.

Working mode shift: pirates and architects

The primary source: Dan Shipper (Every; 20-min talk, transcript captured; surfaced by Lenny’s “Two-slice teams” post and Cory House’s summary, both captured). His framing, verbatim from the transcript:

“the composition that works for us… is I call them pirates and architects. So, one person on the team is the pirate. They are slop cannons who are absolutely obsessed with finding value… doing lots and lots of stuff, and throwing it away, and it’s all messy, but I’m going to find something really interesting. And then architects are people who are looking at a messy system, like maybe something that I built that’s completely vibe coded, and trying to help shape it into something that’s valuable and beautiful and extensible. And pairing those types of people together is very, very powerful in AI.”

The 2023 lineage: Rubin had the pairing first. “The Experimenter and the Finisher” (The Creative Act, Jan 2023 — pages transcribed, source note planted): “many artists lean toward one of two categories: Experimenters or Finishers… Each might find it helpful to borrow from the other” — Experimenter = pirate (dreaming/play, “finding it more difficult to complete and release their work”), Finisher = architect (“move quickly to the end point with immediate clarity”). His completion mechanics are the pirate’s survival kit: narrow ten songs to two, “going from two to three is easier than going from zero to two,” skip stuck items — “the knowledge we gain from finishing the other pieces becomes a key to overcoming earlier obstacles.”

Team-size claim (Cory House’s distillation, captured): “AI makes the optimal team size 1-2 people” — one pirate + one architect (“Two-slice teams,” Lenny). The talk’s deeper structure: run the product org like a research lab — separate exploration from execution, expect the lab side to dispose ~90% of what it makes and the product side to adopt ~10%; “a labs team of one” is affordable because agents give everyone superpowers; identify the org’s early adopters (“we’re living in the future already”) and harness them without distracting everybody else. The transition ritual: when an experiment shows real org value, “this is the time where I would then bring in an architect” to make the messy thing really great.

Corroborating captures:

• swizec (When code is cheap, judgement becomes the job) — why the architect role is newly viable: “most architects burn out” because nobody reads the docs — but agents read guidance every time; one .md update redirects every future PR, repeated review comments become deterministic linters. “Build the system that generates the software.” The economics now have a dollar figure: better file structure cut token cost by 83% (Martin Fowler experiment) — “agents amplify the code quality you’ve got,” so in a company of vibe coders someone must look at the code and enforce quality. His org also names the failure mode of unbalanced pirate energy: product debt — “a big danger when everyone can ship any idea,” every feature a new learning curve, 52 shipped features/year = 52 things to maintain and users who “complain that your software looks clunky but they’re not sure why.”
• tometo (tometo (OSS)) — the same split as runner config: “put a stronger model on the judgment, and a fast one on the edit.”
• The motion bullet above is the trait-level version: pirates are motion; architects are where taste compounds.
• Gergely Orosz’s LDX3 field report (keynote transcript, captured; full report in Software engineering in AI era) — the industry data behind the pairing: Anthropic’s head of eng on why projects cap at “one or two engineers” (“you’re already fighting against your agents, which are stepping on each other’s toes”) — the two-slice claim from inside a frontier lab; ~100% AI-generated code at startups tracking it; 5–10 parallel agents per engineer as the new normal (Linear: “one mouse, one keyboard, one screen used to cap us”); specializations collapsing (OpenAI no longer hires iOS/Android specialists — “anyone can do the work at a good enough level,” and good enough is exactly where taste decides); and Matt Pocock finding that saying the old words (“deep modules,” “tracer bullets”) makes the agent write better code — craft vocabulary as prompting.

The industry state check (Orosz, LDX3 2026): what holds, what breaks

The same talk’s held/broke ledger is the stress test for this note’s roster — full field report in Software engineering in AI era; the trait-relevant entries:

What holds (hype corrections):

• Teams survive. Anthropic’s Caitlin Lesser: “two humans + agents is NOT a team” — teams still own software and on-call; two-pizza shape intact. The pirates-and-architects pair lives inside a team structure, not instead of one.
• The test ratio is unchanged. Bun’s Jarred Sumner: AI writes ~all the code AND ~all the tests, and the time split “has not changed” — verification did not get cheaper with generation.
• Non-engineers are NOT shipping to prod. “They don’t even know what a pull request is” — at every company checked. The chronically-online-operator wave is about judgment + distribution, not literal shipping by anyone.

What breaks (failure modes of the roster’s traits):

• Zombie code reviews — “everyone is playing the theater of doing reviews… if your team is still doing code reviews, you’re probably not doing them”; Ramp phases them out for non-critical code. The tension with swizec’s “review is the highest-leverage activity”: both are true — review that scales is rules + linters + selective human judgment, not per-PR heroics.
• Focus breaks at industry scale. Linear’s Dima: “there’s always another agent that waits your response” — parallel became the baseline expectation, and he ends up with more work than pre-AI. The “Not Yet” discipline is not optional; it is the counterweight to the motion agents create.
• Engineering leadership attrition — founders demanding 2x with half the team, smaller teams needing fewer leaders, fractional-CTO escape hatch. The market can’t hire what it wants (below) partly because its leaders are exiting.

The thesis in one quote — Titus Winters (Software Engineering at Google), verbatim from the talk: “You need intelligence — you know how to do it. Wisdom — you know what to do. Charisma — you convince others to do it. When intelligence becomes commonplace… wisdom and charisma become much more important.” This note’s entire roster is wisdom-and-charisma territory: taste, judgment, focus, review, ownership, distribution, problem-framing.

Why the market can’t hire what it wants (adelwu_ wave, recap captured)

The “chronically online operator” wave (Anna’s recap of adelwu_’s 1.1M-view post, captured) maps the structural reasons the scarce traits stay unhired — the points that matter beyond taste itself:

• The archetype is a founder, not an employee. The skill combination (taste + technical understanding + shipping) is “closer to a founder or independent creator than an employee” — the people who have it already make money and own their output, so a salary is a downgrade. And self-labeling doesn’t create the trait: “calling yourself this person doesn’t make you one.”
• Platform skills don’t transfer. Excellence on X doesn’t make you excellent on YouTube or LinkedIn, even when the job description demands all three — distribution skill is platform-shaped, not generic.
• Hiring for taste means surrendering decisions. “Hiring someone for their taste means letting them make decisions you wouldn’t make yourself” — startups want the expertise, then micromanage it, which destroys exactly what they paid for.
• HR is the filter that rejects the archetype. Resumé-screening processes are built for legible, interchangeable skills; the combination the market wants is illegible by construction. Title inflation (“narrative fairy” roles) is the symptom — the people who fit “aren’t applying.”
• The scarcity is incentives and culture, not skills. The emerging consensus: the role is real, the demand is real, the scarcity is real — but “the companies that most need the skillset are often the ones least equipped to hire, retain, or fully use it.”
• The operator’s side: attract, don’t chase. adelwu_’s own advice to the “this is me, why am I not getting hired” crowd is operational: post multiple times a day, keep improving, interact in person and online, build real relationships — the same distribution discipline startups fail to buy.

Kin

• Vibecoding vs AI engineering — taste vs failure-judgment as the two lanes’ core skills
• Agent knowledge systems — where team knowledge lives when agents forget everything
• AI native teams — the org-design side of who holds these traits
• Coding with agents 2026-10-07 — hands-on state of the craft
• Software engineering in AI era — the profession-level umbrella; Orosz LDX3 field report, swizec, ghuntley arc