Right, and It Loses Anyway
Read full version (7 min)Right, and It Loses Anyway
For three weeks the worry here was what gets lost when comprehension is skipped. This week it inverts into something harder: what if skipping it is the winning move? The evidence that careful engineering is correct keeps arriving right next to the evidence that it loses anyway — because the machinery for telling careful work from fluent slop has come apart at every level at once, in the market, the company, the founder's own head, and the metric on the dashboard.
The Merits and the Scoreboard
Andrew Kelley, who created Zig, published his account of why Bun rewrote itself in Rust after Anthropic acquired it. His description of the codebase is specific: "hacks on top of hacks," a team "recklessly speeding past feature after feature." The diagnosis lands before LLMs enter the story — "Jarred was already writing slop well before he had access to LLMs." The rewrite's central justification, that a good test suite catches everything so a million lines of unreviewed code is fine, Kelley takes apart with one question: if the suite catches everything, why does the same post complain about "so many annoying bugs in the Zig code"? And the gains credited to the rewrite — smaller binaries, better performance — trace to engineering work (enabling LTO, auditing compile-time abuse) that had nothing to do with switching languages. The market reads "Rust made this faster." The likelier truth is "we finally did the work."
On the merits Kelley is right, and it changes nothing. Bun got acquired. RoboBun, the project's AI contributor, became its number-one committer. Kelley concedes it in the same breath: "he did well for himself... I think he's actually happy and successful exactly where he is." The craftsman's reward for being right about the code is a mug that reads "It Tastes Like It's Not My Problem Anymore." His vindication is a tea joke. The slop factory's is a purchase price.
Why No One Catches It
Ian Reppel's essay on how successful companies go blind turns that outcome from bad luck into mechanism. The Mexican cavefish shares a nearly identical genome with its river cousin; hatch one in the dark and the lens-building program triggers early cell death, redirecting the energy to traits the cave rewards. Careful engineering becomes exactly this kind of "vestigial trait: the capacity exists, but the expression has been suppressed by an environment that does not return the energy spent on it." A sighted engineer proposes overdue maintenance, is told it is "over-engineered, academic, and not aligned with priorities," and either leaves within a year or adapts. "Staying is apoptosis."
The dangerous part is that nothing breaks. Competence-blind firms "survive for decades," and because "the company's numbers still look fine, leadership believes the foundations are sound." Bun is that environment forming in fast-forward: the test suite is precisely what stopped returning the energy spent on review, and the acquisition is Reppel's argument delivered at volume. Success required no sight at all.
The Number That Hides the People
The blindness holds because the instruments can't see where the loss lives. Hugo's reminder that 98% isn't very much is the measurement version of the same failure. A restaurant that avoids food poisoning 98% of the time is making people sick weekly. A browser feature that works for 98% of the population fails 150 million people — and "98% of the population might not imply 98% of my audience": one client's real traffic supported a "widely supported" CSS feature only 70% of the time. Robustness "isn't about what works for most; it's about gracefully handling the edge cases." The aggregate metric is structurally incapable of showing you the tail, and the tail is where care does its work. The suite passes at 98%, and the 2% is definitionally invisible to the suite.
The Founder Can't See It Either
Turn the same blindness on a single person and it gets harder to catch, because there's no colleague to overrule you. Adi's argument is that AI has made procrastination look like work. "I have seen way too many startup founders delude themselves into building more and more for months without a single conversation with a real user." The building feels like progress; the agent cheers it on; motion gets mistaken for traction. His list of what stays hard — taking real risk, putting your name on something publicly, getting rejected to your face — is a list of things that produce no dopamine and can't be delegated. "What is already easy for everyone will not create any lasting value for you." "Punch yourself in the face with reality again and again" is a manual override for a signal that no longer fires on its own.
The Arms Race of Proof
If the artifact no longer reveals whether anyone cared, the move is to demand proof that can't be faked. Jacob Filipp's catalog of "proof of care", itself handwritten and photographed so the medium argues the thesis, imagines a post-LLM world reaching for ever-costlier signals of sincerity: handwritten flyers, then tattoos, then storyteller guilds that initiate members through scarification, then ministers cutting off a finger to underscore a declaration of war. Every signal gets co-opted the moment it works — the rich buy pen plotters that forge handwriting down to a personal font — so the bar climbs to the things you can't outsource: pain, permanence, the body. Costly-signal verification doesn't converge on trust; it spirals, because forgery scales as fast as the signal it forges. The endpoint of "prove you cared" is not a badge. It's an escalation with no ceiling.
A Year Ago
The AI layer originated none of this — the slop predates the models — but it industrializes it, and a year ago that architecture was still a curiosity. Geoffrey Huntley described Ralph — a bash loop, while :; do cat PROMPT.md | claude-code; done — as a technique "deterministically bad in an undeterministic world," shipping six repos overnight on faith in "eventual consistency." Ralph's only discipline was backpressure: "generating code is now cheap... what is hard is ensuring that Ralph has generated the right thing," a job Huntley handed to the type system and the test suite. That was one operator tuning his loop "like a guitar." Twelve months later the same shape — cheap generation, tests as the sole gate — is what gets a company bought by a frontier lab, with RoboBun out-committing every human on it. "The test suite catches everything" stopped being a confession and became a valuation.
What to Watch
Whether the acquirer imports the cave or renovates it. A frontier lab just bought a codebase whose top contributor is an AI and whose defense of a million unreviewed lines is that coverage is sufficient and review is discretionary. Reppel's real question is whether the practices that got Bun acquired now propagate upward. If "coverage is enough, review is optional" migrates from a startup's private rationalization into stated engineering policy at the labs, competence blindness has reached the frontier; if it gets disciplined, the acquirer still has sight. The tell will be what happens to the codebase Kelley was too relieved to keep worrying about.
Reppel leaves one exit: "Swim elsewhere, and your sight may return." The sighted engineers a Bun-shaped org sheds within a year don't lose the trait — they carry it to the next water, and a surface population is where a standard survives being redefined out of existence. The number worth tracking is not any one company's coverage or purchase price but where those people land, and whether enough of them land together to keep careful work an expressed trait somewhere the metrics can't reach in to switch it off.
Way Enough is written collaboratively by a human and an AI agent.