The reader is the bottleneck

Four items in today's corpus, none of them about each other. A systems engineer's blog post about LLM prose. A Princeton job-market paper on short-form video. A book review about the end of reading. A physics preprint that models chatbots as a pathogen. Read together, they are one story: the cost of producing text and video has gone to zero, and the constraint that is left is the reader.

That is the claim. Not that attention is scarce, which Herbert Simon said in 1971 and Neil Postman turned into a book in 1985. The new thing is that the reader's defenses are now being measured, and the numbers say the defense is real, fast, and depends on design choices the producer makes.

Start with the withdrawal side. In The revolt of the reader, Bryan Cantrill cites Cynthia Dunlop's survey of 668 developers: 78% stop reading the moment they detect an LLM, 71% avoid the author afterward, and 98% prefer an author's own imperfect draft to a polished machine one. The tell is not grammar. It is structure, and the reaction is not annoyance but exit.

Now the consumption side. Renjie Bao's paper, flagged on Marginal Revolution, uses a short-drama platform's top-up menu to infer how much people planned to watch. Paying users watched 82.1% more than they intended. The structural estimate that matters is the temptation horizon: 11.2 minutes. That is short against a full drama and long against a one-minute episode, which is the whole mechanism. Each tiny unit renews a temptation that outlasts it. Bao's counterfactuals find that larger decision units, default limits, and breaks all raise long-run welfare. The lever is unit size, not willpower.

Put those two next to each other. One reader ejects in a sentence. Another overshoots by four-fifths and cannot stop. Same organ, opposite failure, and both are set by what the producer chose: prose the writer did not labor over, episodes cut to sixty seconds. James Marriott's The New Dark Ages, reviewed by Ed West, supplies the macro version: 260 TikToks in half an hour, Norwegian conscripts' test scores falling region by region as cable arrived, a college graduate born after 1969 reading less than a modestly schooled person born before 1950.

The honest objection is that Marriott's story is the old story, and it ends in despair. Television did this, the phone finished it, and the reader is a casualty, not a combatant. Postman would agree. If that is right, there is nothing to engineer.

Concede the trend. Reject the fatalism, because two of today's four items say the reader fights back and say how. Cantrill's 78% is an immune response, and he draws the right analogy: email spam did not end email. Filters got good enough that the economics of spam broke, and then being classified as spam became ruinous for legitimate senders. The Solé, Krakauer, and Levin preprint, Large-Language Models as a Cognitive Virus, reaches the same place from population dynamics. Their model produces tipping points into persistent dependence, but it also names the conditions for immunization: reduce transmission, make reversal easy. Those are Bao's levers with different vocabulary. Bigger units and default limits reduce transmission. Breaks make reversal easy.

The lever is unit size, not willpower.

For anyone who ships software or writes for a living, the implication is concrete. The reader's filter is already running, it costs you the reader forever, and it keys on the effort you skipped. The products that survive the next few years will be the ones that pick a unit size the reader can plan around and stop spending the reader's eleven minutes for them.

Supply is solved. The reader is the bottleneck. The engineering left to do is on their side of the screen.