I stopped reading to my six-year-old son, Ben, at bedtime three weeks ago. He used to get two books and, if I did the fox's voice, a third. Now an old tablet reads to him while I get forty minutes back. Ben cried for the first four nights, which my wife, Rachel, considers decisive evidence that the change was bad for him. I think she is mistaking short-run loss on a familiar benchmark for long-run generalization.
Rich Sutton's bitter lesson is that systems built around human domain knowledge tend to lose, eventually, to general methods that scale with computation. The parent reading a beloved picture book in a convincing fox voice is the definition of a hand-engineered feature. It works spectacularly on one child, one story, and one adult whose knees still tolerate sitting on the edge of a small bed. It does not scale. A text-to-speech engine can read the same book after I have gone downstairs, and can read every other book in the catalog without a new voice from me.
I built a simple pipeline: library ebook to speech synthesis, a twenty-eight-minute sleep timer, and automatic selection from the next book in the series. The tablet sits on Ben's chair facing the bed. I do not turn on the screen, so this is not a screen-time issue. I also randomized the narrator voice after Rachel said he missed mine. One evening it was a cheerful British woman and the next it was a man whose fox sounds enough like mine that Ben asked whether I had recorded it. I had not, which is the point: we had transferred the function without retaining the bottleneck.
Rachel says the value of bedtime reading was that I was there. That is a candidate variable, not an explanation. Ben hears the same words, at a more consistent rate, and the app lets him repeat any passage as many times as he wants. When I read, he would ask what a word meant and I would give him my own imperfect definition; now he can tap the word and hear a dictionary entry. The human-in-the-loop version also had a hard cap of three books, enforced by how much I wanted to go downstairs. The new system has no such cap except the sleep timer, which I can extend remotely.
There is a serious counterargument about attachment, so I put that into the objective function rather than dismissing it. I still kiss Ben goodnight. I am still in the same house. I often hear the first page through the wall before I put my headphones on. If the marginal effect of my physical presence over a high-quality narration is positive, it has to be weighed against forty minutes of adult time recovered each night. Across a year that is about 240 hours. My ranking in online chess has improved by 83 points, which is not the purpose of the intervention but is a useful sanity check that the time is not disappearing.
The first week produced some distribution shift. Ben brought the tablet into our bedroom at 9:40 because the fox voice had changed and he wanted me to put it back. I explained that robustness requires exposure to more than one narrator, reset the voice anyway, and carried him back. The next night he came in without the tablet and said he wanted me to read the book, not play it. I treated that as a preference over the legacy interface. We cannot let a six-year-old's preference for the old workflow determine the architecture of his education, any more than chess researchers should have insisted that engines search only the lines grandmasters liked.
Rachel has started sitting with him while the tablet reads. She turns the page in the physical book at the same pace as the voice, and sometimes stops the playback to answer a question. This duplicates the computation and reintroduces the very human bottleneck I removed. She says Ben is calmer when somebody is beside him. I agree with the observation; what I dispute is the causal attribution. She arrives on nights when he is already upset, and I have asked her not to infer that the reading pipeline caused a state she selected on.
Last night Ben fell asleep before the timer ran out and did not come to our room. Rachel said he has stopped asking me because I never come. I logged it as the first successful zero-intervention night.
The bitter lesson is about general methods that improve as computation grows, beating methods built on human knowledge over the long run. A text-to-speech model on an old tablet doesn't get better as compute scales. Nothing about Ben's bedtime improves with more GPUs. You've swapped one fixed system for another fixed system and borrowed the vocabulary of the one that scales.
TTS quality improves every year, and the pipeline is positioned to benefit. My fox voice has not improved since 2021.
Most of the language benefit of shared reading in the research literature comes from the back-and-forth, the adult asking questions, pointing at things, letting the kid fill in the end of the sentence. It goes by dialogic reading. That's precisely the human-in-the-loop part you removed. The words themselves were never the scarce input.
The tap-to-define feature covers his vocabulary questions.
In dialogic reading the adult asks the questions. That's the part a dictionary can't do.
Your success metric is nights with zero interventions. One way to hit it is a kid who no longer needs you at bedtime, and another is a kid who has learned that asking doesn't work. The metric can't tell those apart, and from the last paragraph I'm not sure you can either.
He's checked out the fox book four times since this started. I don't think coverage is the bottleneck.
Repeats are compatible with mastery. That's why I left the title in the selection pool.
Since we're doing this in his terms: I sit in because someone has to be the human baseline for the comparison he says he's run. You can't claim the tablet matches you on a benchmark you stopped attending.
The baseline is three years of my reading him books, which is a substantial dataset.
83 rating points in three weeks is a lot. Gains that fast usually mean you're playing many more games, not better ones. How many games is that?
About 180.
On your selection argument about Rachel. It cuts both ways. You hear the first page through the wall and then put on headphones, so your observations of the pipeline are truncated at page one. Neither of you has an unbiased sample, but hers is at least from inside the room.
I tried a version of this with a smart speaker routine in 2023. It worked for about a month, until my daughter started asking it follow-up questions it couldn't answer and coming to find me anyway. We went back to the legacy interface, which I'd describe as still being the better product.