An answer to the two sigma problem
In Nuremberg in 1658, Johann Amos Comenius published Orbis Sensualium Pictus, a book in which every Latin word appeared beside a woodcut of the thing it named. Children until then had been taught by recitation and by the rod. Comenius proposed showing a child the world and giving it names, in an order a child could follow. The book remained in use for the better part of two centuries.
What Comenius could not supply was attention. In 1984 Benjamin Bloom set out the arithmetic plainly. A child tutored one to one performs roughly two standard deviations above a child in a conventional classroom, far enough to carry an average student past the ninety-eighth percentile. Bloom put the question to his field as an open challenge and asked for a group method that could match it. Forty years on, the question is still open.
The reason is not mysterious. There has never been one adult for every child. The cohort, the bell, and the single pace set for thirty all descend from that shortage. And because nobody was free to sit and listen to a child work through a word, we taught a method that spared us the listening. Children were told to look at the picture, weigh the shape of the word, and settle on whatever would make sense in the sentence. This was called reading instruction in most American classrooms for forty years.
English is a code and a child has to be taught to break it. The sequence is known and it is narrow. A child learns which sounds the letters make, practises on text assembled only from the correspondences already covered, and meets a word often enough that its spelling, its sound, and its sense fuse into one act of recognition. After that the word is simply known on sight. The research describing all of this is thirty years old and no longer seriously contested.
Which returns the problem to attention. A tutor's real work is auditory. A tutor hears a child say wabbit and knows to let it stand, hears the stall before a digraph and knows precisely which lesson to return to, hears the difference between a child who cannot decode a word and a child who has simply had enough for one afternoon. Speech recognition trained on adult voices has been poor at all of this. Children speak higher, faster, and less consistently, with substitutions that are developmentally correct at five and would be defects at thirty. A machine could read aloud to a child long before it could listen to one.
That has changed in the last two years. A model can now follow a five year old through a sentence at the level of the individual sound, hold a record of every word she has attempted since the first session, and assemble tomorrow's story out of what she missed on Tuesday and what she has loved all month. It does not tire, and it does not have thirty other children.
Our founding team is from Harvard, with research in artificial intelligence and education at the Berkman Klein Center and the Kempner Institute. Husam has worked on education policy, Jesus on learning platforms used by over a million students. We have both taught children at every age this tutor will serve. We believe a tutor that cannot listen is a worksheet with a voice.
Poma is building the tutor that can. For the first time there can be one for every child.