Teaching a network to move its eyes

The email confirms it: their pre-registration passed, and the China Innovation & Entrepreneurship competition wants them in the Paris division pitch on October 28, whose winners fly to Shenzhen for the real one ▸ 26:17. Global prizes run to $136,000 ▸ 29:01, but the prize he actually names is geographic: finalists get help establishing in Shenzhen, “una puerta real a entrar literalmente a China” ▸ 30:42, the country REAPRA can’t reach.

The entry ticket is the research that’s been frozen since early this year: Query Network’s eyes. Thesis one: current networks swallow whole images at once, while humans see partially, glance here, glance there, close your eyes and reconstruct, “terminamos de llenar los huecos” ▸ 31:56. Thesis two: train the gaze itself using the network’s own error signal as a curiosity drive ▸ 32:30, because novelty is how animals allocate attention: it’s why TikTok works, why the baby drops the familiar teddy and grabs the new duck ▸ 34:15, why you stop improving at bicycle once the error signal goes quiet ▸ 33:04. It’s the zebrafish objection turned constructive: he’s not just saying today’s architectures aren’t brains, he has a sketch of what’s missing.

la curiosidad es una regla de entrenamiento; los ojos son su actuador →

The strategy stays disciplined even while dreaming: he hasn’t touched architectures in months and won’t yet, “el camino más fuerte es primero crear esa empresa, la vaquita de leche… y ya dedicar el resto del tiempo a la investigación” ▸ 35:14, exactly why the research stopped the first time: no funding, no food. The competition costs nothing but a pitch Julia will help polish; if the long shot lands, it funds the very thing it describes. Two marshmallows, properly stacked…