Adam V2 waking up

Show it something

draw one digit, large

it is waking up

Tell it what that was

it answers first, then you correct it. the difference between the two is its learning rate.

What it receives 28 x 28

Seeing layer

Idea cortex

Cortex map clusters, not single neurons

seeing layer, idea cortex, mouth-ear, fast store. a node is a cluster of cells and lights with their spikes. lines are learned weight mass and move as you teach. the fast store is written on every lesson and never read, so it hangs there unconnected, honestly.

How close its answer fizz lands to each word

Adam V2

A Digital Intelligence you teach yourself, live. It runs entire in your browser. Nothing you show it or teach it is sent anywhere.

Adam V2 runs the second DENN, a Digitally Emulated Neural Network built the way a biological brain is built: spiking neurons, local learning, no answers wired in. Where Adam V1 could only watch, V2 has an idea cortex and a voice, and it learns from you while you use it.

How a lesson works

Every lesson has two phases, and the order is the design. First it answers: the picture alone drives its idea cortex, and whatever word its own speaking weights produce is its answer, not a classifier's. Then you tell it the truth. The mismatch between what it said and what you said becomes its learning rate. Surprise makes it learn hard; being right already makes the lesson glance off.

When it is stuck, it asks

Sometimes two answers land almost exactly as close, a four and a nine with the same slant. Inside a measured margin it stops guessing and asks you outright, showing the two digits it is torn between. Within that zone the true digit really is one of its two candidates most of the time; the rate was measured on held-out digits and ships in the model file. Below a floor of real evidence it never quizzes you, it just says it does not know.

It listens harder when you insist

Your lessons carry more weight than its schooling, and if you correct the same digit again while it is still getting it wrong, the next lesson leans in harder, up to a stated cap. An insisted lesson is also pressed into its long memory, the slow copy that normally protects old knowledge from new teaching, so what you hammered in does not wash back out. When it finally says your digit on its own, it tells you, and the extra weight resets. Replayed memories never get this boost; they are the counterweight that keeps your enthusiasm from erasing what it already knows.

Each Adam is its own Adam

The brain you boot lives in this page's memory and nowhere else. What you teach it stays in it, and when you close the tab it is gone. New Adam gives you a fresh one from the shipped brain; Blank brain gives you one that has never been taught a single digit, so you can raise it yourself from nothing.

The map

The map shows the real populations at cluster grain: one hundred seeing cells, a thousand idea cells, the two hundred units that are its mouth and its ear at once, and a fast store of four thousand. Nodes light with real spikes from the presentation you just made. The lines are learned weight mass, and they move when you teach it.

Honesty notes

The shipped brain was lightly schooled and still confuses digits that look alike; that is the point, it leaves you something to teach. When its answer fizz is too far from every word it says it does not know rather than guessing. The model file is scrambled against casual download, and plainly said, that is a deterrent, not a vault. And the fast store really is written on every lesson and never read; it is drawn unconnected because it is unconnected.