CHRIS HAY

IDEAS · SYSTEMS · OBJECTS / LONDON · 2026

CHRIS HAY / THE MECHANISM / INTERACTIVE STUDY

Reading
by address.

Six planted facts. Twenty-four dimensions. One question to follow.

Choose a relation and an entity. Watch which keys activate, then read the answer those activations produce. Switch off the matching neuron to see what the remaining rows write.

01 / CHOOSE AN ADDRESS

These are invented places with planted answers. Choosing an address supplies its preassigned 24-number vector to the FFN.

02 / MATCH KEYS · APPLY RELU

Each row compares its key with the input. Negative matches become zero. Other rows can still activate.

  1. capital of AtlantisWRITES PARIS1.000
  2. currency of AtlantisWRITES EURO0.000
  3. language of AtlantisWRITES LATIN0.000
  4. capital of ZeriviaWRITES CAIRO0.310
  5. currency of ZeriviaWRITES RAND0.277
  6. language of ZeriviaWRITES TAMIL0.000

03 / COMBINE VALUES · READ THE OUTPUT

Paris

Planted answer: Paris. Recovered.

Every active row adds its value vector, weighted by its activation. The answer is the word whose reader scores that combined output highest.

Cairo
0.358
Euro
0.286
Latin
0.297
Paris
1.031
Rand
0.209
Tamil
0.197

DOT-PRODUCT SCORES / NOT PROBABILITIES

Inspect the calculation

activations = ReLU(W_in × address)
output = W_outᵀ × activations
answer = highest dot(output, word_reader)

There are six rows, each with 24 dimensions. This uses two matrix projections and a ReLU. The browser recomputes the result when you change the controls. Switching off a neuron sets its activation to zero before combining the values.

WHAT YOU ARE RUNNING

This is the constructed FFN from The Mechanism, adapted for the browser using the source’s six facts, seed-0 key and value vectors, and output readers. All six baseline answers match the saved Python example. The neuron switch is an added intervention you can explore here. No model server or download is needed.

A CONSTRUCTED MEMORY

The facts and their address vectors were supplied by hand. This example does not parse an arbitrary natural-language question.

Its six successful reads do not establish unlimited capacity, interference-free storage or the organisation of every fact in a trained model.

The value of the toy is that every operation can be inspected.

a constructed memorythe facts and their address vectors were supplied by hand. this example does not parse an arbitrary natural-language question.; its six successful reads do not establish unlimited capacity, interference-free storage or the organisation of every fact in a trained model.. The value of the toy is that every operation can be inspected.

CONTINUE THE INVESTIGATION

The notebook connects this small mechanism to the films, the native-model experiments and the question of how a useful address develops through a transformer.

Reading by addressThe map and its readersSource repository