Lab notebook · AKD1500

An AKD1500 uses prediction errors to gate teaching, filmed live

2026-11-16 · AKD1500on-chip learningprediction
The complete continuous screen capture, encoded at 60 fps (14 seconds). Each row appears after its whole queue of 48 predictions has been read back from the chip. The final readback remains on screen, labelled as finished.

The AKD1500 can use its own prediction errors to gate teaching. In this live repeat of a fixed binary-stream test, it opened 2 of 48 teaching gates and recovered both changed answers. Always-teach opened 48; cutting the gate or disabling learning kept the old answers. Holding history fixed caused errors.

Numbers

One synthetic stream of 48 next-bit predictions, with the rule changing two answers after step 16, run under five conditions:

ConditionTeaching gates openedChanged answers on revisit (steps 26, 27)Errors before change
Error-gated teaching2 of 48 (steps 18, 19)both right0
Always teach48 of 48both right0
Teaching gate cut0 of 48both old answers0
Learning disabledlearning offboth old answers0
History held fixed15 of 48both right7

Every output and every measured memory snapshot in all five conditions matched the earlier physical run byte for byte. The earlier study ran six preselected synthetic streams; error-gated teaching opened 2 of 48 gates and recovered both changed answers in all six.

The part worth knowing

In the error-gated condition, the teaching gate opens only when the prediction and supplied observation disagree. It recovers the same changed answers as always-teach, with 2 gate openings instead of 48. The short history and comparison run as programmed circuits on the chip; the host supplies the stream, initializes the rule and reads results back.

Limits

  • We supply the stream, the starting history and the initial rule. The history and error circuits are programmed on the chip, not learned.
  • This is a repeat of one previously tested synthetic stream, not a new task or a representative sample.
  • Gate openings are not connection writes: error-gated and always-teach each changed connections twice in the reference trace.
  • No timing advantage and no energy saving is claimed or measured.

Evidence and availability