Does useful knowledge remain available?
Retain what matters, with enough context to interpret it later.
ASIC Research / Kanzi Program
How should intelligence learn from experience?
Kanzi investigates how a system can accumulate knowledge, change its understanding, and carry what it learns into new situations.
Explore KanziKANZI / A QUESTION ABOUT LEARNINGILLUSTRATIVE STUDY
Observation 01
A / accepted
B / rejected
New evidence
Ring alone
rejectedBar alone
accepted
rejectedBar alone
accepted
An explanation must be able to change.
01 / Observe02 / Challenge03 / Revise
An explanation is only the beginning.Research asks what happens when experience challenges it.
The Kanzi Program
Learning is more than keeping a record.
Kanzi is a research program within ASIC Research. Its central question is how machine intelligence can build on experience: preserving useful knowledge, revising an explanation when evidence changes, and applying what it learned beyond the original case.
Research question / 01What persists?
FIELD NOTES01
02
03
02
03
01 / OBSERVATIONTwo features appear together.Keep the result and its conditions.
02 / WORKING IDEAThe ring may explain it.Mark the idea as provisional.
03 / PROVENANCEWhich observation supports the idea?Retain the source of the conclusion.
ObservationInterpretationSource
01 / Retain
What should persist?
An observation is useful only when its meaning, conditions, and origin can be carried forward. Kanzi investigates what to retain as knowledge and what should remain temporary context.
The research record distinguishes an observation from a working explanation and keeps a path back to the evidence.
Research question / 02When should an explanation change?
FIRST EXPLANATIONThe ring predicts acceptance.Provisional
CONTRASTING OBSERVATIONS
ring / rejected
bar / accepted
REVISED EXPLANATIONThe bar predicts acceptance.Supported by the new observations
Earlier idea retainedReason for revision visible
02 / Revise
When should understanding change?
New evidence can challenge a confident-looking answer. The question is how a system recognizes that challenge, updates its explanation, and remains able to say why it changed.
In this illustrative study, the first two cases cannot tell us whether the ring or bar matters. Later cases can.
Research question / 03What carries over?
KNOWN CASES
A / accepted
D / accepted
Candidate ruleLook for the bar
NEW CASE
Test the idea here
Different appearanceSame question
03 / Apply
Will the idea travel?
A system can remember an example without learning its underlying relationship. Kanzi asks whether an explanation remains useful when the next situation looks different.
A new case should test the rule, rather than repeat the case that produced it.
Research question / 04What remains reliable?
EVALUATION SHEETRECHECK AFTER LEARNING
Earlier casesCan it still explain what came before?↗
New evidenceCan it use what was just learned?↗
Unseen caseDoes the explanation travel?↗
Revisit the oldProbe the new
04 / Preserve
What must learning preserve?
Learning something new should prompt a fresh look at older capabilities. Research must test both the new behavior and whether earlier understanding remains reliable.
This is an evaluation question, not a reported result for Kanzi or Arkhe.
An illustrative experiment
When the evidence changes.
Move through a small, invented study. Notice what the first cases leave unresolved, then what new evidence forces the explanation to change.
EVIDENCE RECORD / SYNTHETIC EXAMPLESTAGE 01 OF 05
SPECIMEN AAccepted
SPECIMEN BRejected
CURRENT EXPLANATIONTwo observations do not isolate the cause.
OBSERVATIONEXPLANATIONTEST
Step 01 / 05
Two observations, several possible explanations.
A specimen with both a ring and a bar is accepted. A specimen with neither is rejected. Either feature could explain the result.
This sequence explains a research question. It is an illustration, not a demonstration of Kanzi performance or a measured Arkhe result.
What counts as progress
The test is in what remains true.
Research needs ways to distinguish a convincing demonstration from durable learning. These questions guide the work; they do not report Kanzi results.
Does contrary evidence lead to revision?
Show what changed and the observation that caused it.
Does an idea transfer beyond its examples?
Test a genuinely new case, not a repeated prompt.
Are earlier abilities still intact?
Recheck prior cases after introducing new learning.
Within ASIC
Research opens the question. Intelligence makes validated advances useful.
ASIC Research pursues foundational work through Kanzi. ASIC Intelligence develops and operates products, including Arkhe and Pipwick. A research direction becomes a product claim only when its behavior has been validated in the relevant configuration.
Explore ASIC IntelligenceASIC Research