Dontopedia

AnchorKAN

From Dontopedia, the open, paraconsistent wiki. (Last updated 2026-06-06.)

AnchorKAN has 80 facts recorded in Dontopedia across 11 references, with 6 live disagreements.

80 facts·68 predicates·11 sources·6 in dispute

Mostly:rdf:type(4), uses soft attractors(2), tokens seen so far(2)

Maturity scale raw canonical shape-checked rule-derived certified

Inbound mentions (16)

Other subjects in dontopedia point AT this entity as a value. These are inverse relationships — e.g. "X motherOf this subject" — and answer questions the forward facts can't. Grouped by predicate.

areActiveAre Active(1)

classifiedAsUselessClassified As Useless(1)

concernsSummarizingConcerns Summarizing(1)

estimatedSpeedAdvantageEstimated Speed Advantage(1)

exhibitsSimilarPatternToExhibits Similar Pattern to(1)

fasterThanAnchorkanFaster Than Anchorkan(1)

hasSubjectHas Subject(1)

hasWorsePerplexityThanHas Worse Perplexity Than(1)

includesVariantIncludes Variant(1)

involvesEntityInvolves Entity(1)

plannedSuccessorToPlanned Successor to(1)

portsPorts(1)

presentsHeadToHeadComparisonPresents Head to Head Comparison(1)

producesCoherentEnglishWordsProduces Coherent English Words(1)

referencesInComparisonReferences in Comparison(1)

teleologicalForAnchorkanTeleological for Anchorkan(1)

Other facts (75)

The long tail: predicates that appear too rarely to warrant their own section. Filter or scroll to find a specific one. Each row links to its source.

75 facts
PredicateValueRef
Rdf:typeModel Architecture[6]
Rdf:typeModel Architecture[7]
Rdf:typeArchitecture[9]
Rdf:typeMechanism[10]
Uses Soft AttractorsMemory Mechanism[1]
Uses Soft AttractorsSoft Attractors[5]
Tokens Seen So Far~65M (16K × 4096)[2]
Tokens Seen So Far65000000[6]
Learning Statelearning vocabulary and basic structure[6]
Learning Statenot developed strong topical coherence[6]
Is Evidence AgainstSoft Finite Attractors[11]
Is Evidence AgainstFix[11]
Has S2 At20k1.8%[1]
Marginally Worse Than BaselineBaseline[1]
Has Key InsightSoft attractors: active but useless[1]
Lacks Strong Topical Coherence YetAt Ppl 242[2]
Ontologically Similar toGpt 2 Architecture[2]
Matches Ppl at Equivalent ComputeQuadratic Attention[2]
Seen Tokens Vs Gpt2 Completion~30x fewer[2]
Known to MatchQuadratic Attention[2]
Is Learning Vocabulary and Basic StructureClearly[2]
Uses AttentionAnchorkan O L M[2]
Has Context Length4096[2]
Has Data Gap Vs Gpt2~30x less data[2]
Ppl Measured onTraining Loss[2]
Has Params145M[2]
Trained onFineweb Edu[2]
Tokens At2 Epochs~266M[2]
Has Significantly Fewer Tokens ThanGpt 2 Small[2]
Uses Fineweb Edu CorpusStandard Dataset[2]
Learned Punctuation Commontrue[3]
Uses Mx Cumsumplain mx.cumsum (not _gated_cumsum)[3]
Uses Commas Periods Dashes Parenthesesdominate every output[3]
Example Output"the, – for well on the to"[3]
Exists As Compile Compatibletrue[3]
Has No Blockerstrue[3]
Has Not Learned When to Use Punctuationtrue[3]
Heavily Punctuation Loadedtrue[3]
Holds Cum Gv Tensor(8, 2048, 4, 8, 208)[3]
Holds V Anchors Tensor(8, 2048, 4, 8, 208)[3]
Slower Due To5 D TensorsSpectral[3]
Is Fully Compile Compatibletrue[3]
Direction Is Closedtrue[4]
References Prior Experimenttrue[4]
S3 Rebinding Score42.8%[5]
Dc128 Score93.9%[5]
Worse Than Baseline on Multiple MetricsS1 S3 S4[5]
S2 Distractor Score1.8%[5]
S1 Direct Score52.1%[5]
Bpb Score2.105[5]
S5 Scoped Score2.2%[5]
S4 Multi Entity Score20.1%[5]
Parameter Count145000000[6]
Tokens at Two Epochs266000000[6]
Data Volume Compared toGpt 2 Small[6]
Has Relative Data Volume~30x less[6]
Context Length4096[6]
Evaluated onTraining Loss Metric[6]
Uses Attention MechanismAnchorkan O Lm[6]
Has Relative Token Count~30x fewer[6]
Is Bottleneckfalse[6]
Has Perplexity242[6]
Perplexity ContextEval State 242[6]
Target Perplexity185[6]
Expected Quality at Ppl185noticeably better[6]
Status Asimplemented and tested[10]
Close to Native Physics ofAntenna[10]
ProvidesFinite Attractor Mechanism[10]
Described Asmore natural first attempt[10]
Provides FeatureStable Finite Identity Attractors[10]
Characterized Asvery strong fit[10]
Recommended ActionShould Now Be Fired[10]
Uses Anchorstrue[11]
Anchors Did Not Help Persistencetrue[11]
Strengthens Case forDiscrete Identity Coding[11]

Timeline

Timeline axis is valid_time — when each source says the fact was true in the world, not when Dontopedia learned about it. Retracted rows are kept for provenance; coloured stripes indicate the context kind.

80 undated
—
hasS2At20kblah/random/part-38
1.8%
—
marginallyWorseThanBaselineblah/random/part-38
ex:baseline
—
usesSoftAttractorsblah/random/part-38
ex:memory-mechanism
—
hasKeyInsightblah/random/part-38
Soft attractors: active but useless
—
lacksStrongTopicalCoherenceYetblah/watt-activation/part-92
ex:at-ppl-242
—
ontologicallySimilarToblah/watt-activation/part-92
ex:gpt-2-architecture
—
matchesPPLAtEquivalentComputeblah/watt-activation/part-92
ex:quadratic-attention
—
seenTokensVsGpt2Completionblah/watt-activation/part-92
~30x fewer
—
knownToMatchblah/watt-activation/part-92
ex:quadratic-attention
—
isLearningVocabularyAndBasicStructureblah/watt-activation/part-92
ex:clearly
—
usesAttentionblah/watt-activation/part-92
ex:anchorkan-o-l-m
—
hasContextLengthblah/watt-activation/part-92
4096
—
hasDataGapVsGpt2blah/watt-activation/part-92
~30x less data
—
pplMeasuredOnblah/watt-activation/part-92
ex:training-loss
—
hasParamsblah/watt-activation/part-92
145M
—
tokensSeenSoFarblah/watt-activation/part-92
~65M (16K × 4096)
—
trainedOnblah/watt-activation/part-92
ex:fineweb-edu
—
tokensAt2Epochsblah/watt-activation/part-92
~266M
—
hasSignificantlyFewerTokensThanblah/watt-activation/part-92
ex:gpt-2-small
—
usesFineweb-eduCorpusblah/watt-activation/part-92
ex:standard-dataset
—
learnedPunctuationCommonblah/watt-activation/part-232
true
—
usesMxCumsumblah/watt-activation/part-232
plain mx.cumsum (not _gated_cumsum)
—
usesCommasPeriodsDashesParenthesesblah/watt-activation/part-232
dominate every output
—
exampleOutputblah/watt-activation/part-232
"the, – for well on the to"
—
existsAsCompileCompatibleblah/watt-activation/part-232
true
—
hasNoBlockersblah/watt-activation/part-232
true
—
hasNotLearnedWhenToUsePunctuationblah/watt-activation/part-232
true
—
heavilyPunctuationLoadedblah/watt-activation/part-232
true
—
holdsCumGvTensorblah/watt-activation/part-232
(8, 2048, 4, 8, 208)
—
holdsVAnchorsTensorblah/watt-activation/part-232
(8, 2048, 4, 8, 208)
—
slowerDueTo5DTensorsblah/watt-activation/part-232
ex:spectral
—
isFullyCompileCompatibleblah/watt-activation/part-232
true
—
directionIsClosedblah/watt-activation/part-374
true
—
referencesPriorExperimentblah/watt-activation/part-374
true
—
s3RebindingScoreblah/watt-activation/part-376
42.8%
—
dc128Scoreblah/watt-activation/part-376
93.9%
—
usesSoftAttractorsblah/watt-activation/part-376
ex:soft-attractors
—
worseThanBaselineOnMultipleMetricsblah/watt-activation/part-376
ex:s1-s3-s4
—
s2DistractorScoreblah/watt-activation/part-376
1.8%
—
s1DirectScoreblah/watt-activation/part-376
52.1%
—
bpbScoreblah/watt-activation/part-376
2.105
—
s5ScopedScoreblah/watt-activation/part-376
2.2%
—
s4MultiEntityScoreblah/watt-activation/part-376
20.1%
—
typeblah/watt-activation/92
ex:ModelArchitecture
—
labelblah/watt-activation/92
AnchorKAN
—
parameterCountblah/watt-activation/92
145000000
—
tokensSeenSoFarblah/watt-activation/92
65000000
—
tokensAtTwoEpochsblah/watt-activation/92
266000000
—
dataVolumeComparedToblah/watt-activation/92
ex:gpt-2-small
—
hasRelativeDataVolumeblah/watt-activation/92
~30x less
—
contextLengthblah/watt-activation/92
4096
—
evaluatedOnblah/watt-activation/92
ex:training-loss-metric
—
usesAttentionMechanismblah/watt-activation/92
ex:anchorkan-o-lm
—
hasRelativeTokenCountblah/watt-activation/92
~30x fewer
—
isBottleneckblah/watt-activation/92
false
—
hasPerplexityblah/watt-activation/92
242
—
perplexityContextblah/watt-activation/92
ex:eval-state-242
—
learningStateblah/watt-activation/92
learning vocabulary and basic structure
—
learningStateblah/watt-activation/92
not developed strong topical coherence
—
targetPerplexityblah/watt-activation/92
185
—
expectedQualityAtPpl185blah/watt-activation/92
noticeably better
—
typeblah/watt-activation/103
ex:ModelArchitecture
—
labelblah/watt-activation/311
AnchorKAN
—
typeblah/watt-activation/318
ex:Architecture
—
labelblah/watt-activation/318
AnchorKAN
—
labelblah/watt-activation/369
AnchorKAN
—
typeblah/watt-activation/369
ex:Mechanism
—
statusAsblah/watt-activation/369
implemented and tested
—
closeToNativePhysicsOfblah/watt-activation/369
ex:antenna
—
providesblah/watt-activation/369
ex:finite-attractor-mechanism
—
describedAsblah/watt-activation/369
more natural first attempt
—
providesFeatureblah/watt-activation/369
ex:stable-finite-identity-attractors
—
characterizedAsblah/watt-activation/369
very strong fit
—
recommendedActionblah/watt-activation/369
ex:should-now-be-fired
—
labelblah/watt-activation/372
AnchorKAN
—
usesAnchorsblah/watt-activation/372
true
—
anchorsDidNotHelpPersistenceblah/watt-activation/372
true
—
isEvidenceAgainstblah/watt-activation/372
ex:soft-finite-attractors
—
isEvidenceAgainstblah/watt-activation/372
ex:fix
—
strengthensCaseForblah/watt-activation/372
ex:discrete-identity-coding

References (11)

11 references
  1. [1]Part 384 facts
    ctx:discord/blah/random/part-38
  2. [2]Part 9216 facts
    ctx:discord/blah/watt-activation/part-92
  3. [3]Part 23212 facts
    ctx:discord/blah/watt-activation/part-232
  4. [4]Part 3742 facts
    ctx:discord/blah/watt-activation/part-374
  5. [5]Part 3769 facts
    ctx:discord/blah/watt-activation/part-376
  6. [6]9218 facts
    ctx:discord/blah/watt-activation/92
    • full textwatt-activation-92
      text/plain3 KBdoc:agent/watt-activation-92/a597b55a-3d12-478b-951d-f09c655a8870
      Show excerpt
      [2026-03-08 01:46] xenonfun: ``` Direct comparison is tricky but here are the reference points: GPT-2 Small (124M) published benchmarks: - WikiText-103 test: 29.4 PPL - Penn Treebank: 65.9 PPL - Trained on ~8-9B tokens of WebText …
  7. [7]1031 fact
    ctx:discord/blah/watt-activation/103
    • full textwatt-activation-103
      text/plain3 KBdoc:agent/watt-activation-103/6d322edd-8b82-4859-be6f-bc7033a53fe1
      Show excerpt
      [2026-03-08 18:36] xenonfun: It appears your agents have actually already done all this work <@1211062099137265723> in your repo already. https://github.com/MonumentalSystems/harmonic-gpt/blob/master/docs/M3_DEPLOY_NOTES.md (files: Screens…
  8. [8]3111 fact
    ctx:discord/blah/watt-activation/311
    • full textwatt-activation-311
      text/plain2 KBdoc:agent/watt-activation-311/f942f53b-f6c0-497d-a8cc-7bbe7ae3efb9
      Show excerpt
      [2026-03-15 00:39] xenonfun: Key Findings Star topology is best (1,685.6 ppl) — 8% better than baseline anchor_kan (1,756-2,013 range). The hub-and-spoke structure concentrates coupling through one central anchor, similar to the DC mode …
  9. [9]3182 facts
    ctx:discord/blah/watt-activation/318
    • full textwatt-activation-318
      text/plain3 KBdoc:agent/watt-activation-318/f52d95a8-f461-40d1-9360-f08558b18eb1
      Show excerpt
      [2026-03-15 02:47] xenonfun: ⏺ I see you're working on wire encoding / phase modulation — that's a fascinating direction. Let me check what you've got: [2026-03-15 02:47] lisamegawatts: Wire QPSK + Standard: PPL 4.94, Byte Accuracy 51.5% T…
  10. [10]3699 facts
    ctx:discord/blah/watt-activation/369
    • full textwatt-activation-369
      text/plain2 KBdoc:agent/watt-activation-369/0cb2b937-fe59-4554-9b34-62ddb285f694
      Show excerpt
      [2026-03-18 16:16] xenonfun: Yes — this is very relevant, and it changes the ranking. Given: S1 plateauing again at 4.1% DC continuing to rise and the fact that you already have AnchorKAN implemented and tested I would now rank the mec…
  11. [11]3726 facts
    ctx:discord/blah/watt-activation/372
    • full textwatt-activation-372
      text/plain2 KBdoc:agent/watt-activation-372/5df1e4bc-b9b4-4d7d-ad8f-c18916f7e8ae
      Show excerpt
      [2026-03-18 17:55] xenonfun: ``` Recommended panels (3) Panel 1: Anchor Health (line chart, time series) - Y-axis left: anchor_perplexity (range 0 to anchor_count, e.g. 32). Line color: blue. - Y-axis right: anchor_dead count. Line…

See also

Keep researching

Missing something or suspicious of what's here? Kick off a research session — a Claude agent will investigate, cite its sources, and file new facts into a dedicated context you can review before accepting into the shared view.

↑