Dontopedia

AnchorKAN

From Dontopedia, the open, paraconsistent wiki. (Last updated 2026-07-03.)

AnchorKAN has 82 facts recorded in Dontopedia across 14 references, with 5 live disagreements.

82 facts·69 predicates·14 sources·5 in dispute

Mostly:rdf:type(5), dominates outputs with(4), computes(2)

Maturity scale raw canonical shape-checked rule-derived certified

Inbound mentions (55)

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.

dominatesOutputOfDominates Output of(4)

advocatesFiringAdvocates Firing(1)

advocatesForImplementationAdvocates for Implementation(1)

assertsSubjectAsserts Subject(1)

assumesKnowledgeOfAssumes Knowledge of(1)

believesAnchorKanBetterForDeploymentsBelieves Anchor Kan Better for Deployments(1)

benefitsFromConstantSizeBenefits From Constant Size(1)

benefitsFromLinearTimeBenefits From Linear Time(1)

comparesPplCompares Ppl(1)

comparesTokPerSCompares Tok Per S(1)

containsModelContains Model(1)

contrastsWithContrasts With(1)

evaluatesAsBetterThroughputTradeoffEvaluates As Better Throughput Tradeoff(1)

frequentInOutputsOfFrequent in Outputs of(1)

has26PercentPplGapHas26 Percent Ppl Gap(1)

hasArchitecturallyDifferentRepresentationHas Architecturally Different Representation(1)

hasAutoFallbackHas Auto Fallback(1)

hasExperienceWithHas Experience With(1)

hasHigherBetaHas Higher Beta(1)

hasImplementedAndTestedHas Implemented and Tested(1)

hasMemberHas Member(1)

hasSmootherConvergenceHas Smoother Convergence(1)

hasSparserPatternHas Sparser Pattern(1)

hasSyncHubRoleForConfigHas Sync Hub Role for Config(1)

hasWorsePplHas Worse Ppl(1)

hedgesComparisonHedges Comparison(1)

heldByHeld by(1)

impliesWorsePplImplies Worse Ppl(1)

isCoreInnovationIs Core Innovation(1)

isGapOfIs Gap of(1)

isKeyBenefitIs Key Benefit(1)

isNotMasteredByIs Not Mastered by(1)

isOutputOfIs Output of(1)

isStyleOfIs Style of(1)

learnedByLearned by(1)

mentionsMentions(1)

mentionsTopicMentions Topic(1)

partOfPart of(1)

precedesPrecedes(1)

presupposesExistenceOfPresupposes Existence of(1)

reachesHigherFinalOrderReaches Higher Final Order(1)

reachesSameDcSyncStrengthAsReaches Same DC Sync Strength As(1)

referencesKnownModelReferences Known Model(1)

reportsParamReductionReports Param Reduction(1)

reportsSysRReports Sys R(1)

reportsUniformRReports Uniform R(1)

reportsVarianceReports Variance(1)

syncHubPrecedesSync Hub Precedes(1)

trainsVariantTrains Variant(1)

wouldBeFasterThanWould Be Faster Than(1)

wouldHelpWould Help(1)

wouldHelpPerformanceOfWould Help Performance of(1)

Other facts (78)

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.

78 facts
PredicateValueRef
Rdf:typeModel Architecture[11]
Rdf:typeAttention Configuration[12]
Rdf:typeModel Architecture[13]
Rdf:typeModel Kind[14]
Rdf:typeModel[14]
Dominates Outputs WithComma[14]
Dominates Outputs WithPeriod[14]
Dominates Outputs WithDash[14]
Dominates Outputs WithParentheses[14]
ComputesKan Spline Evaluation[4]
ComputesO Nm Causal Cumsum[4]
LearnedPunctuation Commonness Claim[14]
Learnedpunctuation is common[14]
Has Linear Time TrainingTraining[1]
RecoversNear Quadratic Kan Quality[1]
Has Constant Size Decode CacheDecode[1]
Enables PublicationResearch Thesis[1]
Competitive QualityKan[1]
Superior EfficiencyKan[1]
At Similar ScaleTheir Model[2]
Demonstrates Compile BenefitsThroughput[2]
Exists Prior RunAnchor Kan 65k Run[3]
Has Tok S6000[4]
Has Sync HubBlock 9[4]
Has R Hub0.561[4]
Has Slow Tok S6000[4]
Involves Heads12[4]
Involves Dimension768[4]
Involves Anchors32[4]
Has DC Mode0.316[4]
Optimizes ThroughputParameter Budget[5]
Uses Sinusoidal Fallback With Ropetrue[5]
Has Fewer Params24% fewer[5]
Has19 Percent Speed AdvantageSpectral[5]
Has Higher Early Variancetrue[5]
Is SparserSpectral[5]
Is FasterSpectral[5]
Has Dominant H0 in L1true[5]
Represents Same SignalSpectral[5]
Has Uniform Distribution Across BlocksR Values[5]
Has Distributed Representationtrue[5]
Has Bottleneck6K tok/s[6]
Exists As Modeltrue[6]
Provides Better Long Range Coupling ThanSpectralattention[7]
Aligns With Native PhysicsAntenna[8]
Avoids Hard DiscretizationImplementation[8]
Enables Stable AttractorsWithout Discretization[8]
Fits Core IssueTransport Works Identity Does Not[8]
Is Key Conceptual AnswerCurrent Problem[8]
Is Much Closer toAntenna Native Physics[8]
Provides Very Strong FitCurrent Problem[8]
Should Now Be FiredNext Action[8]
Suggests Way to AddStable Finite Identity Attractors[8]
Efficient to Trainway smaller faster to train[9]
Smaller ThanComparable Models[9]
Was Solidpretty solid[9]
References Known ModelAnchor Kan[9]
Scales WellChatgpt[9]
Faster to Train ThanComparable Models[9]
Shows No Negative Thingsdidn't show any really negative things[9]
Is Slower by Large MarginGrouped1[10]
Has Sync Hub LocationBlock 9[12]
Has Sync Strength0.561[12]
Has Baseline Throughput6000[12]
Has Throughput UnitTok S[12]
Has Performance CharacteristicSlow Speed[12]
Has Slowness ReasonCausal Cumsum Operation[12]
Outputs StylePunctuation Loaded Style[14]
Outputs Sample TextAnchorkan Output Sample[14]
Has Learning GapPunctuation Usage Knowledge Gap[14]
Dominates Every Outputtrue[14]
Appears inCode Block[14]
Outputs SampleAnchorkan Output Sample[14]
FollowsSpectral[14]
Output Characteristicheavily punctuation-loaded[14]
Punctuation Typescommas, periods, dashes, parentheses[14]
Example Outputthe, – for well on the to[14]
Has Deficithasn't learned when to use[14]

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.

hasLinearTimeTrainingblah/watt-activation/part-71
ex:training
recoversblah/watt-activation/part-71
ex:near-quadratic-kan-quality
hasConstantSizeDecodeCacheblah/watt-activation/part-71
ex:decode
enablesPublicationblah/watt-activation/part-71
ex:research-thesis
competitiveQualityblah/watt-activation/part-71
ex:kan
superiorEfficiencyblah/watt-activation/part-71
ex:kan
atSimilarScaleblah/watt-activation/part-105
ex:their-model
demonstratesCompileBenefitsblah/watt-activation/part-105
ex:throughput
existsPriorRunblah/watt-activation/part-153
ex:anchor-kan-65k-run
hasTokSblah/watt-activation/part-221
6000
hasSyncHubblah/watt-activation/part-221
ex:block-9
hasRHubblah/watt-activation/part-221
0.561
hasSlowTokSblah/watt-activation/part-221
6000
computesblah/watt-activation/part-221
ex:kan-spline-evaluation
computesblah/watt-activation/part-221
ex:o-nm-causal-cumsum
involvesHeadsblah/watt-activation/part-221
12
involvesDimensionblah/watt-activation/part-221
768
involvesAnchorsblah/watt-activation/part-221
32
hasDcModeblah/watt-activation/part-221
0.316
optimizesThroughputblah/watt-activation/part-236
ex:parameter-budget
usesSinusoidalFallbackWithRopeblah/watt-activation/part-236
true
hasFewerParamsblah/watt-activation/part-236
24% fewer
has19PercentSpeedAdvantageblah/watt-activation/part-236
ex:spectral
hasHigherEarlyVarianceblah/watt-activation/part-236
true
isSparserblah/watt-activation/part-236
ex:spectral
isFasterblah/watt-activation/part-236
ex:spectral
hasDominantH0InL1blah/watt-activation/part-236
true
representsSameSignalblah/watt-activation/part-236
ex:spectral
hasUniformDistributionAcrossBlocksblah/watt-activation/part-236
ex:r-values
hasDistributedRepresentationblah/watt-activation/part-236
true
hasBottleneckblah/watt-activation/part-223
6K tok/s
existsAsModelblah/watt-activation/part-223
true
providesBetterLongRangeCouplingThanblah/watt-activation/part-313
ex:spectralattention
alignsWithNativePhysicsblah/watt-activation/part-371
ex:antenna
avoidsHardDiscretizationblah/watt-activation/part-371
ex:implementation
enablesStableAttractorsblah/watt-activation/part-371
ex:without-discretization
fitsCoreIssueblah/watt-activation/part-371
ex:transport-works-identity-does-not
isKeyConceptualAnswerblah/watt-activation/part-371
ex:current-problem
isMuchCloserToblah/watt-activation/part-371
ex:antenna-native-physics
providesVeryStrongFitblah/watt-activation/part-371
ex:current-problem
shouldNowBeFiredblah/watt-activation/part-371
ex:next-action
suggestsWayToAddblah/watt-activation/part-371
ex:stable-finite-identity-attractors
efficientToTrainblah/watt-activation/part-381
way smaller faster to train
smallerThanblah/watt-activation/part-381
ex:comparable-models
wasSolidblah/watt-activation/part-381
pretty solid
referencesKnownModelblah/watt-activation/part-381
ex:anchor-kan
scalesWellblah/watt-activation/part-381
ex:chatgpt
fasterToTrainThanblah/watt-activation/part-381
ex:comparable-models
showsNoNegativeThingsblah/watt-activation/part-381
didn't show any really negative things
isSlowerByLargeMarginblah/watt-activation/part-103
ex:grouped1
typeblah/watt-activation/71
ex:ModelArchitecture
labelblah/watt-activation/71
anchor-KAN
typeblah/watt-activation/220
ex:AttentionConfiguration
labelblah/watt-activation/220
anchor_kan
hasSyncHubLocationblah/watt-activation/220
ex:block-9
hasSyncStrengthblah/watt-activation/220
0.561
hasBaselineThroughputblah/watt-activation/220
6000
hasThroughputUnitblah/watt-activation/220
ex:tok-s
hasPerformanceCharacteristicblah/watt-activation/220
ex:slow-speed
hasSlownessReasonblah/watt-activation/220
ex:causal-cumsum-operation
labelblah/watt-activation/222
Anchor_kan
typeblah/watt-activation/222
ex:ModelArchitecture
labeldocument/019a8f6a-d43b-40ba-afc1-247f4b73c3a5
AnchorKAN
typedocument/019a8f6a-d43b-40ba-afc1-247f4b73c3a5
ex:model-kind
outputsStyledocument/019a8f6a-d43b-40ba-afc1-247f4b73c3a5
ex:punctuation-loaded-style
outputsSampleTextdocument/019a8f6a-d43b-40ba-afc1-247f4b73c3a5
ex:anchorkan-output-sample
dominatesOutputsWithdocument/019a8f6a-d43b-40ba-afc1-247f4b73c3a5
ex:comma
dominatesOutputsWithdocument/019a8f6a-d43b-40ba-afc1-247f4b73c3a5
ex:period
dominatesOutputsWithdocument/019a8f6a-d43b-40ba-afc1-247f4b73c3a5
ex:dash
dominatesOutputsWithdocument/019a8f6a-d43b-40ba-afc1-247f4b73c3a5
ex:parentheses
learneddocument/019a8f6a-d43b-40ba-afc1-247f4b73c3a5
ex:punctuation-commonness-claim
hasLearningGapdocument/019a8f6a-d43b-40ba-afc1-247f4b73c3a5
ex:punctuation-usage-knowledge-gap
dominatesEveryOutputdocument/019a8f6a-d43b-40ba-afc1-247f4b73c3a5
true
appearsIndocument/019a8f6a-d43b-40ba-afc1-247f4b73c3a5
ex:code-block
outputsSampledocument/019a8f6a-d43b-40ba-afc1-247f4b73c3a5
ex:anchorkan-output-sample
followsdocument/019a8f6a-d43b-40ba-afc1-247f4b73c3a5
ex:spectral
typedocument/019a8f6a-d43b-40ba-afc1-247f4b73c3a5
ex:Model
outputCharacteristicdocument/019a8f6a-d43b-40ba-afc1-247f4b73c3a5
heavily punctuation-loaded
punctuationTypesdocument/019a8f6a-d43b-40ba-afc1-247f4b73c3a5
commas, periods, dashes, parentheses
exampleOutputdocument/019a8f6a-d43b-40ba-afc1-247f4b73c3a5
the, – for well on the to
learneddocument/019a8f6a-d43b-40ba-afc1-247f4b73c3a5
punctuation is common
hasDeficitdocument/019a8f6a-d43b-40ba-afc1-247f4b73c3a5
hasn't learned when to use

References (14)

14 references
  1. [1]Part 716 facts
    ctx:discord/blah/watt-activation/part-71
  2. [2]Part 1052 facts
    ctx:discord/blah/watt-activation/part-105
  3. [3]Part 1531 fact
    ctx:discord/blah/watt-activation/part-153
  4. [4]Part 22110 facts
    ctx:discord/blah/watt-activation/part-221
  5. [5]Part 23611 facts
    ctx:discord/blah/watt-activation/part-236
  6. [6]Part 2232 facts
    ctx:discord/blah/watt-activation/part-223
  7. [7]Part 3131 fact
    ctx:discord/blah/watt-activation/part-313
  8. [8]Part 3719 facts
    ctx:discord/blah/watt-activation/part-371
  9. [9]Part 3817 facts
    ctx:discord/blah/watt-activation/part-381
  10. [10]Part 1031 fact
    ctx:discord/blah/watt-activation/part-103
  11. [11]712 facts
    ctx:discord/blah/watt-activation/71
    • full textwatt-activation-71
      text/plain2 KBdoc:agent/watt-activation-71/82bde084-e631-42b5-99fe-0a0a8898ac2b
      Show excerpt
      [2026-03-07 17:14] xenonfun: ``` [E_mhkan_h5] iter 4000/10000 | avg_loss=4.7613 | PPL=116.9 | r=0.000 | E=0.000 | clusters=0 | 130 it/s (66.7K tok/s) │ │ [E_mhkan_h5] iter 6000/10000 | avg_loss=4.6
  12. [12]2208 facts
    ctx:discord/blah/watt-activation/220
    • full textwatt-activation-220
      text/plain3 KBdoc:agent/watt-activation-220/5c7f4a28-90e7-46de-ae1e-9e19a58c8d65
      Show excerpt
      [2026-03-11 04:42] xenonfun: FFN DFT — much richer specialization than spectral: ``` ┌─────┬───────┬────────┬────────────────┬────────────────┐ │ blk │ r │ FFN DC │ dominant mode │ pattern │ ├─────┼───────┼────────┼───────
  13. [13]2222 facts
    ctx:discord/blah/watt-activation/222
    • full textwatt-activation-222
      text/plain3 KBdoc:agent/watt-activation-222/d8201f0f-b5d1-4b50-9f4e-2aca2c0d4c1e
      Show excerpt
      [2026-03-11 05:02] xenonfun: ⏺ mx.compile with RotationalAdamW is a dead end — the optimizer creates new array objects on each step, so inputs=[model.state] captures stale references. The error "array without primitive" is exactly what CL
  14. ctx:claims/document/019a8f6a-d43b-40ba-afc1-247f4b73c3a5
    • full textxenonfun: well not that much speed up 46K now peak, think its memory bound already. 8K voc
      text/plain680 Bdiscord:msg/a3126764-fdd1-42f8-9653-a5170ea5bdef
      Show excerpt
      xenonfun: well not that much speed up 46K now peak, think its memory bound already. 8K vocab is signifigantly worse output at same training but makes sense, model was too lopsided with that much embeddings, but at 100K it did observe intere

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.