Dontopedia
Explore

Lohe Spherical Attention

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

Lohe Spherical Attention has 86 facts recorded in Dontopedia across 13 references, with 4 live disagreements.

86 facts·76 predicates·13 sources·4 in dispute

Mostly:has parameter(7), parameter name(3), outputs(2)

Maturity scale raw canonical shape-checked rule-derived certified

Has Parameterin disputehasParameter

  • H=4[1]all time · Part 201
  • n_layers=4[1]all time · Part 201
  • batch=4[1]all time · Part 201
  • d_model=256[1]all time · Part 201
  • vocab=1024[1]all time · Part 201
  • seq_len=128[1]all time · Part 201
  • G=4[1]all time · Part 201

Outputsin disputeoutputs

Projectsin disputeprojects

  • K[9]all time · Part 197
  • Q[9]all time · Part 197

Parameter Namein disputeparameterName

  • n_ffn_harmonics[12]all time · 258
  • n_ffn_groups[12]all time · 258
  • heads[12]all time · 258

Had Prior PplhadPriorPPL

  • 1350[8]all time · Part 318

Has Best Ppl Times Tok S Efficiency at Rot0 1hasBestPPLTimesTokSEfficiencyAtRot0-1

  • fast + good quality[8]all time · Part 318

Has Loss at Rot0 1hasLossAtRot0-1

  • 6.37[8]all time · Part 318

Has Params at Rot0 1hasParamsAtRot0-1

  • 12.3M[8]all time · Part 318

Has Ppl at Rot0 1hasPPLAtRot0-1

  • 582[8]all time · Part 318

Has Tok Per S at Rot0 1hasTokPerSAtRot0-1

  • 83625[8]all time · Part 318

Is3 2x Faster Than Topo at Rot0 1is3-2xFasterThanTopoAtRot0-1

Is47 Percent Faster Than Spectral At0 1is47PercentFasterThanSpectralAt0-1

Inbound mentions (20)

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.

advocatesUsageAdvocates Usage(1)

attentionTypeAttention Type(1)

composedOfComposed of(1)

concernsImplementationConcerns Implementation(1)

discussesLoheSphericalAttentionDiscusses Lohe Spherical Attention(1)

focusesOnModelTrainingFocuses on Model Training(1)

hasAttentionTypeHas Attention Type(1)

hasNaturalPairedAttentionHas Natural Paired Attention(1)

hasParameterHas Parameter(1)

hasSubcomponentHas Subcomponent(1)

implementsAttentionTypeImplements Attention Type(1)

involvesSketchingInvolves Sketching(1)

isMorePureVersionIs More Pure Version(1)

isPurerIs Purer(1)

isRunningIs Running(1)

requiresRewriteRequires Rewrite(1)

usesAttentionUses Attention(1)

verifiesBasicFunctionalityVerifies Basic Functionality(1)

wasCompletelyDanglingInWas Completely Dangling in(1)

wasNotForwardingIntoWas Not Forwarding Into(1)

Other facts (64)

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.

64 facts
PredicateValueRef
Overtakes at Higher RotTopo Star Kan Attention[8]
Overtakes Topo Star Kan at Rot0 1Topo Star Kan Attention[8]
Uses27 Percent Fewer Params Than SpectralSpectral Attention[8]
Within27 Percent Ppl of Spectral At0 1Spectral Attention[8]
Does Not Use HeadsHeads[7]
Has Nffn GroupsN Ffn Groups[7]
Has Nffn HarmonicsN Ffn Harmonics[7]
Does Not Hard Ban Non4 H in Forward Inferencenull[6]
Uses Exact Quaternion Sync atH Equals 4[6]
Uses Generic Spectral Sync Otherwisenull[6]
Applied toMultimodal Harmonicmlx[2]
Described in AttachmentScreenshot 2026 03 11 at 10 17 19 Pm Png[4]
Has Live VisualizationLohe Sph Live2 Png[4]
Exists As OptionAttn Type[1]
Has Loss Descent0.176[1]
Has Lower ThroughputSpectral Plus Lohe V2[1]
Has Mean R Global During Training0.61[1]
Has R Global at Random Init0.69[1]
Has Throughput29.5K tok/s[1]
Is Attention TypeModel Config[1]
Is Small Model ConfigG=4, H=4[1]
Is Tested inSmoke Test[1]
Is Usable Viaattn_type="lohe_spherical"[1]
Maintains R Global StabilityValid Range[1]
Matches Loss DescentSpectral Plus Lohe V2[1]
Documented inDocs Claude Lohe Attention Md[5]
Has Fewer Attention Params48%[5]
Has Same Dynamics AsFfnv2[5]
Has Throughput Cost~10% from sync step[5]
Planned Implementation{}[5]
Reduces Params by48%[5]
Eliminates Need forR Equals 0 5 Proxy[9]
Is Natural PairingLohe Ffnv2[9]
Outputs R forOptimizer Gate[9]
Projects OntoS H Minus 1[9]
RunsSpectral Lohe Sync[9]
Dominates Efficiency Frontier at Rot0 1Spectral Attention[8]
Is Componentnull[11]
Has Default Geometry8x4[6]
Uses Models Requested Geometrynull[6]
Exists As Pure Varianttrue[10]
Is PurerLohe Spherical Attention[10]
Achieves SyncWithin Block[1]
Contains R GlobalAttention Infos[1]
Dominates byProj V Proj Out[1]
Exhibits Within Block SyncAlready Present[1]
Expects Smaller Throughput GapGpt 2 Scale[1]
Is Viable AlternativeSpectral Attention[1]
Provides R Global MetricAttention Infos[1]
Scales to Gpt2Gpt 2[1]
Has Parameter Efficiencyusing 27% fewer params[3]
Has Speed Advantage47% faster[3]
Compared toSpectral Attention[3]
Has Perplexity Proximitywithin 27% PPL[3]
Has Tokens Per Second83625[3]
Has Speed Multiplier3.2× faster[3]
Has Perplexity582[3]
Overtakes Ranking ofTopo Star Plus Kan Attention[3]
Has Current Perplexity582[3]
Has Previous Perplexity1350[3]
Perplexity Change57% better[3]
Rdf:typeAttention Type[3]
Rdfs:labellohe_spherical[3]
Related toLohe Ffn V1[13]

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.

achievesSyncblah/watt-activation/part-201
ex:within-block
appliedToblah/watt-activation/part-240
ex:multimodal-harmonicmlx
comparedToblah/watt-activation/316
ex:spectral-attention
containsRGlobalblah/watt-activation/part-201
ex:attention-infos
describedInAttachmentblah/watt-activation/part-239
ex:screenshot-2026-03-11-at-10-17-19-pm-png
documentedInblah/watt-activation/part-199
ex:docs-claude-lohe-attention-md
doesNotHardBanNon4HInForwardInferenceblah/watt-activation/part-261
null
doesNotUseHeadsblah/watt-activation/part-260
ex:heads
dominatesByblah/watt-activation/part-201
ex:proj-v-proj-out
dominatesEfficiencyFrontierAtRot0-1blah/watt-activation/part-318
ex:spectral-attention
eliminatesNeedForblah/watt-activation/part-197
ex:r-equals-0-5-proxy
exhibitsWithinBlockSyncblah/watt-activation/part-201
ex:already-present
existsAsOptionblah/watt-activation/part-201
ex:attn-type
existsAsPureVariantblah/watt-activation/part-238
true
expectsSmallerThroughputGapblah/watt-activation/part-201
ex:gpt-2-scale
hadPriorPPLblah/watt-activation/part-318
1350
hasBestPPLTimesTokSEfficiencyAtRot0-1blah/watt-activation/part-318
fast + good quality
hasCurrentPerplexityblah/watt-activation/316
582
hasDefaultGeometryblah/watt-activation/part-261
ex:8x4
hasFewerAttentionParamsblah/watt-activation/part-199
48%
hasLiveVisualizationblah/watt-activation/part-239
ex:lohe-sph-live2-png
hasLossAtRot0-1blah/watt-activation/part-318
6.37
hasLossDescentblah/watt-activation/part-201
0.176
hasLowerThroughputblah/watt-activation/part-201
ex:spectral-plus-lohe-v2
hasMeanRGlobalDuringTrainingblah/watt-activation/part-201
0.61
hasNffnGroupsblah/watt-activation/part-260
ex:n-ffn-groups
hasNffnHarmonicsblah/watt-activation/part-260
ex:n-ffn-harmonics
hasParameterblah/watt-activation/part-201
H=4
hasParameterblah/watt-activation/part-201
n_layers=4
hasParameterblah/watt-activation/part-201
batch=4
hasParameterblah/watt-activation/part-201
d_model=256
hasParameterblah/watt-activation/part-201
vocab=1024
hasParameterblah/watt-activation/part-201
seq_len=128
hasParameterblah/watt-activation/part-201
G=4
hasParameterEfficiencyblah/watt-activation/316
using 27% fewer params
hasParamsAtRot0-1blah/watt-activation/part-318
12.3M
hasPerplexityblah/watt-activation/316
582
hasPerplexityProximityblah/watt-activation/316
within 27% PPL
hasPPLAtRot0-1blah/watt-activation/part-318
582
hasPreviousPerplexityblah/watt-activation/316
1350
hasRGlobalAtRandomInitblah/watt-activation/part-201
0.69
hasSameDynamicsAsblah/watt-activation/part-199
ex:ffnv2
hasSpeedAdvantageblah/watt-activation/316
47% faster
hasSpeedMultiplierblah/watt-activation/316
3.2× faster
hasThroughputblah/watt-activation/part-201
29.5K tok/s
hasThroughputCostblah/watt-activation/part-199
~10% from sync step
hasTokensPerSecondblah/watt-activation/316
83625
hasTokPerSAtRot0-1blah/watt-activation/part-318
83625
is3-2xFasterThanTopoAtRot0-1blah/watt-activation/part-318
ex:topo-star-kan-attention
is47PercentFasterThanSpectralAt0-1blah/watt-activation/part-318
ex:spectral-attention
isAttentionTypeblah/watt-activation/part-201
ex:model-config
isComponentblah/watt-activation/part-246
null
isNaturalPairingblah/watt-activation/part-197
ex:lohe-ffnv2
isPurerblah/watt-activation/part-238
ex:lohe-spherical-attention
isSmallModelConfigblah/watt-activation/part-201
G=4, H=4
isTestedInblah/watt-activation/part-201
ex:smoke-test
isUsableViablah/watt-activation/part-201
attn_type="lohe_spherical"
isViableAlternativeblah/watt-activation/part-201
ex:spectral-attention
maintainsRGlobalStabilityblah/watt-activation/part-201
ex:valid-range
matchesLossDescentblah/watt-activation/part-201
ex:spectral-plus-lohe-v2
outputsblah/watt-activation/part-197
ex:attention-values
outputsblah/watt-activation/part-197
ex:sync-state
outputsRForblah/watt-activation/part-197
ex:optimizer-gate
overtakesAtHigherRotblah/watt-activation/part-318
ex:topo-star-kan-attention
overtakesRankingOfblah/watt-activation/316
ex:topo-star-plus-kan-attention
overtakesTopoStarKanAtRot0-1blah/watt-activation/part-318
ex:topo-star-kan-attention
parameterNameblah/watt-activation/258
n_ffn_harmonics
parameterNameblah/watt-activation/258
n_ffn_groups
parameterNameblah/watt-activation/258
heads
perplexityChangeblah/watt-activation/316
57% better
plannedImplementationblah/watt-activation/part-199
{}
projectsblah/watt-activation/part-197
ex:k
projectsblah/watt-activation/part-197
ex:q
projectsOntoblah/watt-activation/part-197
ex:s-h-minus-1
providesRGlobalMetricblah/watt-activation/part-201
ex:attention-infos
labelblah/watt-activation/316
lohe_spherical
typeblah/watt-activation/316
ex:AttentionType
reducesParamsByblah/watt-activation/part-199
48%
relatedToblah/watt-activation/part-458
ex:lohe-ffn-v1
runsblah/watt-activation/part-197
ex:spectral-lohe-sync
scalesToGpt2blah/watt-activation/part-201
ex:gpt-2
uses27PercentFewerParamsThanSpectralblah/watt-activation/part-318
ex:spectral-attention
usesExactQuaternionSyncAtblah/watt-activation/part-261
ex:h-equals-4
usesGenericSpectralSyncOtherwiseblah/watt-activation/part-261
null
usesModelsRequestedGeometryblah/watt-activation/part-261
null
within27PercentPPLOfSpectralAt0-1blah/watt-activation/part-318
ex:spectral-attention

References (13)

13 references
  1. [1]Part 20127 facts
    customctx:discord/blah/watt-activation/part-201
  2. [2]Part 2401 fact
    customctx:discord/blah/watt-activation/part-240
  3. customctx:discord/blah/watt-activation/316
    • full textwatt-activation-316
      text/plain3 KBdoc:agent/watt-activation-316/c4f50044-9de8-4f34-94c9-40d82203b46b
      Show excerpt
      [2026-03-15 01:57] xenonfun: ``` rot_strength = 0.1 (new default) ┌────────────────┬──────┬─────┬────────┬────────┐ │ Attention │ Loss │ PPL │ Tok/s │ Params │ ├────────────────┼──────┼─────┼────────┼────────┤ │ spectral
  4. [4]Part 2392 facts
    customctx:discord/blah/watt-activation/part-239
  5. [5]Part 1996 facts
    customctx:discord/blah/watt-activation/part-199
  6. [6]Part 2615 facts
    customctx:discord/blah/watt-activation/part-261
  7. [7]Part 2603 facts
    customctx:discord/blah/watt-activation/part-260
  8. [8]Part 31813 facts
    customctx:discord/blah/watt-activation/part-318
  9. [9]Part 1979 facts
    customctx:discord/blah/watt-activation/part-197
  10. [10]Part 2382 facts
    customctx:discord/blah/watt-activation/part-238
  11. [11]Part 2461 fact
    customctx:discord/blah/watt-activation/part-246
  12. customctx:discord/blah/watt-activation/258
    • full textwatt-activation-258
      text/plain3 KBdoc:agent/watt-activation-258/f49597e9-7ccc-4374-b019-f99f5ba153ea
      Show excerpt
      [2026-03-12 16:15] xenonfun: ``` Where I'd push back: Codex is critiquing the rhetoric more than the architecture. The claim isn't "text is fundamentally phase-based" — it's "we can project text into a phase-valued latent space and u
  13. [13]Part 4581 fact
    customctx:discord/blah/watt-activation/part-458

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.