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.
Mostly:rdf:type(5), dominates outputs with(4), computes(2)
Maturity scale
raw canonical shape-checked rule-derived certifiedInbound 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)
- Comma
ex:comma - Dash
ex:dash - Parentheses
ex:parentheses - Period
ex:period
advocatesFiringAdvocates Firing(1)
- Xenonfun
ex:xenonfun
advocatesForImplementationAdvocates for Implementation(1)
- Xenonfun
ex:xenonfun
assertsSubjectAsserts Subject(1)
- Takeaway Anchor Kan
ex:takeaway-anchor-kan
assumesKnowledgeOfAssumes Knowledge of(1)
- Xenonfun
ex:xenonfun
believesAnchorKanBetterForDeploymentsBelieves Anchor Kan Better for Deployments(1)
- Xenonfun
ex:xenonfun
benefitsFromConstantSizeBenefits From Constant Size(1)
- Decode Cache
ex:decode-cache
benefitsFromLinearTimeBenefits From Linear Time(1)
- Training
ex:training
comparesPplCompares Ppl(1)
- Row 1
ex:row-1
comparesTokPerSCompares Tok Per S(1)
- Row 2
ex:row-2
containsModelContains Model(1)
- Code Block
ex:code-block
contrastsWithContrasts With(1)
- Spherical Vq
ex:spherical-vq
evaluatesAsBetterThroughputTradeoffEvaluates As Better Throughput Tradeoff(1)
- Key Takeaway
ex:key-takeaway
frequentInOutputsOfFrequent in Outputs of(1)
- Punctuation
ex:punctuation
has26PercentPplGapHas26 Percent Ppl Gap(1)
- Spectral
ex:spectral
hasArchitecturallyDifferentRepresentationHas Architecturally Different Representation(1)
- Row 4
ex:row-4
hasAutoFallbackHas Auto Fallback(1)
- Parallel Prefill
ex:parallel-prefill
hasExperienceWithHas Experience With(1)
- Xenonfun
ex:xenonfun
hasHigherBetaHas Higher Beta(1)
- Spectral
ex:spectral
hasImplementedAndTestedHas Implemented and Tested(1)
- Recipient
ex:recipient
hasMemberHas Member(1)
- Code Block
ex:code-block
hasSmootherConvergenceHas Smoother Convergence(1)
- Spectral
ex:spectral
hasSparserPatternHas Sparser Pattern(1)
- Row 4
ex:row-4
hasSyncHubRoleForConfigHas Sync Hub Role for Config(1)
- Block 9
ex:block-9
hasWorsePplHas Worse Ppl(1)
- Spectralattention
ex:spectralattention
hedgesComparisonHedges Comparison(1)
- Xenonfun
ex:xenonfun
heldByHeld by(1)
- Punctuation Commonness Claim
ex:punctuation-commonness-claim
impliesWorsePplImplies Worse Ppl(1)
- Row 1
ex:row-1
isCoreInnovationIs Core Innovation(1)
- Low Rank Learned Routing Structure
ex:low-rank-learned-routing-structure
isGapOfIs Gap of(1)
- Punctuation Usage Knowledge Gap
ex:punctuation-usage-knowledge-gap
isKeyBenefitIs Key Benefit(1)
- Low Rank Learned Routing Structure
ex:low-rank-learned-routing-structure
isNotMasteredByIs Not Mastered by(1)
- Punctuation Usage Knowledge Gap
ex:punctuation-usage-knowledge-gap
isOutputOfIs Output of(1)
- Anchorkan Output Sample
ex:anchorkan-output-sample
isStyleOfIs Style of(1)
- Punctuation Loaded Style
ex:punctuation-loaded-style
learnedByLearned by(1)
- Punctuation Commonness Claim
ex:punctuation-commonness-claim
mentionsMentions(1)
- Source Utterance
ex:source-utterance
mentionsTopicMentions Topic(1)
- Log Entry 2026 03 11 20 31
ex:log-entry-2026-03-11-20-31
partOfPart of(1)
- Current Training Run
ex:current-training-run
precedesPrecedes(1)
- Spectral
ex:spectral
presupposesExistenceOfPresupposes Existence of(1)
- Text
ex:text
reachesHigherFinalOrderReaches Higher Final Order(1)
- Spectral
ex:spectral
reachesSameDcSyncStrengthAsReaches Same DC Sync Strength As(1)
- Spectral
ex:spectral
referencesKnownModelReferences Known Model(1)
- Anchor Kan
ex:anchor-kan
reportsParamReductionReports Param Reduction(1)
- Row 2
ex:row-2
reportsSysRReports Sys R(1)
- Row 2
ex:row-2
reportsUniformRReports Uniform R(1)
- Row 3
ex:row-3
reportsVarianceReports Variance(1)
- Row 1
ex:row-1
syncHubPrecedesSync Hub Precedes(1)
- Spectral
ex:spectral
trainsVariantTrains Variant(1)
- Current Architecture
ex:current-architecture
wouldBeFasterThanWould Be Faster Than(1)
- Grouped1
ex:grouped1
wouldHelpWould Help(1)
- Compile
ex:compile
wouldHelpPerformanceOfWould Help Performance of(1)
- Mx Compile
ex:mx-compile
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.
| Predicate | Value | Ref |
|---|---|---|
| Rdf:type | Model Architecture | [11] |
| Rdf:type | Attention Configuration | [12] |
| Rdf:type | Model Architecture | [13] |
| Rdf:type | Model Kind | [14] |
| Rdf:type | Model | [14] |
| Dominates Outputs With | Comma | [14] |
| Dominates Outputs With | Period | [14] |
| Dominates Outputs With | Dash | [14] |
| Dominates Outputs With | Parentheses | [14] |
| Computes | Kan Spline Evaluation | [4] |
| Computes | O Nm Causal Cumsum | [4] |
| Learned | Punctuation Commonness Claim | [14] |
| Learned | punctuation is common | [14] |
| Has Linear Time Training | Training | [1] |
| Recovers | Near Quadratic Kan Quality | [1] |
| Has Constant Size Decode Cache | Decode | [1] |
| Enables Publication | Research Thesis | [1] |
| Competitive Quality | Kan | [1] |
| Superior Efficiency | Kan | [1] |
| At Similar Scale | Their Model | [2] |
| Demonstrates Compile Benefits | Throughput | [2] |
| Exists Prior Run | Anchor Kan 65k Run | [3] |
| Has Tok S | 6000 | [4] |
| Has Sync Hub | Block 9 | [4] |
| Has R Hub | 0.561 | [4] |
| Has Slow Tok S | 6000 | [4] |
| Involves Heads | 12 | [4] |
| Involves Dimension | 768 | [4] |
| Involves Anchors | 32 | [4] |
| Has DC Mode | 0.316 | [4] |
| Optimizes Throughput | Parameter Budget | [5] |
| Uses Sinusoidal Fallback With Rope | true | [5] |
| Has Fewer Params | 24% fewer | [5] |
| Has19 Percent Speed Advantage | Spectral | [5] |
| Has Higher Early Variance | true | [5] |
| Is Sparser | Spectral | [5] |
| Is Faster | Spectral | [5] |
| Has Dominant H0 in L1 | true | [5] |
| Represents Same Signal | Spectral | [5] |
| Has Uniform Distribution Across Blocks | R Values | [5] |
| Has Distributed Representation | true | [5] |
| Has Bottleneck | 6K tok/s | [6] |
| Exists As Model | true | [6] |
| Provides Better Long Range Coupling Than | Spectralattention | [7] |
| Aligns With Native Physics | Antenna | [8] |
| Avoids Hard Discretization | Implementation | [8] |
| Enables Stable Attractors | Without Discretization | [8] |
| Fits Core Issue | Transport Works Identity Does Not | [8] |
| Is Key Conceptual Answer | Current Problem | [8] |
| Is Much Closer to | Antenna Native Physics | [8] |
| Provides Very Strong Fit | Current Problem | [8] |
| Should Now Be Fired | Next Action | [8] |
| Suggests Way to Add | Stable Finite Identity Attractors | [8] |
| Efficient to Train | way smaller faster to train | [9] |
| Smaller Than | Comparable Models | [9] |
| Was Solid | pretty solid | [9] |
| References Known Model | Anchor Kan | [9] |
| Scales Well | Chatgpt | [9] |
| Faster to Train Than | Comparable Models | [9] |
| Shows No Negative Things | didn't show any really negative things | [9] |
| Is Slower by Large Margin | Grouped1 | [10] |
| Has Sync Hub Location | Block 9 | [12] |
| Has Sync Strength | 0.561 | [12] |
| Has Baseline Throughput | 6000 | [12] |
| Has Throughput Unit | Tok S | [12] |
| Has Performance Characteristic | Slow Speed | [12] |
| Has Slowness Reason | Causal Cumsum Operation | [12] |
| Outputs Style | Punctuation Loaded Style | [14] |
| Outputs Sample Text | Anchorkan Output Sample | [14] |
| Has Learning Gap | Punctuation Usage Knowledge Gap | [14] |
| Dominates Every Output | true | [14] |
| Appears in | Code Block | [14] |
| Outputs Sample | Anchorkan Output Sample | [14] |
| Follows | Spectral | [14] |
| Output Characteristic | heavily punctuation-loaded | [14] |
| Punctuation Types | commas, periods, dashes, parentheses | [14] |
| Example Output | the, – for well on the to | [14] |
| Has Deficit | hasn'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.
References (14)
ctx:discord/blah/watt-activation/part-71ctx:discord/blah/watt-activation/part-105ctx:discord/blah/watt-activation/part-153ctx:discord/blah/watt-activation/part-221ctx:discord/blah/watt-activation/part-236ctx:discord/blah/watt-activation/part-223ctx:discord/blah/watt-activation/part-313ctx:discord/blah/watt-activation/part-371ctx:discord/blah/watt-activation/part-381ctx:discord/blah/watt-activation/part-103ctx:discord/blah/watt-activation/71- full textwatt-activation-71text/plain2 KB
doc:agent/watt-activation-71/82bde084-e631-42b5-99fe-0a0a8898ac2bShow 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…
ctx:discord/blah/watt-activation/220- full textwatt-activation-220text/plain3 KB
doc:agent/watt-activation-220/5c7f4a28-90e7-46de-ae1e-9e19a58c8d65Show excerpt
[2026-03-11 04:42] xenonfun: FFN DFT — much richer specialization than spectral: ``` ┌─────┬───────┬────────┬────────────────┬────────────────┐ │ blk │ r │ FFN DC │ dominant mode │ pattern │ ├─────┼───────┼────────┼───────…
ctx:discord/blah/watt-activation/222- full textwatt-activation-222text/plain3 KB
doc:agent/watt-activation-222/d8201f0f-b5d1-4b50-9f4e-2aca2c0d4c1eShow 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…
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 voctext/plain680 B
discord:msg/a3126764-fdd1-42f8-9653-a5170ea5bdefShow 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
- Training
- Near Quadratic Kan Quality
- Decode
- Research Thesis
- Kan
- Their Model
- Throughput
- Anchor Kan 65k Run
- Block 9
- Kan Spline Evaluation
- O Nm Causal Cumsum
- Parameter Budget
- Spectral
- R Values
- Spectralattention
- Antenna
- Implementation
- Without Discretization
- Transport Works Identity Does Not
- Current Problem
- Antenna Native Physics
- Next Action
- Stable Finite Identity Attractors
- Comparable Models
- Chatgpt
- Grouped1
- Model Architecture
- Attention Configuration
- Tok S
- Slow Speed
- Causal Cumsum Operation
- Model Kind
- Punctuation Loaded Style
- Anchorkan Output Sample
- Comma
- Period
- Dash
- Parentheses
- Punctuation Commonness Claim
- Punctuation Usage Knowledge Gap
- Code Block
- Model
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.