Dontopedia

AR decoder

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

AR decoder has 41 facts recorded in Dontopedia across 7 references, with 3 live disagreements.

41 facts·25 predicates·7 sources·3 in dispute

Mostly:produces(14), conditions byte on(2), must be rewritten using(2)

Maturity scale raw canonical shape-checked rule-derived certified

Producesin disputeproduces

Inbound mentions (5)

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.

causedByCaused by(1)

describesAsMassiveImprovementDescribes As Massive Improvement(1)

introducedByIntroduced by(1)

partOfPart of(1)

shouldNotAppearShould Not Appear(1)

Other facts (26)

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.

26 facts
PredicateValueRef
Conditions Byte onCode[2]
Conditions Byte onPrevious Bytes[2]
Must Be Rewritten UsingLohe Spherical Attention Blocks[4]
Must Be Rewritten UsingLohesphericalattention Blocks[7]
Enables Reconstructionnull[1]
Superior toBaseline Single Step Decoder[1]
Has Bpb3.08[1]
Improves Bpb by1.87[1]
Compared toBaseline Single Step Decoder[1]
Reconstructs More Info Percent38[1]
Achieves Lower BpbBaseline Single Step Decoder[1]
Conditioned on Codenull[2]
Does Not ProduceComplete Sentences[2]
Enables Word Fragmentsnull[2]
Improves OverSingle Step Decoder[2]
Produces Lots ofWhitespace[2]
Produces RecognizableEnglish Word Fragments[2]
Is Autoregressive{}[3]
Is Only ComponentBreaking Manifold Assumption[4]
Requires RewriteLohe Spherical Attention[4]
Introduces InconsistencyInconsistent Optimization Geometry[4]
Deontically Must UseLohe Blocks[4]
Rdf:typeModel Component[5]
Produces OutputEnglish Word Fragments[6]
Prediction ModeConditional Byte Prediction[6]
RealizesLexical Content[6]

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.

41 undated
—
enablesReconstructionblah/watt-activation/part-299
null
—
superiorToblah/watt-activation/part-299
ex:baseline-single-step-decoder
—
hasBpbblah/watt-activation/part-299
3.08
—
improvesBpbByblah/watt-activation/part-299
1.87
—
comparedToblah/watt-activation/part-299
ex:baseline-single-step-decoder
—
reconstructsMoreInfoPercentblah/watt-activation/part-299
38
—
achievesLowerBpbblah/watt-activation/part-299
ex:baseline-single-step-decoder
—
producesblah/watt-activation/part-300
ex:word-this
—
conditionedOnCodeblah/watt-activation/part-300
null
—
conditionsByteOnblah/watt-activation/part-300
ex:code
—
conditionsByteOnblah/watt-activation/part-300
ex:previous-bytes
—
doesNotProduceblah/watt-activation/part-300
ex:complete-sentences
—
enablesWordFragmentsblah/watt-activation/part-300
null
—
improvesOverblah/watt-activation/part-300
ex:single-step-decoder
—
producesblah/watt-activation/part-300
ex:word-earch
—
producesblah/watt-activation/part-300
ex:word-for
—
producesblah/watt-activation/part-300
ex:word-fragments
—
producesblah/watt-activation/part-300
ex:word-imported
—
producesblah/watt-activation/part-300
ex:word-one
—
producesblah/watt-activation/part-300
ex:word-project
—
producesblah/watt-activation/part-300
ex:word-protec-tion
—
producesblah/watt-activation/part-300
ex:word-state
—
producesblah/watt-activation/part-300
ex:word-the
—
producesblah/watt-activation/part-300
ex:word-with
—
producesLotsOfblah/watt-activation/part-300
ex:whitespace
—
producesRecognizableblah/watt-activation/part-300
ex:english-word-fragments
—
producesblah/watt-activation/part-300
ex:hex-like-artifacts
—
producesblah/watt-activation/part-300
ex:word-and
—
producesblah/watt-activation/part-300
ex:word-context
—
isAutoregressiveblah/watt-activation/part-301
{}
—
isOnlyComponentblah/watt-activation/part-302
ex:breaking-manifold-assumption
—
mustBeRewrittenUsingblah/watt-activation/part-302
ex:lohe-spherical-attention-blocks
—
requiresRewriteblah/watt-activation/part-302
ex:lohe-spherical-attention
—
introducesInconsistencyblah/watt-activation/part-302
ex:inconsistent-optimization-geometry
—
deonticallyMustUseblah/watt-activation/part-302
ex:lohe-blocks
—
typeblah/watt-activation/297
ex:ModelComponent
—
labelblah/watt-activation/298
AR decoder
—
producesOutputblah/watt-activation/298
ex:english-word-fragments
—
predictionModeblah/watt-activation/298
ex:conditional-byte-prediction
—
realizesblah/watt-activation/298
ex:lexical-content
—
mustBeRewrittenUsingblah/watt-activation/300
ex:lohesphericalattention-blocks

References (7)

7 references
  1. [1]Part 2997 facts
    ctx:discord/blah/watt-activation/part-299
  2. [2]Part 30022 facts
    ctx:discord/blah/watt-activation/part-300
  3. [3]Part 3011 fact
    ctx:discord/blah/watt-activation/part-301
  4. [4]Part 3025 facts
    ctx:discord/blah/watt-activation/part-302
  5. [5]2971 fact
    ctx:discord/blah/watt-activation/297
    • full textwatt-activation-297
      text/plain2 KBdoc:agent/watt-activation-297/ad91f718-f038-464f-a6d2-ba91d77fe4e3
      Show excerpt
      [2026-03-14 05:23] xenonfun: 600K context, the UI crazy scroll thing and memory leaks are annoyin claude is sucking up 6GB now. [2026-03-14 05:24] xenonfun: ``` ⏺ Launched in rjs:longgen. This trains with longer context (512) and more steps…
  6. [6]2984 facts
    ctx:discord/blah/watt-activation/298
    • full textwatt-activation-298
      text/plain2 KBdoc:agent/watt-activation-298/f5cde311-fd9a-43e7-a746-9177b5a91fee
      Show excerpt
      [2026-03-14 05:53] xenonfun: ``` What Changed The AR decoder produces recognizable English word fragments: "the", "and", "for", "with", "this", "one", "protec(tion)", "earch", "context", "project", "state", "imported". These are real m…
  7. [7]3001 fact
    ctx:discord/blah/watt-activation/300
    • full textwatt-activation-300
      text/plain3 KBdoc:agent/watt-activation-300/3b6edccf-3524-4608-838f-25890efaea15
      Show excerpt
      [2026-03-14 06:34] xenonfun: ``` 3. Manual attention (lines 110-128) — Hand-rolled softmax attention instead of using mx.fast.scaled_dot_product_attention. MLX's fused attention kernel is significantly faster for small sequence lengths. …

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.

↑