Dontopedia

Problem 1

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

Problem 1 has 14 facts recorded in Dontopedia across 8 references, with 1 live disagreement.

14 facts·12 predicates·8 sources·1 in dispute

Mostly:causes ineffectiveness(1), describes lr reduction(1), is gradient normalization killing magnitude(1)

Maturity scale raw canonical shape-checked rule-derived certified

Inbound mentions (2)

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.

foundProblemFound Problem(1)

hasProblemHas Problem(1)

Other facts (12)

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.

12 facts
PredicateValueRef
Causes IneffectivenessLr Reduction[1]
Describes Lr ReductionNot Sticking[1]
Is Gradient Normalization Killing MagnitudeLohe Optimizer[2]
Caused by Averaging All Modality Step Timesnull[3]
Fixed by Using Text Only Step Timesnull[3]
Involves Tok S Inflatednull[3]
Leads to Worse BehaviorNear Zero Vectors[4]
Involves Mx Maximum Vs Additive EpsDifferent Numerical Behavior[4]
Rdf:typeTechnical Issue[5]
Has Descriptioncoupling, log_adjacency, harmonic_coeffs are never updated[6]
Involves Gradient Normalizationgrad = mx.where(g_norm > 1.0, grad / g_norm, grad)[7]
Located at Line236[7]

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.

14 undated
—
causesIneffectivenessblah/watt-activation/part-39
ex:lr-reduction
—
describesLrReductionblah/watt-activation/part-39
ex:not-sticking
—
isGradientNormalizationKillingMagnitudeblah/watt-activation/part-192
ex:lohe-optimizer
—
causedByAveragingAllModalityStepTimesblah/watt-activation/part-244
null
—
fixedByUsingTextOnlyStepTimesblah/watt-activation/part-244
null
—
involvesTokSInflatedblah/watt-activation/part-244
null
—
leadsToWorseBehaviorblah/watt-activation/part-276
ex:near-zero-vectors
—
involvesMxMaximumVsAdditiveEpsblah/watt-activation/part-276
ex:different-numerical-behavior
—
typeblah/watt-activation/26
ex:TechnicalIssue
—
labelblah/watt-activation/188
Problem 1
—
hasDescriptionblah/watt-activation/188
coupling, log_adjacency, harmonic_coeffs are never updated
—
involvesGradientNormalizationblah/watt-activation/192
grad = mx.where(g_norm > 1.0, grad / g_norm, grad)
—
locatedAtLineblah/watt-activation/192
236
—
labelblah/watt-activation/274
Problem 1: mx.maximum vs + eps

References (8)

8 references
  1. [1]Part 392 facts
    ctx:discord/blah/watt-activation/part-39
  2. [2]Part 1921 fact
    ctx:discord/blah/watt-activation/part-192
  3. [3]Part 2443 facts
    ctx:discord/blah/watt-activation/part-244
  4. [4]Part 2762 facts
    ctx:discord/blah/watt-activation/part-276
  5. [5]261 fact
    ctx:discord/blah/watt-activation/26
    • full textwatt-activation-26
      text/plain3 KBdoc:agent/watt-activation-26/33435d36-2781-4f3b-ab35-e0cdf2197a23
      Show excerpt
      [2026-03-06 16:23] xenonfun: On making conversational: ``` Right — no EOS token means the model just generates forever until you hit max_new_tokens. A few paths to fix this, in order of effort: 1. Heuristic stop (zero retraining, works …
  6. [6]1882 facts
    ctx:discord/blah/watt-activation/188
    • full textwatt-activation-188
      text/plain3 KBdoc:agent/watt-activation-188/0b24c5f9-ca6d-47b7-9d97-98b6fac36e0c
      Show excerpt
      [2026-03-10 03:16] xenonfun: well I imagine data from working RotAdamW will be informative for it as to how to correct behavior / step issues in LoheOptimizer [2026-03-10 03:17] xenonfun: also that will be recorded [2026-03-10 03:38] xenonf…
  7. [7]1922 facts
    ctx:discord/blah/watt-activation/192
    • full textwatt-activation-192
      text/plain3 KBdoc:agent/watt-activation-192/c3ca4e62-2524-47f4-9e3e-6cb28f08f78c
      Show excerpt
      [2026-03-10 04:10] xenonfun: ⏺ Working correctly now. Full comparison: ```┌──────────────────────┬────────────────────┬─────────────────────────┐ │ │ RotAdamW fine-tune │ LoheOptimizer fine-tune │ ├─────────────────…
  8. [8]2741 fact
    ctx:discord/blah/watt-activation/274
    • full textwatt-activation-274
      text/plain1 KBdoc:agent/watt-activation-274/2c5f3f8e-4dd2-4aea-8be5-e9d163af5028
      Show excerpt
      [2026-03-13 20:06] xenonfun: It tried its fuckery again but we learned an antipattern to remember. ```⏺ Here's the difference: _normalize (attention.py:136-141): norm = mx.sqrt(mx.maximum(sum(x²), eps)) # eps=1e-6 for float32 x / no…

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.

↑