From Dontopedia, the open, paraconsistent wiki. (Last updated 2026-06-06.)
Rotational Adam W has 33 facts recorded in Dontopedia across 8 references, with 4 live disagreements.
Mostly:rdf:type(8), rdfs:label(5), preserves(2)
rdfs:labelpreserveshasWeightDecayequivalentTousesUpdateTypereliesOnisGeometryRespectinghasKeyDistinctionFromupdatesViaoperatesOnisDescribedAsOther 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.
hasOptimizerHas Optimizer(2)ex:anchor-kan-rotadam-v3-configex:rotational-adam-plus-anchor-configcontainsCustomOptimizerContains Custom Optimizer(1)ex:harmonic-gpt-repodiscussedExceptionDiscussed Exception(1)ex:xenonfunextendsExtends(1)ex:KuramotoRotationalAdamWinvolvesOptimizerInvolves Optimizer(1)ex:step-pretrain-rotadamwinvolvesTechniqueInvolves Technique(1)ex:fixing-optimizerzeroedOutByZeroed Out by(1)ex:adjacencyThe 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 |
|---|---|---|
| Has Parameter Zero Mean | true | [5] |
| Has Learning Rate | 0.0001 | [5] |
| Lacks Feature | captured-forward | [6] |
| Acts As | Implicit Regularization | [1] |
| Has Norm Preservation | true | [1] |
| Has Initialization Parameter | zero_mean | [4] |
| Has Feature | No Weight Decay Needed | [3] |
| Performs Angular Updates | Per Neuron Sphere | [3] |
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.
doc:agent/watt-activation-160/83f1326e-5f35-47f2-901e-bb1cc61a1eae[2026-03-09 16:37] xenonfun: ⏺ With dropout=0.0, the dropout layers are no-ops — they pass through unchanged during both training and inference. So model.train(False) vs model.train(True) makes no difference for our current config. It …
doc:agent/watt-activation-212/6835fc9f-e8f3-4cfe-b6ab-3f16b5dbc7d2[2026-03-11 04:12] xenonfun: ``` ⏺ The sidecar data is very revealing! Let me respond to the designer message while the run finishes. --- On Omega's optimizer question: RotationalAdamW is exactly the geometry-aware rotation optimizer d…
doc:agent/watt-activation-117/93da7c3e-762b-4710-96fb-cdcbffaa6ad1[2026-03-08 23:25] xenonfun: okay it still is using Adam finally pointed it at your optimizers: ``` ⏺ This is really interesting. There are 4 custom optimizers in harmonic-gpt/harmonic_gpt/optim/: 1. RotationalAdamW — Angular updates on …
doc:agent/watt-activation-138/9bd08756-4bea-4055-bcfd-5d3ed62434a1[2026-03-09 06:23] xenonfun: ``` ⏺ No — softmax is on self.log_adjacency which is a static (G, G) parameter tensor, completely independent of T. It runs once per forward pass in O(G²) = O(64). The sequence-length work is entirely in _gate…
doc:agent/watt-activation-204/c8f804b0-ce31-4b4d-9ff6-960a682fa16d[2026-03-10 07:26] xenonfun: those are all with `opt = RotationalAdamW(model, lr=1e-4, weight_decay=0.01, zero_mean=True)` optimizer. ``` ⏺ Right — the whole point of Form 2 was to give LoheOptimizer a real order parameter instead of the …
doc:agent/watt-activation-181/a38e89b1-8c19-4e64-aa47-644f1d1e399d[2026-03-10 01:48] xenonfun: ``` Performance hit ┌────────────────────────────────────────────────┬────────────────────────────────┐ │ Step │ Cost │ ├─────────────…
doc:agent/watt-activation-214/6ebedb80-7ad3-4d5e-8432-4c87da7e9ad5[2026-03-11 04:17] xenonfun: ``` Model: lohe_spherical+lohe_v3 768d 12L 12H vocab=100277 Steps: 1000 BS=4 SEQ=1024 (4.1M tokens/run) Phase metrics sidecar → logs/phase_metrics_lohe_spherical_20260311_001319.jsonl ====================…
doc:agent/watt-activation-230/e16b9c93-10f7-44c7-8fd1-827e0be1da24[2026-03-11 07:15] xenonfun: <@1211062099137265723> ⏺ anchor_r is the Kuramoto order parameter for the anchor oscillators — the mean-field coherence of the m anchor points on the unit circle in each attention head. The anchors are com…
Dontopedia is in a read-only public launch. Follow the references and disputed branches now; contributions will open after durable identity and moderation are in place.