LoheFFNv3
From Dontopedia, the open, paraconsistent wiki. (Last updated 2026-06-06.)
LoheFFNv3 has 33 facts recorded in Dontopedia across 10 references, with 4 live disagreements.
Mostly:rdf:type(5), computes feature(4), implementation change(3)
Maturity scale
raw canonical shape-checked rule-derived certifiedInbound mentions (7)
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.
connectsConnects(1)
- Unifying Structure
ex:unifying-structure
hasComponentHas Component(1)
- Harmonic Block
ex:harmonic-block
modifiesModifies(1)
- Clifford Readout
ex:clifford-readout
mostlyRedundantWithMostly Redundant With(1)
- Grade 0 Features
ex:grade-0-features
portsPorts(1)
- Direction Mlx Faithful
ex:direction-mlx-faithful
sameDecompositionAsSame Decomposition As(1)
- Depth Wise Dft
ex:depth-wise-dft
usesUses(1)
- Harmonic Block X2 Frontend
ex:harmonic-block-x2-frontend
Other facts (27)
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 | Synchronization Method | [4] |
| Rdf:type | Operation | [6] |
| Rdf:type | Feed Forward Network | [7] |
| Rdf:type | Architecture | [8] |
| Rdf:type | Model Version | [9] |
| Computes Feature | Raw Synced Spectra | [9] |
| Computes Feature | Dft Ring Eigenmodes | [9] |
| Computes Feature | R Global | [9] |
| Computes Feature | Jacobi Polynomials | [9] |
| Implementation Change | Add K Probes Param | [9] |
| Implementation Change | Add Probe Weights | [9] |
| Implementation Change | Append Features | [9] |
| Refers to Architecture | Intra Block Ring Sync | [1] |
| Syncs | Ring | [2] |
| Has Ring Graph Design | True | [3] |
| Classification | sparse graph | [4] |
| Uses Decomposition Within Block | Dft Modes | [5] |
| Operates Within | Each Block | [5] |
| Has Duration | 1.5 | [6] |
| Has Percentage of Block | 8 | [6] |
| Is More Parameter Efficient Than | Standard Transformer Ffn | [7] |
| Has Dynamics | V3 Dynamics | [9] |
| Estimated Code Change Size | ~10 lines of change | [9] |
| Parameter Count Increase | 256 | [9] |
| Latency Increase | ~0.2ms extra | [9] |
| Has Feature | Clifford Readout | [10] |
| Flag | Clifford Readout Flag | [10] |
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 (10)
ctx:discord/blah/watt-activation/part-226ctx:discord/blah/watt-activation/part-305ctx:discord/blah/watt-activation/part-313ctx:discord/blah/watt-activation/225- full textwatt-activation-225text/plain2 KB
doc:agent/watt-activation-225/9b15ed6a-4fd9-4458-9e17-1e0372b5a3e0Show excerpt
[2026-03-11 05:15] xenonfun: ⏺ Ablation complete. The result is more interesting than "hub wins": blk acc lift note 0 0.575 +0.217 ← input (BEST) 4 0.567 +0.208 3 0.558 +0.200 7 0.550 +0.192 ... 9…
ctx:discord/blah/watt-activation/228- full textwatt-activation-228text/plain3 KB
doc:agent/watt-activation-228/218c4ece-61aa-47aa-a16e-b77f7f994d5fShow excerpt
[2026-03-11 05:23] xenonfun: ⏺ It is, and the math backs it up cleanly. What the network discovered spontaneously is essentially a depth-wise DFT — the same decomposition that LoheFFNv3 uses within each block (ring → DFT modes), but now o…
ctx:discord/blah/watt-activation/291- full textwatt-activation-291text/plain3 KB
doc:agent/watt-activation-291/883c968e-0b03-4e27-aca9-027dfc155696Show excerpt
[2026-03-14 03:41] xenonfun: Why Keeping the Spherical Head Still Makes Sense (performace at 600K parm scale is effectively same as euclidian head) ``` Even if performance is the same, the spherical head is still the better design. Reasons…
ctx:discord/blah/watt-activation/301- full textwatt-activation-301text/plain3 KB
doc:agent/watt-activation-301/3e98420c-16db-4c57-9cad-a987cec9976bShow excerpt
[2026-03-14 06:51] xenonfun: ``` ⏺ Lohe vs Softmax AR Decoder Comparison ┌───────────────┬───────────────┬────────────┐ │ Metric │ Softmax d=384 │ Lohe d=384 │ ├───────────────┼───────────────┼────────────┤ │ BPB │…
ctx:discord/blah/watt-activation/318- full textwatt-activation-318text/plain3 KB
doc:agent/watt-activation-318/f52d95a8-f461-40d1-9360-f08558b18eb1Show excerpt
[2026-03-15 02:47] xenonfun: ⏺ I see you're working on wire encoding / phase modulation — that's a fascinating direction. Let me check what you've got: [2026-03-15 02:47] lisamegawatts: Wire QPSK + Standard: PPL 4.94, Byte Accuracy 51.5% T…
ctx:discord/blah/watt-activation/380- full textwatt-activation-380text/plain3 KB
doc:agent/watt-activation-380/84a046b5-55b5-4f12-a277-387a85fc8228Show excerpt
[2026-03-19 00:07] xenonfun: ``` ❯ would we want to make a LoheFFNv4 that just replaces that readout or does that add more to it? ⏺ Just the readout. Here's why: v3 dynamics are sound — ring sync, DFT modes, Lohe tangent step all work. …
ctx:discord/blah/watt-activation/403- full textwatt-activation-403text/plain2 KB
doc:agent/watt-activation-403/d55f91e3-d6ae-47d0-a701-fcf659140c12Show excerpt
[2026-03-19 05:31] xenonfun: ❯ is any of this applicable to WireLM lohe_v3 style stuff? ``` ⏺ Yes — the decay gates specifically. The gated cumsum monoid (g, v) ⊗ (g', v') = (g·g', g'·v + v') is exactly what _gated_cumsum already does in at…
See also
- Intra Block Ring Sync
- Ring
- True
- Synchronization Method
- Dft Modes
- Each Block
- Operation
- Feed Forward Network
- Standard Transformer Ffn
- Architecture
- Model Version
- V3 Dynamics
- Raw Synced Spectra
- Dft Ring Eigenmodes
- R Global
- Jacobi Polynomials
- Add K Probes Param
- Add Probe Weights
- Append Features
- Clifford Readout
- Clifford Readout Flag
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.