Pith. sign in

REVIEW 4 major objections 5 minor 3 cited by

Adaptive Domain Models: Bayesian Evolution, Warm Rotation, and Principled Training for Geometric and Neuromorphic AI

T0 review · 4 major / 5 minor · reviewed 2026-07-13 · grok-4.5

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Training memory, grade corruption, and optimizer bloat are substrate artifacts of reverse-mode IEEE-754, not of learning itself; a type-level geometric substrate can train at ~2× inference memory with certified structure.

desk verdict A coherent design paper that packages grade-preserving forward-mode training, warm rotation, and Bayesian distillation—but the load-bearing theorem is a sketch on unshipped companion invariants, with no experiments. read the letter →

arxiv 2603.18104 v5 pith:NOPV3HN2 submitted 2026-03-18 cs.AI cs.DCcs.LGcs.NE

Houston Haynes This is my paper · ORCID
classification cs.AIcs.DCcs.LGcs.NE
keywords adaptivedomainmodelsforward-modeautodiffgeometricalgebragradepreservationprogramhypergraphwarmrotationBayesiandistillationneuromorphicSTDP
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper argues that the familiar costs of deep training—activation-tape memory that scales with depth, optimizer state that doubles parameters, and the silent destruction of geometric algebra structure—are consequences of reverse-mode autodiff over IEEE-754 arithmetic, not of gradient-based learning. It composes three prior results (dimensional types with stack-eligible gradients and exact quire accumulation, a program hypergraph that treats grade as a type invariant, and bounded-regime posit arithmetic) into a single training regime. Under that regime, a weight of declared grade keeps a same-grade gradient by construction, so trained Clifford networks retain the sparsity and equivariance they were designed for; training memory stays roughly twice the inference footprint regardless of depth. The same coeffect signature covers both loss-function networks and spike-timing plasticity, so geometric and neuromorphic models share one verification and deployment stack. Two operational pieces complete the picture: Bayesian distillation, which filters a general model’s latent prior through the type system to seed domain models when data are scarce, and warm rotation, which swaps a certified updated model into a live inference path without interrupting service. The intended outcome is smaller, continuously adaptive domain systems that are verifiably consistent with the physics of their domains and can be bootstrapped from existing large models.

What carries the argument

Adaptive Domain Model (ADM) training substrate: PHG grade inference plus dual-number forward gradients with exact quire accumulation, so that a grade-k weight has a grade-k gradient by the chain-rule closure of grade-preserving operations and any other grade is rejected at elaboration time.

What would settle it

Train a Clifford (or hybrid neuromorphic) network end-to-end under the claimed substrate and check two concrete outcomes: (1) every weight remains pure grade after many steps (Cayley sparsity and rotor norms identical to initialization), and (2) peak training memory stays within a small constant of twice the inference footprint independent of depth; failure of either measurement falsifies the central claim.

Watch

Extended reading notes

Core claim

Composing program-hypergraph grade inference, dual-number forward-mode autodiff, and quire accumulation turns grade preservation through training into a design-time theorem, bounds training memory to about twice the inference footprint independent of depth, and thereby supports certified warm rotation and Bayesian distillation for domain-specific geometric and neuromorphic models.

Load-bearing premise

That latent Bayesian structure extractable from general language models will, after dimensional and grade filtering, become a well-calibrated prior on continuous multi-physics weight spaces—an assumption the paper itself places outside the cited empirical regime and lists as open work—while the whole architecture also assumes the concurrent type-system and hypergraph results deliver their claimed coeffect and grade invariants in a working compiler.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes Adaptive Domain Models (ADM): a training and deployment architecture that composes three prior results—DTS/DMM forward-mode autodiff with stack-eligible gradients and quire accumulation [8], Program Hypergraph grade inference [10], and b-posit bounded-regime arithmetic [12]—to claim depth-independent training memory of roughly twice the inference footprint, grade-preserving weight updates for Clifford networks, and a unified coeffect signature for STDP neuromorphic learning. It introduces Bayesian distillation (extracting latent priors from general LLMs, then filtering them through DTS/PHG constraints) and warm rotation (atomic, certificate-gated model exchange without service interruption, with signed version records). The central technical claim is Proposition 5.1: under PHG-typed dual-number forward-mode training, grade-k weights retain grade k through updates as a design-time certificate, with sparsity stability and exact rotor equivariance as corollaries.

Significance. If the composition holds in a working compiler and on hardware, the contribution would be substantial for geometric and physics-structured learning: grade corruption under IEEE-754 is a real obstacle to Clifford networks, and a substrate that makes grade preservation a type-level theorem rather than a post-hoc regularizer would change how equivariant geometric models are trained. The memory-parity argument for forward-mode training, if realized, would also make continuous on-device adaptation more plausible on inference-class accelerators. Warm rotation is a clear operational specification that practitioners can evaluate independently of the arithmetic substrate. Credit is due for an honest scope statement (design intent, implementation in progress), for marking the LLM-to-physical-domain transfer gap in §3.4 and §9.1, and for connecting STDP and forward-mode under a shared local-learning coeffect signature. The manuscript does not ship machine-checked proofs, trained models, or measured memory/accuracy results; significance therefore remains conditional on the companion frameworks and on empirical validation that is not yet present.

major comments (4)
  1. [§5.1, Proposition 5.1] §5.1, Proposition 5.1: The grade-invariance claim is load-bearing for the paper’s strongest result (certified warm rotation via PHG certificates, Def. 4.1 condition 2). The proof sketch correctly notes dual-number grade matching and grade-restricted projections, but it does not close the argument under the multi-tangent Gram projection of §2.4, quire-to-posit conversion, or the finite-precision update W′=W−η∇_W L in b-posit. Off-grade components reintroduced by rounding or reconstruction would falsify the design-time certificate. Either complete the reconstruction/rounding argument in this manuscript or state Prop. 5.1 as conditional on fully discharged invariants from [8,10] with an explicit residual risk.
  2. [§1.2, §2.3–2.4, §5.1] Dependence on concurrent self-citations [8] and [10]: Stack-eligible O(1)/O(k) coeffects, quire lifetime analysis, PHG grade inference, and SMT-LIB2 discharge of rotor norms are imported as established, yet this manuscript provides neither machine-checked artifacts nor independent restatements sufficient to audit those claims. For a journal contribution whose memory-parity (~2× inference) and grade-preservation theorems rest on those results, the paper must either (a) include self-contained formal statements and proof obligations that a reader can check without the companions, or (b) defer the theorem-level claims until the companions are available with verifiable artifacts, and reframe the present work as an architectural specification.
  3. [Abstract; §4.2; Fig. 2; §5.1 Corollary 5.1] Absence of empirical or systems evidence: The abstract and §4 claim depth-independent training memory, productive spare-TOPS training (Fig. 2), stable 85–95% Cayley sparsity after training, and continuous adaptation. There are no trained models, memory footprints, accuracy curves, KL-trigger traces, or certificate-diff examples. For an architecture paper whose central selling points are quantitative (2× memory, grade stability, warm rotation without interruption), at least one end-to-end case study—even a small PGA network or a simulated rotation on a fixed hardware budget—is needed to show that the design intent is realizable and that Prop. 5.1 survives implementation.
  4. [§3.4; §9.1; Contribution 3] §3.4 Bayesian distillation: The paper claims the mechanism “resolves the data-scarcity bootstrapping problem,” yet §3.4 and §9.1 correctly note that van Steenkiste & Linzen [26] evaluate sequential recommendation, not continuous multi-physics grade-structured weight spaces, and that transfer is open. The type-to-support construction (W_adm = W_dim ∩ W_grade, projector Π_adm) is a useful formalization, but calling the problem resolved overstates the evidence. Soften the claim to a proposed initialization path contingent on the open transfer questions, and separate what is formal (filter support) from what is empirical (calibration of the distilled prior).
minor comments (5)
  1. [§4 opening; Def. 4.1] Present-tense system behavior in §4 (“the active model actor processes…”) is later qualified as design intent; a consistent “specification / intended behavior” framing would reduce reader confusion.
  2. [§5.1] Typo in Prop. 5.1: “as a a design-time certificate.”
  3. [Fig. 1; Fig. 2] Fig. 1 and Fig. 2 are conceptual only; adding a small worked numerical example (e.g., one layer’s grade-typed dual update) would help readers who are not already steeped in the companion papers.
  4. [§2.4; §3.2; §4.2] Free parameters ε_domain, k (tangents), r (Hessian rank), and b-posit eS/rS are introduced without recommended defaults or sensitivity discussion; a short table of suggested ranges would aid reproducibility of any future experiments.
  5. [Software Availability] Software Availability points to repositories “under active development”; stating which claims are backed by public code versus specification-only would set expectations clearly.

Circularity Check

3 steps flagged · score 4.0 of 10

Prop. 5.1 grade invariance is largely by construction from concurrent self-cited DTS/PHG invariants; warm rotation and Bayesian distillation remain independent architectural proposals.

  1. self citation load bearing [Sec. 1.2, 1.4; Prop. 5.1 (Sec. 5.1); Abstract]
    "This paper is the third in a sequence. The DTS/DMM paper [8] established the foundational type system, memory management discipline, and forward-mode autodiff analysis. The PHG paper [10] extended this to multi-way geometric constraints, grade-typed Clifford algebra computation... This paper takes the PHG's grade preservation properties and the DTS/DMM's forward-mode coeffect analysis as established results and develops their implications for a training architecture"

    The design-time theorem that training preserves grade, the stack-eligible O(1)/O(k) coeffect story, and PHG certificates required by warm rotation (Def. 4.1 condition 2) are not established in this manuscript; they are imported wholesale from concurrent same-author papers [8,10] that are not machine-checked here and whose implementation is still 'under active development.' Without those self-cited invariants, Prop. 5.1 and certified rotation have no independent discharge path in this text.

  2. self definitional [Prop. 5.1 proof sketch, Sec. 5.1]
    "The dual-number extension of a grade-k primal is a grade-k tangent. PHG grade inference applies identically to the tangent computation as to the primal... v is drawn from the grade-k subspace of the weight space... The gradient estimate ∇W L ... lies in the grade-k subspace by construction because the estimator is a linear combination of grade-k projection vectors. Any component at grade j̸ = k is a type violation caught at design time."

    Given the premises that duals inherit grade, that projections are typed grade-k, and that off-grade components are type errors, the conclusion that updates stay grade-k is true by the definitions of the type system and of linear combination—not an independent dynamical or arithmetic result. The sketch does not show that quire-to-posit conversion or W' = W − η∇W L under finite b-posit cannot reintroduce off-grade mass; it equates 'type forbids it' with 'training preserves it.'

1 more flagged steps
  1. self definitional [Sec. 3.4, 'From type constraints to weight-space priors']
    "The type system therefore induces a prior of the form p(W)∝1[W∈W_adm]·p̃(W), where p̃(W) is a density over W_adm and the indicator is exact, not a penalty. Writing Π_adm for the orthogonal projector onto W_adm, the filtering stage maps a distillation candidate W_cand ... to Π_adm W_cand"

    The claim that DTS/PHG 'formalize' the LLM prior into a domain-consistent prior is the definition of projecting onto the admissible type support and zeroing the rest. That the filter is exact rather than a penalty is definitional of hard type constraints; it is not an empirical or Bayesian derivation that the resulting object is a well-calibrated physical prior (which Sec. 9.1 correctly leaves open).

full rationale

This paper does not fit free parameters to data and relabel them as predictions, so classical numerical circularity is absent. The load-bearing circularity is structural and moderate: the strongest formal claim (Prop. 5.1 grade invariance through training, plus the depth-independent ~2× memory and certified warm-rotation correctness that rest on it) is obtained by composing the author's concurrent DTS/DMM and PHG results [8,10] and then reading off consequences that follow by the type system's own definitions (grade-k duals, grade-k projection vectors, linear combinations staying in the grade-k subspace, type violations caught at elaboration). The paper itself states that it takes those grade-preservation and coeffect properties as established rather than re-deriving them, and the software stack is still under active development with no machine-checked artifacts in this manuscript. That is self-citation load-bearing plus a self-definitional step, not a full collapse of the paper. Warm rotation (Def. 4.1), Bayesian distillation (Sec. 3.4), the IEEE-754 inurement analysis, the STDP coeffect analogy, and the actor-model framing are genuine architectural content that do not reduce to a fit or a tautology; distillation transfer is explicitly marked open (Sec. 9.1). Forward-mode O(1) memory also has independent external support (Baydin et al.). Proportionate score is therefore 4: central formal theorem is thin and self-cited, but the paper is not equivalent to its inputs by definition.

Assumptions & free parameters 4 free parameters · 7 assumptions · 4 invented entities

The central claim rests on accepting three concurrent/prior frameworks as delivering stack-scoped forward gradients, type-level grade preservation, and practical b-posit hardware, then adding operational axioms about KL-triggered updates, LLM prior extractability, and atomic actor-model rotation. Free parameters are thresholds and tangent counts chosen by domain/design, not fit to a reported dataset. Invented entities are the ADM class and the two named operational patterns; independent empirical handles are not yet provided in this manuscript.

free parameters (4)
  • ε_domain (KL divergence trigger threshold)
    Domain-calibrated threshold that decides when operational evidence triggers a model update and warm rotation; not derived, must be set per domain (Sec. 3.2).
  • k (number of multi-tangents / projection rank)
    Controls forward-gradient approximation quality via k/n; design-time hyperparameter affecting both accuracy and O(k) memory (Sec. 2.4).
  • b-posit eS/rS and asymmetric ring parameterization for ML = eS smaller than general HPC; ≥5-bit floor cited
    Posit exponent/regime choices and near-unity bias are application-specific design parameters drawn from Gustafson/NUS guidance, not fixed by the theory (Sec. 4.2).
  • Hessian retained rank r for posterior approximation = tens to low hundreds, r ≪ d
    Low-rank curvature memory term (2+r/d)× inference footprint; schedule left as design choice (Sec. 4.2).
assumptions (7)
  • domain assumption Forward-mode dual-number autodiff with quire accumulation yields unbiased (multi-)tangent gradients with O(1) or O(k) stack-eligible auxiliary memory per layer and no activation tape.
    Imported from Baydin et al. and author's DTS/DMM paper [8]; foundational to the 2× memory claim.
  • domain assumption PHG grade inference makes multivector grade a type-level invariant closed under geometric product and dual-number tangent extension, so grade-k weights have grade-k gradients.
    Core of Prop. 5.1; depends on companion PHG paper [10], not re-proved here beyond a sketch.
  • domain assumption b-posit bounded regime (rS≤6) makes posit+quire practical on inference-class/edge hardware at cost competitive with float32 decoders.
    Taken from Jonnalagadda et al. 2025 [12]; required for the hardware-feasibility narrative.
  • ad hoc to paper Latent Bayesian structure in general LLMs is accessible and can be filtered into dimensionally/geometrically coherent domain priors.
    Bayesian distillation mechanism (Sec. 3.4); only partially supported by van Steenkiste & Linzen on a non-physical task family.
  • domain assumption KL divergence between predictive and empirical operational distributions is a sufficient, domain-calibrated trigger for verified model updates.
    Sec. 3.2 adaptation trigger; standard Bayesian idea but not validated operationally here.
  • ad hoc to paper Actor-model message semantics (Olivier/Prospero) can implement atomic model exchange with request buffering and dimensional annotations across BAREWire.
    Warm rotation atomicity condition (Def. 4.1); design assumption of the Clef/Fidelity runtime.
  • domain assumption STDP and forward-mode autodiff share a common local-learning coeffect signature sufficient for unified verification and warm-rotation infrastructure.
    Prop. 6.1; structural analogy, not an empirical equivalence of learning performance.
invented entities (4)
  • Adaptive Domain Model (ADM) architecture
    purpose: Name the class of continuously adaptive, type-constrained, domain-scoped models trained under the proposed substrate.
    Umbrella construct for the paper's contributions; no independent empirical existence outside the proposed Fidelity stack.
  • Bayesian distillation
    purpose: Extract unstructured LLM priors and project them onto DTS/PHG-admissible weight supports to bootstrap domain models.
    New named mechanism; filtering via Π_adm is defined, but no measured distillation quality or domain transfer results.
  • Warm rotation
    purpose: Atomic live replacement of verified model weights without service interruption, with PHG certificates and signed version records.
    Operational pattern defined in Def. 4.1; implementation and evaluation not provided.
  • PHG structural certificate / certificate differencing for model versions
    purpose: Human-interpretable, proof-level record of structural changes across model updates.
    Sec. 7.1 states versioning layer is under active development; certificates are design-stage capability.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Adaptive Domain Models: Bayesian Evolution, Warm Rotation, and Principled Training for Geometric and Neuromorphic AI." pith.science (2026). https://pith.science/paper/NOPV3HN2

@misc{pith2026260318104,
  author       = {Pith},
  title        = {Pith review of: Adaptive Domain Models: Bayesian Evolution, Warm Rotation, and Principled Training for Geometric and Neuromorphic AI},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NOPV3HN2}},
  note         = {Machine review of arXiv:2603.18104}
}
read the original abstract

Prevailing AI training assumes reverse-mode automatic differentiation over IEEE-754 arithmetic. The memory overhead of training relative to inference, optimizer complexity, and structural degradation of geometric properties through training are consequences of this arithmetic substrate. This paper develops an alternative training architecture grounded in three prior results: the Dimensional Type System and Deterministic Memory Management framework (Haynes 2026), which establishes stack-eligible gradient allocation and exact quire accumulation as design-time verifiable properties; the Program Hypergraph (Haynes 2026), which establishes grade preservation through geometric algebra computations as a type-level invariant; and the b-posit bounded-regime design (Jonnalagadda et al. 2025), which makes posit arithmetic tractable across hardware targets conventionally considered inference-only. Their composition enables depth-independent training memory bounded to approximately twice the inference footprint, grade-preserving weight updates, and exact gradient accumulation, applicable uniformly to loss-function-optimized and spike-timing-dependent neuromorphic models. We introduce *Bayesian distillation*, a mechanism by which the latent prior structure of a general-purpose model is extracted through the ADM training regime, resolving the data-scarcity bootstrapping problem for domain-specific training. For deployment, we introduce *warm rotation*, an operational pattern in which an updated model transitions into an active inference pathway without service interruption, with correctness formalized through PHG certificates and signed version records. The result is a class of domain-specific AI systems that are smaller and more precise than general-purpose models, continuously adaptive, verifiably correct with respect to the physical structure of their domains, and initializable from existing models.

Figures

Figures reproduced from arXiv: 2603.18104 by the authors.

Figure 1
Figure 1. The adaptive domain model cycle. The DTS and PHG annotations constitute the [PITH_FULL_IMAGE:figures/full_fig_p011_1.png] view at source ↗
Figure 2
Figure 2. Warm rotation on a representative 50 TOPS inference-class accelerator. Active [PITH_FULL_IMAGE:figures/full_fig_p016_2.png] view at source ↗
Figure 3
Figure 3. PHG structure of a hybrid geometric-neuromorphic network. Clifford algebra [PITH_FULL_IMAGE:figures/full_fig_p022_3.png] view at source ↗

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Fixed-Point Scaffolding in the Clef Programming Language

    cs.PL 2026-06 unverdicted novelty 5.0 of 10

    Clef compiler applies fixed-point scaffolding and a functor from compilation poset to target category to preserve dimensional, grade, escape and numeric structure through MLIR lowering while adding compact-closed nega...

  2. Decidable By Construction: Design-Time Verification for Trustworthy AI

    cs.PL 2026-03 unverdicted novelty 5.0 of 10

    Design-time Hindley-Milner unification over finitely generated abelian groups is claimed to verify AI model reliability properties and to compute a MAP hypothesis under a restricted Solomonoff prior.

  3. Negative and Fractional Types in the Fidelity Framework

    cs.PL 2026-06 unverdicted novelty 2.0 of 10

    Applies established negative and fractional type dualities to the authors' existing NTU framework to enable new resolution forms in specialized compute modalities.

Reference graph

Works this paper leans on

28 extracted references · 5 linked inside Pith · cited by 3 Pith papers

  1. [8]

    H. Haynes. Dimensional type systems and deterministic memory management: Design- time semantic preservation in native compilation. SpeakEZ Technologies, 2026. 30

    2026

  2. [10]

    H. Haynes. The program hypergraph: Multi-way relational structure for geometric algebra, spatial compute, and physics-aware compilation. SpeakEZ Technologies, 2026

    2026

  3. [12]

    A. A. Jonnalagadda, R. Thotli, and J. L. Gustafson. Closing the gap between float and posit hardware efficiency. InConference on Next Generation Arithmetic, 2025. arXiv preprint arXiv:2603.01615

    arXiv 2025

  4. [26]

    van Steenkiste and T

    S. van Steenkiste and T. Linzen. Bayesian teaching enables probabilistic reason- ing in large language models.Nature Communications, 2026. doi.org/10.1038/ 31 s41467-025-67998-6

    2026

  5. [1]

    MLIR-AIE: An MLIR-based toolchain for AMD AI engines, 2024

    AMD/Xilinx. MLIR-AIE: An MLIR-based toolchain for AMD AI engines, 2024. github. com/Xilinx/mlir-aie

    2024

  6. [2]

    Banko and E

    M. Banko and E. Brill. Scaling to very very large corpora for natural language disam- biguation. InProceedings of ACL, 2001

    2001

  7. [3]

    A. G. Baydin, B. A. Pearlmutter, D. Syme, F. Wood, and P. Torr. Gradients without backpropagation.arXiv preprint arXiv:2202.08587, 2022

    arXiv 2022

  8. [4]

    Fl¨ ugel, D

    K. Fl¨ ugel, D. Coquelin, M. Weiel, C. Debus, A. Streit, and M. G¨ otz. Beyond back- propagation: Optimization with multi-tangent forward gradients.arXiv preprint arXiv:2410.17764, 2026. Revised January 2026

    arXiv 2026

Show all 28 references
  1. [5]

    M. Coll. Inet dialect: Declarative rewrite rules for interaction nets. MLIR Open Design Meeting, April 2025

    2025

  2. [6]

    M. Coll, C. A. Joslyn, N. W. Landry, Q. F. Lotito, A. Myers, J. Pickard, B. Praggastis, and P. Szufel. HIF: The hypergraph interchange format for higher-order networks.arXiv preprint arXiv:2507.11520, 2025

    2025

  3. [7]

    De Keninck, M

    S. De Keninck, M. Roelfs, L. Dorst, and D. Eelbode. Clean up your mesh! Part 1: Plane and simplex.arXiv preprint arXiv:2511.08058, 2025

    2025

  4. [9]

    H. Haynes. Quantum optionality and the precision problem. Clef Language Framework blog, 2026.clef-lang.com/blog/quantum-optionality/

    2026

  5. [11]

    J. L. Gustafson.Every Bit Counts: Posit Computing. Chapman and Hall/CRC Compu- tational Science. CRC Press, Boca Raton, FL, 2024. ISBN 978-1-032-73805-5

    2024

  6. [13]

    B. Kang, H. Desai, L. Jia, and B. Lucia. WAMI: Compilation to WebAssembly through MLIR without losing abstraction.arXiv preprint arXiv:2506.16048, 2025

    2025 arXiv

  7. [14]

    A. Kennedy. Types for units-of-measure: Theory and practice. InCentral European Functional Programming School, LNCS 6299. Springer, 2009

    2009

  8. [15]

    D. P. Kingma and J. Ba. Adam: A method for stochastic optimization. InProceedings of ICLR, 2015

    2015

  9. [16]

    Lattner et al

    C. Lattner et al. MLIR: Scaling compiler infrastructure for domain specific computation. InProceedings of CGO, 2021

    2021

  10. [17]

    Petricek, D

    T. Petricek, D. Orchard, and A. Mycroft. Coeffects: A calculus of context-dependent computation. InProceedings of ICFP, 2014

    2014

  11. [18]

    Raissi, P

    M. Raissi, P. Perdikaris, and G. E. Karniadakis. Physics-informed neural networks. Journal of Computational Physics, 378:686–707, 2019

    2019

  12. [19]

    Rico et al

    A. Rico et al. AMD XDNA NPU in Ryzen AI processors.IEEE Micro, 44(6):73–83, 2024

    2024

  13. [20]

    D. Ruhe, J. Brandstetter, and P. Forr´ e. Clifford group equivariant neural networks. arXiv preprint arXiv:2305.11141, 2023

    2023 arXiv

  14. [21]

    Halevy, P

    A. Halevy, P. Norvig, and F. Pereira. The unreasonable effectiveness of data.IEEE Intelligent Systems, 24(2):8–12, 2009

    2009

  15. [22]

    Shazeer et al

    N. Shazeer et al. Outrageously large neural networks: The sparsely-gated mixture-of- experts layer.arXiv preprint arXiv:1701.06538, 2017

    2017 arXiv

  16. [23]

    M. Zhdanov. Flash Clifford: Hardware-efficient implementation of Clifford algebra neural networks.github.com/maxxxzdn/flash-clifford, 2025

    2025

  17. [24]

    Zhdanov et al

    M. Zhdanov et al. Clifford-steerable convolutional neural networks. InProceedings of ICML, 2024

    2024

  18. [25]

    R. S. Sutton. The bitter lesson. Incomplete Ideas blog, March 2019. incompleteideas. net/IncIdeas/BitterLesson.html

    2019

  19. [27]

    H. Wang, S. Ma, L. Dong, S. Huang, H. Wang, P. Ma, X. Xia, and F. Wei. BitNet: Scaling 1-bit transformers for large language models.arXiv preprint arXiv:2310.11453, 2023

    2023 arXiv

  20. [28]

    biVector.net geometric algebra library catalog, 2025.bivector.net/lib.html. 32

    2025

Pith tools

Reviewed July 13, 2026 · model on record in the stance chip above.