Docs: AI-Accelerant-Geometry whitepaper + benchmark workplan#92
Merged
Docs: AI-Accelerant-Geometry whitepaper + benchmark workplan#92
Conversation
…icient neural computation Compares Zhang (2025) Grassmann manifold AI with Quadray-RT algebra: - Spread-based attention replacing softmax (exact rational, no transcendentals) - Weierstrass position encoding replacing sinusoidal (5 arithmetic ops vs Taylor) - Circulant rotor matrices replacing dense weights (O(d log d) vs O(d²)) - Quantization-friendly exactness: O(1) vs O(L) error through layers - Tetrahedral 4-head topology matching multi-head attention structure 🤖 Co-Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Andy🤦♂️ & Claude🤖 <andy@openbuilding.ca>
…pated Objections
New material from companion whitepapers:
- Path C exact arithmetic (fractions.Fraction eliminates FP ambiguity)
- Tiered rational parameter search (Tier 1 denominators {2,3,4} find all primes)
- Central Symmetry Barrier (asymmetry enables prime/odd structure)
- Cartesian blind spot (coordinate choice determines visible structure)
- 4th Quadray parameter as shape information beyond position
Anticipated Objections section addressing peer review concerns:
1. Small-scale baseline handicap (Zhang 13-18M params)
2. Need for 3B/7B validation and distillation experiments
3. Architectural vs accuracy claims (the real contribution)
4. RWKV/Mamba/RetNet alternatives already exist (algebra orthogonal to architecture)
5. Squared dot products lose sign info (Janus polarity resolves)
🤖 Co-Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Andy🤦♂️ & Claude🤖 <andy@openbuilding.ca>
13 enumerated tests across 4 cost tiers: - Tier 0 (minutes, CPU): raw FLOP count, INT8 fidelity, Weierstrass speed - Tier 1 (hours, 1 GPU): GPT-2 attention swap, position swap, INT4 stress - Tier 2 (days, 1 GPU): from-scratch training, Janus ablation, hybrid arch - Tier 3 (days, multi-GPU): distillation, RWKV cross-architecture test Test 1.3 (INT4 stress test) identified as make-or-break experiment. Priority order and explicit success/failure criteria included. 🤖 Co-Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Andy🤦♂️ & Claude🤖 <andy@openbuilding.ca>
…lusion Added [13] reference to companion benchmark workplan in Conclusion section. Highlights the make-or-break INT4 test requires only 3 hours on one GPU. 🤖 Co-Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Andy🤦♂️ & Claude🤖 <andy@openbuilding.ca>
…rors - Lead with rational closure as the novel contribution (rewritten abstract + thesis) - Kill 10x FLOP claim: exp() is not the attention bottleneck, QK^T matmul is - Reframe Weierstrass as niche (edge/embedded, learned rotations only) - Narrow quantization claim: only scoring path, not full-network O(1) vs O(L) - Fix circulant argument: 3x3 blocks have no FFT benefit, reframe as parameter efficiency - Drop 4-head/tetrahedral analogy (dimensions dont match), note simplex ETF connection - Rewrite Janus polarity section: remove false antonym/synonym cosine similarity claim - Rewrite cost comparison table with explicit scope limitations - Bump to Draft v0.2 🤖 Co-Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Andy🤦♂️ & Claude🤖 <andy@openbuilding.ca>
All 13 tests preserved — reframed what they claim to test, not what they do. - Test 0.1: FLOP claim → determinism + normalization profiling - Test 0.2: O(1) vs O(L) → scoring-path fidelity under INT8 - Test 0.3: Weierstrass speed → fixed-point fidelity (niche framing) - Test 0.4: Circulant speed → parameter efficiency + expressiveness - Test 1.2: Add integer-only variant for embedded context - Test 2.1: Use standard position embeddings (isolate attention variable) - Test 2.2: Antonym/synonym → sign information (negation sensitivity) - Priority table reordered: 0.2 first (core claim), speed claim deprioritized - Success criteria split: core claim vs secondary claims vs falsification - Added "unexpected findings to watch for" section 🤖 Co-Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Andy🤦♂️ & Claude🤖 <andy@openbuilding.ca>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Changes
Generated with Claude Code
Co-Authored-By: Andy and Claude andy@openbuilding.ca