AI/ML●●Solid
We beat Gemini Embedding 2 by training only 16M params (open weights)
16M trained params beats Gemini Embedding 2 on audio-text retrieval.
Big BrainNiche Gem
abtonmoy
7019d ago
Back to browse
GitHub Repository
Harmonic coherence as a universal relationship detection operator
9 starsPython
Harmonic embeddings beat random init and work frozen – no tokenizer
by atech-77·Feb 16, 2026·2 points·1 comment
AI Analysis
●MidBig BrainBold Bet
Ancient geometry meets Fourier analysis—neat math, but application to LLMs and databases unproven.
Strengths
- •Novel mathematical insight: cos(n×(θ_a − θ_b)) unifies symmetric/asymmetric/fuzzy relationships in one operator
- •Reproducible validation: 25 tests with corrective findings, code in Rust and Python
- •Connects database query operators to transformer attention structure—conceptually interesting bridge
Weaknesses
- •No implemented system: no benchmarks against SQL JOINs or actual LLM integration; theoretical only
- •Unclear how non-geometric relationships scale beyond lookup tables; unbounded frequency claim needs proof
Category
Target Audience
ML researchers, database theorists, LLM architecture explorers
Similar To
Positional encoding schemes (RoPE, ALiBi) · Fourier feature networks · Vector similarity search (FAISS, Pinecone)
Similar Projects
Developer Tools●●Solid
Graph-Oriented Generation – Beating RAG for Codebases by 89%
Replaces vector RAG with dependency graphs; 89% fewer tokens but benchmark setup is contrived.
Big BrainWizardry
dchisholm125
1224mo ago
Design●Mid
Opening lines of famous literary works
Beautifully typeset quotes, but a static list lacks technical depth.
CozyEye Candy
plicerin
18310615d ago
AI/ML●●●●Gem
A 6M-token movable window on a single 46GB GPU
6M-token context window on one GPU when vLLM caps at 30K tokens.
WizardryZero to OneBig Brain
Wetime
6121d ago
Education●●Solid
Interactive visualizer for Karpathy's 243-line microGPT
Type a name and you can literally watch characters turn into IDs, 16‑dim embeddings get added with positional encodings, and causal attention matrices animate per head — all matched numerically to Karpathy's 244‑line microGPT. The implementation is pure TypeScript (no ML libs) and includes a helpful scrollable sidebar with the reference math, which makes this an excellent, low‑friction learning tool — more pedagogical deep dive than research innovation.
Rabbit HoleNiche GemEye Candy
Sayyed23
115mo ago
Developer Tools●●Solid
ngrep – grep plus word embeddings
Semantic grep with word embeddings when traditional grep only does syntax.
Big BrainWizardry
xnan
324mo ago