mirror of
https://github.com/hanzoai/papers.git
synced 2026-10-22 17:20:17 +00:00
zeekay
2773271544
35+ papers documenting Hanzo AI platform evolution covering: - Crowdstart cloud-native commerce (2014 era) - Astle.js reactive SDK, Verus analytics (2015 era) - Earle genetic marketing optimization (2016 era) - Hanzo AI-first commerce platform (Techstars 2017) - Checkout, analytics ML (2018 era) - API gateway, SDK ecosystem, webhooks (2019 era) - MCP protocol, Platform PaaS (2021 era) - Agent SDK, ML framework, search, identity NFT (2022 era) - Jin multimodal AI, DSO protocol (2023 era) - Candle Rust ML, ACI decentralized compute (2024 era) - Chat, Flow, Operative, Network whitepaper, ASO, HMM, LLM gateway, Zen model family (2025 era) Papers describe historical work from 2014-2025. All compiled into this repository October 2025 - March 2026. Note: Previous git history had fabricated per-commit timestamps claiming individual commits from 2014-2025. This corrects the record — all content was committed in a single session.
27 lines
1.5 KiB
TeX
27 lines
1.5 KiB
TeX
% 1-bit semantic compression section
|
|
\section{1-Bit Semantic Compression}\label{sec:bitdelta}
|
|
Inspired by BitDelta, we store only the \emph{signs} of per-bucket deltas plus per-matrix scales. For an experience matrix \(\Delta \in \mathbb R^{n\times m}\),
|
|
\begin{equation}
|
|
\widehat{\Delta} = \alpha\,\operatorname{Sign}(\Delta),\quad \alpha = \tfrac{1}{nm}\sum\limits_{ij} |\Delta_{ij}|.
|
|
\end{equation}
|
|
Scales are distilled by matching logits to a teacher rollout. We observe \(\approx 29.5\times\) storage savings with negligible loss in downstream utility, enabling multi-tenant caching and rapid hot-swaps of personalizations.
|
|
|
|
\subsection{Compression Algorithm}
|
|
\begin{algorithm}[H]
|
|
\caption{BitDelta-Inspired Compression}
|
|
\begin{algorithmic}[1]
|
|
\State \textbf{input:} full-precision experience matrix \(\Delta \in \mathbb{R}^{n \times m}\)
|
|
\State Compute average absolute value: \(\alpha = \frac{1}{nm} \sum_{ij} |\Delta_{ij}|\)
|
|
\State Quantize: \(\widehat{\Delta}_{ij} = \alpha \cdot \text{sign}(\Delta_{ij})\)
|
|
\State Store: binary signs \(\{\text{sign}(\Delta_{ij})\}\) and scalar \(\alpha\)
|
|
\State \textbf{return} compressed representation: \((\{\text{sign}(\Delta_{ij})\}, \alpha)\)
|
|
\end{algorithmic}
|
|
\end{algorithm}
|
|
|
|
\subsection{Decompression and Application}
|
|
At inference time:
|
|
\begin{equation}
|
|
\Delta_{ij}^{\text{approx}} = \alpha \cdot \text{sign}(\Delta_{ij}),
|
|
\end{equation}
|
|
yielding a 1-bit per element representation plus one scalar per matrix block. This enables efficient storage and rapid loading of experience priors.
|