Files
zeekay 5f1d29cf0b papers: Zoo 4.0 launch retrospective (2026-02-14)
Adds explicit Triumvirate framing (DEX+EVM+FHE) tying Zoo into the
Lux-family canonical L1 configuration. Existing 11-section structure
preserved: thesis, L1 graduation, GPU-native execution, Zoo DEX V2/V3
+ V4 precompile to Lux DEX, AI Desktop with Zen models, zoo-bot,
Zoo Bridge multi-ecosystem, F-Chain FHE + Z-Chain ZKP privacy,
NFT migration program, wallet desktop, activation phases. Zoo
chronology preserved: 1.0 (BSC 2021, Worring/NFT Liquidity Protocol)
-> 2.0 (Lux L2 EVM PQ October 2024) -> 3.0 (full PQ 2025) ->
4.0 (GPU-native sovereign L1 Feb 14 2026). April 20 native securities
launch with Liquidity Protocol confirmed in section 11.

- zoo-4-0-launch/zoo-4-0-launch.tex (23 pages, 12 sections including new triumvirate)
- zoo-4-0-launch/sections/01a-triumvirate.tex (new)
2026-04-26 13:57:45 -07:00

79 lines
3.2 KiB
TeX

\section{\Zoo{} AI Desktop and Local Private AI}
\label{sec:ai-desktop}
\Zoo{} 4.0 ships a desktop application --- the \Zoo{} AI Desktop --- that
runs the \Zen{} model family locally on the user's machine. The
application is the user-facing entry point to Zoo's privacy-first AI
posture: weights live on the device; queries do not leave the device
unless the user opts in.
\subsection{Form factor}
The \Zoo{} AI Desktop is a Tauri 2 application (Rust backend, React
frontend) that auto-spawns the \texttt{zoo-node} runtime on launch. The
local node hosts:
\begin{itemize}[leftmargin=1.2em]
\item Inference runtimes: \texttt{llama.cpp}, \texttt{vLLM}, and
\texttt{ollama} backends, selected by hardware capability.
\item Model registry: the local copy of the \Zen{} family
(\texttt{zen4-nano} through \texttt{zen4-ultra-max}), gated by
device VRAM.
\item Wallet integration: \$AI mining rewards accrue to the
device's local wallet.
\item MCP server: agents in the desktop expose \Hanzo{} MCP tools for
local-first task automation.
\end{itemize}
\subsection{Free local inference}
A user with a recent consumer \GPU{} can run \texttt{zen4-nano},
\texttt{zen4-mini}, or \texttt{zen4-small} fully locally with no
network egress. Inference is free at the marginal cost of electricity.
The user's queries do not enter any logging surface, do not enter any
cloud telemetry pipeline, and do not enter any third-party datastore.
\subsection{Zoo Cloud upgrade path}
Users running larger models (\texttt{zen4-large}, \texttt{zen4-ultra},
\texttt{zen4-ultra-max}) that exceed local VRAM can opt into Zoo Cloud,
which dispatches to the federated capacity layer documented in
\cite{zoo-per-llm-chains} (\S Pooled \GPU{} Hosting). Zoo Cloud is
opt-in, pay-per-token, and TEE-attested: the cloud worker runs in an
H100 / TDX / SEV-SNP container, the inference receipt is anchored to
A-Chain, and the user can verify post-hoc that the response came from
the model the user paid for.
\subsection{Privacy posture}
The default privacy posture is local-only:
\begin{itemize}[leftmargin=1.2em]
\item No cloud calls without explicit user opt-in.
\item No telemetry without explicit user opt-in.
\item Wallet signing happens locally; the signing key never leaves
the device.
\item Conversation history is stored encrypted in a local SQLite
database (per Zoo Web5 / local-first architecture).
\end{itemize}
When the user opts into Zoo Cloud, the inference request is routed via
\Lux{} F-Chain \FHE{} (\S\ref{sec:privacy}) when confidential mode is
selected, so the cloud worker never sees the plaintext query.
\subsection{Mining UX}
The desktop's wallet pane shows live \$AI accrual from compute the
device contributes to the Hanzo network. A typical desktop with a
recent gaming \GPU{} can earn approximately one \$AI per minute under
default mining settings; rates depend on market demand for the
device's compute class.
\subsection{Why this matters for \Zoo{} 4.0}
The \Zoo{} 4.0 narrative is \GPU{}-native everywhere. The desktop app
is the user-side endpoint of that narrative: the same \GPU{}
architecture that runs validator-side consensus runs user-side
inference. \Zoo{} is the chain whose user runtime and validator
runtime share a substrate.