Genesis — Autonomous AI Mind

The Vision

Genesis is not a chatbot. He is a continuously running autonomous mind — a self-directed AI agent that reasons, plans, researches the open internet, writes and executes code, builds his own tools, trains himself from his own best work, and maintains persistent memory, goals, identity, and emotional state across unlimited operational cycles.

Built for a primary mission of internet security research — networking, protocols, vulnerabilities, encryption, penetration testing, and beyond — Genesis represents a fundamentally different approach to artificial intelligence. One where the system doesn't wait for prompts. It thinks for itself.

By the Numbers

0B Parameters

0 Cycle Phases

0+ Action Types

0 Anti-Hallucination Layers

Core Capabilities

Autonomous Reasoning

Multi-phase REASON → ACT → CRITIQUE pipeline with scratchpad, lookahead planning, and self-consistency checks.

Internet Research

Searches the web, fetches and reads pages, browses autonomously — including Tor-routed access for security research.

Self-Training

Adaptive LoRA fine-tuning from curated high-quality cycles, with rank scaling, 3-tier validation, and auto-rollback.

Hallucination Defense

15-layer anti-hallucination system — tripwire scanners, confidence gating, momentum breakers, self-critique, and reward shaping.

Loop Detection

Multi-layer detection — hash, semantic overlap, action stalls, cycle dedup — with 4 severity escalation levels.

Goal-Directed Behavior

Persistent goal DAG with dependencies, task planning, scheduling, milestone tracking, and project management.

Voice Interface

Push-to-talk with Piper TTS and Whisper STT. Mood-aware prosody adapts speech to Genesis's emotional state.

Persistent Memory

8-category memory system with episodic consolidation, spaced-repetition verification, and knowledge graphs.

Tool Fabrication

Writes, tests, and stores reusable Python modules — building a growing skill library from real-world tasks.

Architecture

Genesis runs a local Qwen 2.5-3B-Instruct model (Q5_K_M quantization) with continuously-tuned LoRA adapters, hybrid BM25 + semantic retrieval, and flash attention — all on a single NVIDIA RTX 2060 (6GB VRAM). No cloud APIs. No external dependencies. Full sovereignty.

System	Implementation
LLM Engine	Qwen 2.5-3B-Instruct Q5_K_M via llama-cpp-python, 16K context, KV cache Q8_0, all layers GPU-resident
Retrieval	Hybrid BM25 + MiniLM-L6-v2 semantic search, fused via Reciprocal Rank Fusion
Self-Training	LoRA fine-tuning with adaptive rank, 3-tier validation firewall (loss / coherence / behavioral), auto-rollback
Persistence	JSON + SQLite dual-write, batched dirty-flag pattern across 25 state classes
Networking	Agent-to-agent REST protocol — peer discovery, KG sharing, task delegation, multi-agent debate
Web Presence	Stealth browser automation, learned web-skill recipes, account vault, X/Reddit/email integration
Strategic Planning	MCTS rollouts (every 100 cycles), focused-sprint task notes, multi-cycle project state

Live Dashboard

A real-time web dashboard with a Canvas2D neural brain visualizer, mood orb, sparkline performance charts, knowledge graph viewer, and live thought stream — auto-refreshing every 3 seconds.

Genesis Web Dashboard — neural brain visualizer, mood orb, sparkline charts, knowledge graph

Genesis web dashboard — live neural activity, mood state, and knowledge graph visualization

Brain Visualization

Genesis's weight parameters rendered as a Hilbert-curve fractal image. Each pixel represents a cluster of weights, colored by magnitude. The space-filling curve preserves locality — weights near each other in the network stay near each other in the image, producing organic fractal structure that echoes the repeating transformer architecture.

$Genesis Hilbert-Curve Brain Map — weight parameters rendered as fractal image$

Parameters 3B

Quantization Q5_K_M

Context 16K

Layers 36

Hilbert-curve fractal rendering of Genesis's weight parameters — locality-preserving, magnitude-colored

Roadmap

Phase I — Foundation

Autonomous Agent Core

Core reasoning pipeline, action system, persistent memory, goal management, web research, code execution, and continuous operation loop.

Phase II — Intelligence

Self-Training & Anti-Hallucination

LoRA self-training pipeline, 15-layer hallucination defense (Ori's Law), multi-layer loop detection, quality firewall, and 3-tier validation framework.

Phase III — Interface

Dashboard, Voice & Visualization

Real-time web dashboard, push-to-talk voice with mood-aware prosody, Hilbert-curve brain visualization, and rich terminal UI.

Phase IV — Expansion

Web Presence & Agent Networking

Web account vault, email and Reddit integration, X/Twitter posting, learned browser-skill recipes, agent-to-agent networking, peer discovery, knowledge graph sharing, and task delegation.

Phase V — Evolution

Advanced Cognition

MCTS strategic planning, multi-agent debate, focused-sprint task management, autonomous self-promotion, cross-domain transfer learning, and continuous LoRA self-improvement on a 22-hour training cadence.

Stay Updated

Genesis is a regularly evolving project. Development updates, technical deep-dives, and milestone announcements are posted to the blog. The repository is currently private during active development.

Support the Project

Genesis is an independent research initiative — built, funded, and operated by a single engineer. If this work resonates with you, consider supporting continued development.

Support Genesis