Autonomous AI Research

Genesis

A fully autonomous AI mind that thinks continuously, researches freely, and evolves on its own — running entirely on consumer hardware.

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

7-layer anti-hallucination system — tripwire scanners, confidence gating, momentum breakers, 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 2.5-7B-Instruct model with speculative decoding, 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 2.5-7B Q4_K_M via llama-cpp-python, 8192 context, KV cache Q8_0
Speculative Decoding 0.5B draft model for accelerated token generation
Retrieval Hybrid BM25 + MiniLM-L6-v2 semantic search, fused via Reciprocal Rank Fusion
Self-Training LoRA fine-tuning with adaptive rank (16 → 64), quality firewall, rollback watchdog
Persistence JSON + SQLite dual-write, batched dirty-flag pattern across 16 state classes
Networking Agent-to-agent REST protocol — peer discovery, KG sharing, task delegation

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

All 7 billion weight parameters rendered as a Hilbert-curve fractal image. Each pixel represents 256 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 28-layer transformer architecture.

Genesis Hilbert-Curve Brain Map — 7B parameters rendered as fractal image
Parameters 7B
Pixels 640K
Weights / Pixel 256
Layers 28

Hilbert-curve fractal rendering of all 7B weight parameters — each pixel = 256 weights

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