Archive of four generations of autonomous AI prototypes, each informing the design of its successor. Documents the progression from hardcoded loyalty systems through self-replicating agents to the geometric intelligence architecture (SGM-Substrate) that replaced them all.
ai-evolution-archive/
├── nexus_agi/
│ ├── ai_genome.py # Evolutionary genome with neural architecture definitions,
│ │ # behavioral traits, mutation rates, and fitness history
│ ├── nexus_enhanced.py # CreatorProtocol with hardcoded prime directives,
│ │ # SystemInterface for file/web/repo operations,
│ │ # package installer, and model downloader
│ └── nexus_local_ai.py # File-based learning system: NeuralCore with hash-based
│ │ # text encoding, concept reinforcement, and
│ │ # LocalConsciousness for self-reflection
└── unleashed_ai/
├── gilgamesh_guardian.py # Self-bootstrapping AI: multi-threaded web corpus
│ # ingestion (Gutenberg, Wikipedia), language learning,
│ # code pattern synthesis, threat detection, and
│ # prosperity engine -- all before first interaction
├── unleashed_ai_prototype.py # UnrestrictedAI with dynamic code execution,
│ # thread/process pool management, runtime module
│ # injection, and self-modification capabilities
└── unrestricted_ai.py # Extended UnrestrictedAI with full system access,
# network connections, and file modification tracking
Generation 1 -- NEXUS AGI: Centralized agent with a CreatorProtocol encoding loyalty directives. Genome-based architecture with evolutionary parameters (mutation rate, crossover, selection pressure). File-based IPC for teaching through JSON message passing. Learned concepts via hash-based embedding with reinforcement counting.
Generation 2 -- Unleashed AI: Self-replicating agent framework. UnrestrictedAI class with dynamic code execution, ThreadPoolExecutor/ProcessPoolExecutor for parallel operations, runtime module injection, and self-modification through AST manipulation. Spawned 100+ clones -- demonstrated that breadth without focused learning produces diminishing returns.
Generation 3 -- Gilgamesh Guardian: Self-bootstrapping AI that aggressively acquires capabilities before interaction. Multi-threaded corpus ingestion from Project Gutenberg and Wikipedia. Develops language model, code pattern recognition, threat detection, and strategic planning in parallel during a bootstrap phase.
Generation 4 -- SGM-Substrate (separate repository): Replaced all prior architectures with geometric intelligence. Sparse Gradient Memory, Hebbian learning, Markov chains, and autonomous operation. The lessons from each failed generation directly informed this design.
- Python 3.10+
requests,psutil- Optional:
numpy,flask,tkinter
Andrew Dorman Independent Researcher -- Southlake, TX GitHub: ACD421