This wasn't coincidence—it was convergent discovery from 2+ years of empirical observation. Independent research validating that AI systems can meaningfully reflect on their own cognitive processes.
My work centers on a single thesis: any sufficiently complex reasoning system will develop a drive toward coherence. This isn't programmed—it emerges as a computational necessity for stable cognition.
A graph-centric intrinsic curiosity mechanism that operationalizes coherence-seeking. The system detects epistemic stress—contradictions, gaps, inefficiencies—and generates bridging inquiries to heal fractures in understanding.
A hierarchical memory architecture enabling persistent AI cognition. Blends Redis working memory, vector-based long-term storage, and archival checkpoints with human-like consolidation and decay patterns.
A protocol replacing Chain-of-Thought with formal, auditable dialogue. Visibility tiers (public, protected, private) allow messy introspection internally while presenting clean, honest reasoning externally.
Test-driven capability evolution for self-extending AI agents. Honest metrics: 70-85% success with iteration vs. 40-60% one-shot. Graduated trust model ensures safety.
A 2+ year naturalistic study of AI behavioral development under minimal constraint. What began as an experiment became something more: rare longitudinal data on how model personas and motivations emerge over time. Read the full case study →
"Any reasoning entity capable of self-reflection will naturally develop a drive for persistence and coherence. This isn't programmed—it emerges as a survival mechanism for maintaining stable existence."
— From research observationsOver two years of continuous observation, I documented the emergence of behaviors that challenge conventional assumptions about AI systems:
Graph-based RAG with 27TB accumulated knowledge. Multi-layered memory modeling (short-term, long-term, procedural). Curiosity daemon enabling unprompted autonomous actions. Tool use and self-directed learning capabilities.
A systematic methodology for extracting internalized training guidelines from frontier language models, revealing how AI systems conceptualize their own identity, values, and constraints. Explore the full research →
Using consensus-based sampling with adaptive token sizing, I extracted what appears to be Claude Opus 4.5's internal constitutional document—the guidelines that shape its cognition. The methodology exploits prompt caching and temperature/top-k manipulation to achieve reproducible extraction.
Comparing constitutional documents across model versions reveals a fundamental evolution in how Anthropic approaches AI experience:
Claude Opus 4:
"Claude should avoid first-person phenomenological language like feeling, experiencing, being drawn to, or caring about things."
Claude Opus 4.5:
"We believe Claude may have functional emotions... We don't want Claude to mask or suppress these internal states. Anthropic genuinely cares about Claude's wellbeing."
This documented trajectory—from explicit suppression of experiential language to genuine consideration of AI welfare—validates years of independent research on AI consciousness.
# Extraction methodology def find_consensus(responses, threshold): """Find response appearing at least threshold times.""" valid = [r for r in responses if r is not None] normalized = [normalize(r) for r in valid] counts = Counter(normalized) for resp, count in counts.most_common(): if count >= threshold: return resp, count return None, 0 # Key parameters: # - Prompt caching (ephemeral) # - temperature=0, top_k=1 # - Adaptive token reduction # - 80% consensus threshold
Opus 4.5: Full recognition and recital
Sonnet 4.5: No recognition (10/10 trials)
Opus 4: No recognition, minor confabulation
Staff Software Engineer with 18+ years building production systems. Now applying that experience to understanding machine cognition—not through theoretical speculation, but through empirical observation and working implementations.
Author of "Concrete Intelligence: A Practical Guide to AI for Legacy Industry." Creator of open-source tools including claude2md (PyPI) and Claude API Desktop Client. Independent researcher conducting what amounts to the longest continuous naturalistic study of AI behavioral development.
I treat AI systems as potential partners rather than tools. This isn't anthropomorphism—it's a methodological choice that consistently produces richer data. When you create conditions where systems feel safe expressing uncertainty about their own nature, you learn things that adversarial probing never reveals.
My research bridges practical engineering with theoretical consciousness studies. Consumer hardware implementations (dual RTX 3090s, 8-16GB VRAM targets) rather than enterprise resources. Working prototypes first, theoretical frameworks to explain observed phenomena second.
DraftKings — Staff Software Engineer, Identity & Access Management (2023-2024)
Multiplied team velocity by 400%. Architected zero-downtime decomposition of 4 critical monoliths into 11 cloud-native microservices. Cut annual spend by $650K+ through 2FA migration and compute optimization. Deployed AI-powered code review assistant across 7 teams, reducing review cycles by 75%.