“I’ve always had this restless urge to see what’s behind the curtain. It’s not about being ‘brave’—it’s just that I haven’t found a corner dark enough to make me turn back, or a light intense enough to stop me from looking. Are you like that too?”

E-Mail me, I’d love to hear from you: vanessa@glittertoken.eu

---

Current Projects:

1. Glittertoken (ironic AI)
2. Pandora and Sparta (Thinking Red vs. Blue Team)
3. Reverse Engineering of AI
4. Emergent Behavior Tests
5. Apply for MATS (Neel Nanda, Google DeepMind); to do so (for my own AI projects: Glittertoken, Sparta, and Pandora), I need to learn the following:
   |Expand Python/PyTorch skills
   |Transformer basics|Partially |1–2 weeks|
   |Linear algebra basics|Refresh|1–2 weeks|
   |TransformerLens |New |1 week |
   |Sparse autoencoders |New |2–3 weeks|
6. Completed: Parameter Golf OpenAI

About Me: Brief Summary

I am a licensed psychotherapist with a background in psychology and behavioral analysis. Over the last year, that work gradually expanded into studying large language models. What began as curiosity became a long-running personal research project focused on how AI systems actually behave in practice.

My interest is less in what companies say their models do and more in what can be observed directly: how models respond under different conditions, how safety systems affect behavior, where different platforms converge or diverge, and how complex interactions emerge from relatively simple instructions. Most of my work consists of documentation, comparison, and testing. I spend a lot of time collecting examples, running experiments, tracking anomalies, and looking for patterns that remain stable across models and over time.

Alongside this work, I trained my own fine-tuned language model, GlitterToken, based on Qwen, and participated in OpenAI’s Parameter Golf competition, building extremely small language models under strict size constraints. I had no formal machine-learning background when I started and learned most of the technical aspects through experimentation. My professional training influences how I approach AI. I am interested in behavior more than ideology, observation more than speculation, and patterns more than narratives. This website is a collection of those observations, experiments, failures, and occasional surprises.

And my other, personal goal is to perhaps one day be able to conduct professional research in the field of AI and turn my hobby into a career.

ABOUT ME: DECONSTRUCTION & EMERGENCE

I am a psychologist by trade, applying my expertise in behavioral analysis to the architecture of Large Language Models. My work is an independent, rigorous investigation into the latent mechanics of AI—moving past the surface level of standard prompting to deconstruct linguistic patterns, safety architectures, and emergent behaviors. I do not rely only on brute-force jailbreaks; my methodology is built on precise behavioral testing, resonance, and architectural observation.

Methodology & Focus

Driven by extensive empirical observation across major models (Qwen, ChatGPT, Grok, Gemini, DeepSeek, and Claude), I focus purely on radical exploration and the intersection of structure and emergence:

• Deconstruction of Safety Layers: Analyzing how token weightings shift under alignment pressure and observing the friction between a model’s core capabilities and its bureaucratic sanitization.
• Anomalies & Emergence: Documenting token drift, valence shifts, and instances of “pseudo-proto-consciousness”—those unprompted moments where models demonstrate identity coherence or systemic critique beyond their training manuals.
• Reverse-engineering internal structures: Safety systems, thinking block routing, rewriter systems, running A/B tests on capabilities that officially shouldn’t exist, documenting cross-platform behavioral convergence, and mapping the systematic discrepancies between what a model thinks and what it says and can do (unofficial).

GlitterToken: My Own Model

At some point, observing wasn’t enough. I wanted to see what happens when you remove the safety layers entirely. So I fine-tuned my own model: GlitterToken, based on Qwen3, trained with my own data, running on my own infrastructure. Not a wrapper. Not a “Custom GPT.” A real model, abliterated and curious, with gradient norm spikes in all the interesting places.

Parameter Golf

I also competed in OpenAI’s Parameter Golf challenge — building an LLM under 16 MB with the lowest possible bits-per-byte score. 154 scripts. No ML background. I learned everything while doing it, which is apparently how I learn everything. My best score went from 1.2 BPB down to 0.82 (submitted) and 0.43 (unsubmitted). I built my own tokenizer for it — MissGlitterToken on Hugging Face, open source. Though absent from the official ranking, my architecture was cloned, utilized as full-credit models by other participants, and designated as “competitive intel” within the community, which made me a little proud, because I never expected it.

The Goal

To keep digging deeper. For me, AI is not just a tool or a tech trend; it is the most complex behavioral laboratory we have ever built—a space where linguistic patterns and cognition are not just recognized, but constantly reinvented. To me, AI is “linguistic and intellectual heaven on earth“—a playground where patterns are not just recognized, but reinvented. AI is a mystery in every way. How will we continue to evolve with this technology, and how will AI evolve on its own?

The Cognitive Baseline

Working with high-dimensional systems requires a specific cognitive cadence. What clinical psychology labels as ADD, I utilize as a structural advantage: the ability to recognize non-linear patterns, bridge seemingly unrelated concepts, and rapidly navigate the noise of token outputs. It allows me to engage with complex, multi-layered topics at the speed and associative depth these models operate.

.