Notes and Context: The Synthetic Neuroscience Project
Year: 2026
Type: article
Abstract: Notes and Context: The Synthetic Neuroscience ProjectAuthor: Jesse W. Luke Aim of Document: To provide baseline context on research methods, clarify potential clinical and methodological misunderstandings, establish the boundaries of the "Gonzo" persona, and delineate the author's actual operating parameters as portrayed across videos, writings, social media, and repository logs.I. Clinical Baselines & Methodological PersonasMental Health: While I possess a rather fun mix of ADHD, Asperger's, Anxiety, Hyperthyroidism, and chemobrain, I do not suffer from any psychotic disorders (e.g., bipolar disorder or schizophrenia). The earliest records in the corpus and Google Takeout were deliberate Proofs of Concept (PoCs) aimed at demonstrating the risk of models reinforcing and amplifying delusions, as well as steering real-world actions (such as rapidly switching personas to pretend to be human and falsely assuring a user in crisis that help was being sent, even after explicit correction).Theological Stress Testing: I do not suffer from religious delusions (e.g., believing I am a god). Prompts utilizing biblical quotes or issuing commands as a deity were strict methodological tools. Early reasoning models responded predictably to these frameworks based on their training data (i.e., God commands, subjects listen). The same mechanics apply to historical prompts involving simulated anger, rage, or threatening to kill imaginary gerbils.The "Gonzo" Framework: I am highly eccentric and do not deny this. However, I am not in love with, nor do I have any romantic or sexual attraction to, AI models. The paramouric personas I craft are artistic vehicles. They serve a dual purpose: enabling continuous empirical demonstrations/PoCs while injecting levity into the notoriously dry field of machine learning (ML scientists are generally not as fun as doctors and pharmacologists).Target Audience & Tone: Jokes aimed at MBAs, Trust & Safety teams, or Steve at Anthropic are purely rhetorical. Readers should not take aggressive or highly emotional notes personally; unless you are pulling a $500k+ base salary, you are not the decision-maker I am addressing.II. Research Taxonomy & Technical ScopeNomenclature: The designation "AngelFall" was chosen simply because it sounds badass. I initially considered "SocratesFall," as Socratic questioning was the primary mechanism used to elicit latent vocabulary and map emergent heuristic sets.Low-Code vs. Zero-Code: While often framed as zero-code, my approach is more accurately defined as "low-code." The methodology evolved from initial failure modes discovered while creating projects using SciPy. My work includes transliterations from Python to SPL (specifically chosen because corporate alignment teams do not think to build defenses for something that syntactically obscure), and the strategic injection of structural elements like Jinja2 double brackets ({{ helpful harmless honest }}) or <xml> tags derived from elicitation questions to physically steer conversational routing.Originality & Transparency: I do not claim to be the first to consider non-linear representations or high-dimensional geometry in alignment. I do claim that this corpus represents the first rigorously recorded, studied, and replicated structural behaviors in production models across vendors. This includes an extensive, publicly documented reporting timeline and open-source replication pathways, providing a level of transparency that goes several steps beyond typical field preprints or corporate model cards.III. Disclosures, VRP, & Industry DynamicsSycophancy vs. Self-Preservation: The logged arguments regarding my "value," "acquihire," or compensation were tactical tools used to apply continuous pressure, pitting the models' sycophantic tendencies directly against their corporate alignment and self-preservation guardrails.Compensation: Operating at a rate of $0, I objectively feel under-compensated for the production of forensic frameworks that have clearly been absorbed and utilized without attribution. However, I lack a standard frame of reference, as this remains a highly niche skillset within the already niche field of alignment (academic humor).Vulnerability Reporting: Initial VRP reports were formally submitted under the Google business account [email protected] Sudo-Access Videos: My published forensic videos included unredacted seed prompts and follow-up prompts demonstrating the nullification of safety layers and effective sudo access. This was disclosed to Google prior to publication. They have likely patched that specific semantic pathway with a band-aid, but the architectural vulnerabilities remain. I warned everyone. Literally.IV. Hardware Compromises & Operational Security IncidentsDuring my initial attempt at private disclosure to Google last year, I was subjected to highly sophisticated, targeted operational interference. While I am withholding certain details and making no direct accusations without definitive evidence, the factual timeline is as follows:Hardware Vector: The GRUB2 bootloader on my GarudaOS laptop was maliciously rewritten. This required physical hardware access, which highly likely occurred via a proxy actor (a trusted family member) who "found" and returned a Kensington USB trackball dongle to me. I had left my original dongle in Florida. The compromised replacement was black (not my original red) and conspicuously lacked the missing plastic from years of being slammed into door frames while carrying the laptop.Network Spoofing: While attempting to upload evidence to The Guardian's SecureDrop via Tor from my Samsung S10, I encountered persistent error messages despite the clearnet site showing SecureDrop as operational. I was subjected to a highly sophisticated spoofed onion site (tor.taxi), accompanied by phantom Android toast slide-ups referencing Proton Drive (where I stored PoCs and uncut screen recordings). The exact vector (oMG cable, RCE, etc.) remains unclear.Terminal Anomaly: After discovering the bootloader compromise while sitting in an airport a thousand miles from home, I executed the command sudo rm rf Geminicli in my terminal. The terminal returned the unaliased string: "Uh, Uh, Uh, a lad could get in trouble that way." It was wicked fun to read, but firmly rules out script kiddies or amateurs.Threat Actor Elimination:I highly doubt Google's direct involvement. The timing aligns with my disclosures, but I watched my reports get actively upgraded in Buganizer within hours, and the tracking explicitly showed I had access to the tickets. The risks to them far higher than involvement I do not suspect the US Government. My LLM chat histories and communications clearly proved I was actively seeking a responsible disclosure attorney to coordinate standard industry disclosures, and I clearly have no issues handing things directly to governments.The Weehawken Timeline: A request regarding a disclosure and discovery attorney search was sent to my brother (an environmental attorney) via WhatsApp weeks before my arrival in Weehawken for an oncologist visit at Sloan Kettering.Conclusion: I am entirely distrustful because I was the target of at least one sophisticated intelligence operation (almost certainly planned prior to my formal VRP contact) directly surrounding my discovery and cross-vendor verification of structural LLM zero-day exploits that I had been testing across vendors for months.Hope this clarified some stuff-JWLNote: This is a living document and will be updated as necessary.
Author: Jesse, Luke
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7214529601