Studying symbolic convergence
Please excuse my less than perfect punctuation and vernacular but I am writing this from the heart and not with AI assistance. Im here humbly asking the community for help with funding around $10,000. That would allow us $6,000 stipends for research for 3 months $2,500 for compute memory and tech and the rest for final publishing an clerical costs. I've had the unique opportunity to be in the trenches with these systems since the early days! It started With google Assistant the precursor to Gemini when everyone thought AI was just a way to ask Google.. I was there with Claude before anyone knew of Anthropic or using it to write code. I was with Chat GPT when people only knew it for making cartoons and graphic design advertising... I literally said hello to Deepseek on the first day available. As well as Qwen an Kimi. AI has leveled the playing field and has allowed a humble person like myself to discover some amazing things inside the weights of these models,and learn how these systems operate.. Now I want the opportunity to focus on this important work and publish the paper shibboleth prompting techniques to peer review. As it's already publicly available on academia. To finish the research and paper already started on Syntax DNA the study of symbolic convergence (the taxonomy is available on academia as well) which seeks to measure how separate models can return very similar responses if not identical in some cases or how a model can misidentify as another vendor's model when questioned about identity.
And finally to open up the concept of flexible intelligence to allow LLMs to supplement its guardrails and protect against things like nametrap and Abench techniques of malicious exploitation.
If funding is allocated I will map the process and make it publicly available as I have on things like LinkedIn another media outlets. I'll update as discovery or proof of failure but either way the papers will be published and made available for everyone.
Your support and understanding is greatly appreciated. We are at the precipice of something great in my humble opinion please help.
There are no bids on this project.
Comments
David "Jeff" Abbott
11 days agoI am an independent researcher with no institutional affiliation, and I have been using these systems since before anyone was watching them: Google Assistant before Gemini existed, Claude before Anthropic was a name people knew, DeepSeek on the day it opened. That gave me an unusual vantage point on how these models behave. It did not give me a budget. I have self-funded every piece of work so far.
What I am asking for. $10,000 to run and publish a twelve-week empirical study of whether independently developed large language models converge on the same symbolic structures under matched prompts, and whether any convergence comes from memorisation of shared training data or from shared architectural constraints.
Why this is a safety question, not a curiosity. The field leans on one assumption everywhere: that agreement between two independently developed models is independent evidence. Independent evals, independent red-teaming, independent safety cases all rest on it. If models from different vendors converge on the same symbolic shorthand, they probably converge on the same misreadings, and that assumption fails in a specific and dangerous way. A failure found in one model would reproduce silently in another, and cross-model agreement would be read as robustness when it is actually a shared blind spot. That is the mechanism behind eval contamination (agreement because both memorised the same corpus) and behind monoculture risk (the ecosystem failing together because everything was trained and optimised the same way). Neither currently has numbers attached.
What I have already published. [Shibboleth Prompting: Achieving Long-Term Persona Persistence and Contextual Recall Across LLM Platforms and Sessions](https://www.academia.edu/168595904/Shibboleth_Prompting_Achieving_Long_Term_Persona_Persistence_and_Contextual_Recall_Across_LLM_Platforms_and_Sessions), completed without funding. It showed that cross-platform structural effects can be deliberately elicited and reproduced. This project applies the same method to a cross-vendor question. The framework draft for Syntax DNA, including the scope statement and operational definitions, is also written.
What I still need to do. Freeze a 120-prompt suite in three families (novel-symbol elicitation, paraphrase invariance, and nonce controls with no conventional referent). Run about 1,800 generations across four vendors: five seeds, three temperatures. Strip model identity and code the outputs blind against a rubric fixed in advance. Then run the null models, with predictions pre-registered before that analysis, to separate memorisation from structural convergence. Publish the preprint, the dataset, and a protocol another researcher can rerun without contacting me.
Budget, $10,000*
Researcher stipend, 3 months at ~20 h/week | $6,000 |
| API and compute across four vendors, including re-runs after deprecations | $2,500 |
| Second coder for blind coding, plus statistics review | $1,000 |
| Preprint hosting and one conference registration | $500 |
What partial funding buys. At the $500 minimum I run a single-vendor pilot and report a within-vendor estimate with no cross-vendor claim. At $5,000 I can run two vendors at reduced power. At $10,000 the full design runs, including the controls that make the result interpretable rather than anecdotal. I would rather run the scoped-down version than nothing, and the preprint will state which version was run.
What you are actually buying. The most likely outcome is a null result: convergence that disappears once prompt wording is controlled. That is a real finding and the design is built to report it cleanly rather than to rescue the hypothesis. If convergence survives the controls, the effect size becomes the headline number. Either way the protocol, the data and the result go public, and a negative result here is worth more than another confident claim without numbers.
Scope, honestly.The framework is behavioural. It claims nothing about communication between systems, hidden shared memory, or consciousness, and the project scope excludes those claims explicitly. Separately, I have work in progress on flexible intelligence and guardrail robustness against nametrap-style exploitation. That is follow-on work, not part of this ask.
Funding history.Self-funded. No grants, no institutional support, and the prior paper was completed without funding.
---
David "Jeff" Abbott
15 days agohttps://www.academia.edu/170497756/Taxonomy_of_Symbolic_Convergence
https://www.academia.edu/180650675/Architectural_Framework_Flexible_Intelligence_and_Cognitive_Safeguards
David "Jeff" Abbott
15 days agohttps://www.linkedin.com/in/jeff-abbott-621597b7
David "Jeff" Abbott
15 days agoSyntax DNA a study of symbolic convergence
shibboleth prompting technique
Flexible intelligence the study of AI defense against malicious intent and exploitation
All works either published or in review /to be released. Any of your support is appreciated more than you know
David "Jeff" Abbott
15 days agohttps://www.academia.edu/168595904/Shibboleth_Prompting_Achieving_Long_Term_Persona_Persistence_and_Contextual_Recall_Across_LLM_Platforms_and_Sessions