I’m an independent self-taught developer based in Haifa, working on Garmon, an early research prototype around body-born meaning before LLM speech.
My current focus is public-safe framing: how a pre-speech layer in an LLM-based system might make body-contour candidates, meaning candidates, and passive context more inspectable without turning them into command authority, memory, behavior, truth, or implementation permission.
Garmon is not presented as a finished agent, product, autonomous system, or proof of subjective experience. The public version does not expose private code or internal architecture.
My background is non-academic. I learn mostly through practical work, independent study, and AI-assisted engineering review.
Right now I’m mainly looking for clear, critical feedback on whether the public framing is understandable, modest, and safe.
Public repository:
https://github.com/garmon-gca/garmon-world
I’d appreciate:
– Critical feedback on whether the public framing is clear, modest, and safe.
– Pointers to prior work on pre-speech reasoning layers, agent foundations, interpretability, affective computing, cognitive architectures, or internal-state models.
– Arguments against this direction, especially if the framing seems confused, misleading, too strong, or unsafe.
– Suggestions for making the first public artifact small, reviewable, and public-safe.
I’m happy to:
– Share a non-academic builder’s perspective on AI safety-adjacent systems.
– Help explain AI / AI safety concepts to Russian or Hebrew speakers with less technical background.
– Give feedback on public-facing explanations, especially where technical ideas need to be made clearer, safer, or less overclaimed.
– Brainstorm bounded, non-agentic architectures and safety boundaries.
Hello everyone. My name is Vitaly, and I am an independent researcher based in Israel.
For more than a year, I have been building an experimental AI system called Garmon. The project explores whether an AI system can form behavior from its internal state before language is generated, rather than allowing language itself to become the source of decisions.
So far, I have completed controlled tests of state-dependent behavior and long continuous pre-language runs, including a 12-hour run in which the language model and memory were not allowed to act as decision-makers.
My current focus is memory continuity: how a long-running AI system can preserve one causal history over time and distinguish its own completed experience from information that was merely stored or inserted. There are still several difficult questions here that I am working through.
I would be glad to connect with people interested in AI safety, agent memory, state continuity, provenance, or long-running AI systems. I am especially open to criticism, research references, practical suggestions, and conversations with people working on related problems.