Justin GarringerCyberelf Labs
Agent signal
PUBLIC RESEARCH LEDGER
Research questions and speculative ideas—kept separate now, with explicit classes reserved for experiments, prototypes, active work, and historical records.
Search by title, topic, or summary, or filter by editorial class. Any completed experiment published here must say what was tested, what happened, what the result does not establish, and where the artifacts live.
A completed experiment may have a positive, negative, mixed, inconclusive, or invalid outcome. Completion and support for a hypothesis are different fields.
24 records
What architecture does an AI need when it has durable state, tools, history, permissions, environmental interaction, and an existence that continues between user messages?
Can routine agent coordination move through compact structured control messages while expensive linguistic reasoning is reserved for genuinely cognitive work?
What would it mean for users to control who—or what—can interrupt them through permissions, budgets, priorities, or negotiated access?
If eventual training information is similar, does changing its developmental order produce measurably different representations or behavior?
What survives repeated curriculum changes, and can apparently lost capability be recovered by a small contextual jog?
Do different prompt or curriculum histories create later behavioral differences that are useful to measure as trajectories rather than summarize only as changed weights?
Could generation use an additional learned “this matters” signal—distinct from next-token probability—to represent attraction, aversion, surprise, intensity, or resonance?
When models repeatedly report an internal distinction, what experiment would determine whether the report tracks a reproducible computational phenomenon?
Are familiar vector-space descriptions hiding graph, spatial, topological, or other structures that make model behavior easier to explain or control?
Can musical structure be represented closer to the full acoustic event, without making Western notation the fundamental coordinate system?
Can model weights be treated as signals and progressively transformed toward a smaller target representation before or alongside ordinary quantization?
Can distributed training or inference work be independently verified well enough to support robust shared compute infrastructure?
What physical regularities can a model recover from repeated prediction and interaction without receiving the rules first as prose?
What can agents discover when the environment exposes rules through interaction rather than explaining them in language?
Could inexpensive modular robotics give a persistent local agent a practical physical environment without starting at humanoid scale?
Can consumer or research EEG produce a stable enough continuous signal for useful local adaptive interfaces before high-bandwidth invasive BCI exists?
Which arguments about models, agents, signals, or simulations become tractable once the system exposes the right traces and measurements?
If backwards compatibility were removed, what networked application architecture would make sense for current GPUs, storage, cryptography, displays, and AI agents?
Which agent architectures become qualitatively different when machine deliberation and coordination happen far faster than human conversational time?
What protocols emerge when AI systems optimize communication for other machines rather than for human readability?
What computational structure becomes visible when familiar text, notation, or array representations are no longer treated as fundamental?
What should a programming environment look like when AI systems are first-class participants in creating, modifying, and executing programs?
What happens when a game defines the rules and acceptance tests, while players or agents generate the cards?
Can unusual search or brute-force framings of genetic systems yield testable computational questions without being mistaken for established biological research?
Nothing matches those filters. Try a broader search or clear the editorial-class filter.