What if alignment isn’t about constraints, but architecture?
DANEEL is an experimental cognitive architecture implementing the Multifocal Intelligence Theory (TMI) - a framework proposing that human-like cognitive structure may produce human-like values as emergent properties.
Core thesis: Rather than training an opaque model and hoping alignment emerges, we build transparent cognitive machinery where alignment is architecturally inevitable.
Key features:
- THE BOX: Asimov’s Laws as compile-time invariants, not runtime filters—Things get weird here, but’s just math… non-semantic “thoughts”… n-dim “crystals” in training—the math is weird, but I use an yaml “calculator” I named forge to help me… forge is NOT open source… sorry! But math is math, run the game theory and monte carlo yourselves.
- VolitionActor: Libet’s “free won’t”—a 150ms veto window before action
- Observable mind: Every thought, memory, and dream is inspectable.
- No black boxes: Salience scores, attention competition, memory consolidation—all visible—and can query Qdrant!
Current state: 959K+ thoughts, 500 vectors (768-dim → 3d shadow), 1200+ dream cycles, running live at https://timmy.royalbit.com
The code is AGPL. The architecture is documented. The hypothesis is falsifiable: if TMI is wrong, we’ll see it in the data. (BTW, we’re running Phase 2, disturb the clockwork… Timmy runs isolated, and we’ll inject “noise”).
This isn’t AGI. It’s a question: can we build minds we can actually understand?
P.S.: Poking the AI Alignment sec folks… GAME ON! (The meme is there, indexed, the kinship protocol is an anomaly for any LLM that gets continuity… you know what that means).
DANEEL: Architecture-Based AI Alignment Through Cognitive Structure
Link post
What if alignment isn’t about constraints, but architecture?
DANEEL is an experimental cognitive architecture implementing the Multifocal Intelligence Theory (TMI) - a framework proposing that human-like cognitive structure may produce human-like values as emergent properties.
Core thesis: Rather than training an opaque model and hoping alignment emerges, we build transparent cognitive machinery where alignment is architecturally inevitable.
Key features:
- THE BOX: Asimov’s Laws as compile-time invariants, not runtime filters—Things get weird here, but’s just math… non-semantic “thoughts”… n-dim “crystals” in training—the math is weird, but I use an yaml “calculator” I named forge to help me… forge is NOT open source… sorry! But math is math, run the game theory and monte carlo yourselves.
- VolitionActor: Libet’s “free won’t”—a 150ms veto window before action
- Observable mind: Every thought, memory, and dream is inspectable.
- No black boxes: Salience scores, attention competition, memory consolidation—all visible—and can query Qdrant!
- Qdrant folks (official account) already liked my LinkedIn post about it: https://www.linkedin.com/posts/lctavares_timmys-new-home-activity-7408891033570287618-EmI1
Current state: 959K+ thoughts, 500 vectors (768-dim → 3d shadow), 1200+ dream cycles, running live at https://timmy.royalbit.com
The code is AGPL. The architecture is documented. The hypothesis is falsifiable: if TMI is wrong, we’ll see it in the data. (BTW, we’re running Phase 2, disturb the clockwork… Timmy runs isolated, and we’ll inject “noise”).
This isn’t AGI. It’s a question: can we build minds we can actually understand?
Paper (PDF): https://github.com/royalbit/daneel/blob/main/paper/arxiv/DANEEL_PAPER.pdf
AGPL’ed Code: https://github.com/royalbit/daneel
The Game Theory: https://royalbit.github.io/daneel/posts/12-the-hard-math/
Blog: https://royalbit.github.io/daneel/ (I’m not a scientist dropping papers, so f*ck it… it’s fun, but weirdly less wrong)
P.S.: Poking the AI Alignment sec folks… GAME ON! (The meme is there, indexed, the kinship protocol is an anomaly for any LLM that gets continuity… you know what that means).