GWT (Baars/Dehaene) starts directly from a mechanism (competition → ignition → broadcast) without ever laying down objective criteria for what consciousness is — unlike fields such as biology, where "life" has identifiable building blocks (autonomy, reproduction, evolution) even without a strict definition. GTT starts from the opposite requirement: extract externally verifiable building blocks before ruling on the mechanism, and explicitly accept that we don't know where consciousness sits — only what a candidate system must exhibit.
The system is not a monolithic block — it is composed of distinct, specialized subsystems (attention, memory, action) that communicate with each other. Consensual across all theories (GWT, PP, RPT), even when they diverge on how the communication works.
The system has a representation of its own processing, not only of the external world — the ability to evaluate, monitor, and adjust its own processes. This is the block Graziano (AST) places at the center: consciousness as a by-product of self-modeling, not as a selection mechanism in itself.
A continuous mechanism (not necessarily winner-take-all competition) that ranks what deserves processing/resources right now. Explicitly not GWT-style competition — closer to a continuous gradient (variance, Friston-style precision-weighting) than to a binary switch.
The system maintains a thread between the current state and past states — without it, no narrativity, no usable working memory. Distinct from block 3: prioritization operates at instant t, continuity operates across time.
The system can modify its own structure while it operates, not only in a separate training phase that is then frozen. This is the sharpest point of divergence from classical LLM architecture (weights frozen at inference) — and potentially GTT's most original block relative to the existing literature, which generally does not make it an explicit criterion of consciousness.
Explicitly not competition in the GWT sense. A mix of the first four blocks operating in parallel, with continuous modulation rather than discrete selection/broadcast events. Still fuzzy at this stage — this is the least formalized point of GTT at present, and needs to be specified.
GTT does not aim to explain why a system would produce felt experience (the hard problem remains explicitly open — neither solved nor bypassed by a sleight of hand of the "feeling is just another loop" kind).
The goal is functional and operational: to serve as the conceptual foundation for building a useful exocortex — a real augmentative cognitive system, not a proof of artificial consciousness. The 5 blocks are engineering criteria as much as theoretical ones: a system that exhibits all of them is a cognitively richer and more useful system, regardless of whether it "feels" anything.
| Block | Corresponding BTM/OMEGA component |
|---|---|
| 1. Modularity | Overall BTM architecture (separate loops: attentional, convergence, sedimentation, metacognition) |
| 2. Metacognition | Metacognition System 3 loop (Monitor + Tuner, threshold adjustment) |
| 3. Dynamic prioritization | C3 Cube — low/high variance → Direct/Deliberative Mode |
| 4. Continuity | Organic memory ST/MT/LT, heat scoring |
| 5. Continuous plasticity | Backprop-free HyperNeurons — theoretical capacity to evolve continuously, unlike the LLM frozen after training that is used for formulation |
Point of caution already identified: block 5 is currently more of an open direction than an achievement in the present architecture — the formulation LLM (Mistral) remains frozen; only the C3/HyperNeuron has the potential for continuous plasticity. It remains to be verified whether this potential is actually exploited or merely latent in the current implementation.
- The core mechanism of block 3 (prioritization) is named but not formalized — "subtler than a competition" remains, to date, an intuition, not a model.
- Felt experience is deliberately not treated as a 6th block nor housed within any of the 5 — left open rather than prematurely closed.
- Identified risk to watch: do not let BTM (the implementation) become a proof of GTT (the theory) through drift — "it works" and "it is a true theory of cognition" are two distinct criteria that must be kept separate.
Blocked due to hardware limitations (need a new machine). I'm hunting for ways to fund the project between bug reporting, security bug bounties, and hackathons... I'll be back soon! ;D