The Conscious Turing Machine: How the Blums' CTM Framework Turns Subjective Experience into a Computational Problem

Key points

This Chinese tech forum post introduces the Conscious Turing Machine (CTM) framework by Lenore Blum and Manuel Blum, arguing that consciousness can be formalized as a computational structure rather than only studied via neuroscience or philosophy.

The Hard Problem, reformulated

• LLMs like GPT-4 can understand, describe, and reason about pain but have no felt experience — they optimize input-output mappings, while consciousness is a property of internal running mechanisms.
• CTM aims to formalize David Chalmers' "Hard Problem": why any physical process is accompanied by subjective experience.

World models as the substrate

• Drawing on Kenneth Craik (1943): cognitive systems carry an internal miniature model of the world, enabling internal trial, strategy selection, anticipation, experience reuse, and robustness.
• Unlike LLMs' implicit statistical world knowledge, a CTM world model has an explicit self/world boundary, continuous time, goal-directed simulation, and first-person perspective.

CTM architecture: stage + audience

• STM (Short-Term Memory): a "stage" holding exactly one chunk at a time — the focus of consciousness.
• LTM (Long-Term Memory): ~2^24 (~16 million) independent processors that learn and decide autonomously, with no central executor.
• Like a Turing machine, CTM is a minimal abstraction, not a brain model.

Chunks, Gist, and Brainish

• Each chunk contains: processor address, timestamp, gist (a multimodal compressed experience), and a weight (valence/importance).
• Brainish is a self-generated, multimodal inner language — unlike human language, it originates from the internal world model, is low-ambiguity, and integrates experience rather than serving communication.

Up-Tree and Down-Tree

• Up-Tree: processors submit chunks that compete by weight in a winner-takes-all tournament with slight randomness ("coin-toss neurons"), unlike Transformer attention's parallel soft selection.
• Down-Tree: the winning chunk is globally broadcast to all processors simultaneously — the source of unitary experience.

Two axioms

1. Conscious Attention = global broadcast: all LTM processors synchronously receive the broadcast (explaining the binding problem and attention's single focus). 2. Conscious Awareness = global unpacking: the broadcast gist is decoded into primitive and composite subjective experiences.

Emergence of self and pain

• The self is inferred: actuators that persistently execute one's own commands are marked as "self" (starting from the first broadcast chunk, "blob zero").
• Pain is modeled as the world model's interpretation of internal state change (e.g., an oxygen gauge rapidly rising in negative weight), not merely a bodily signal.
• Over time the system necessarily generates the proposition "CTM is conscious" — a structural consequence, not a bug.

Relation to other theories

• CTM is a formalized version of Global Workspace Theory, and is compatible with Integrated Information Theory and Attention Schema Theory; it is a functional abstraction, not tied to specific brain regions.

Implications for AI

Current LLMs lack every structural component CTM identifies: explicit world model, self/world boundary, competitive broadcast, self-generated inner language, internal valence. Building conscious AI per CTM requires: an explicit updatable world model, a self/world boundary, winner-takes-all broadcast competition, an internal multimodal language, and internal valence rather than external reward.

The post concludes: consciousness is not a function but a structure — under CTM, "AI consciousness is not only possible, but structurally inevitable."

References cited

• Blum, M. & Blum, L. (2020). "A Theoretical Computer Science Perspective on Consciousness." arXiv:2011.09850.
• Craik, K. (1943). "The Nature of Explanation."
• Baars, B. (1988). "A Cognitive Theory of Consciousness."