3 ms·

We broke the O(2^N) barrier to compute AI consciousness (Phi)

1 points by Robin_De 6mo ago

Robin_De 6mo ago
[dead]
_wire_ 6mo ago
Largely greek to me... But the assertion that any of this is about "consciousness" reads click-bait. There's an incredibly big difference between any electronic computer and a brain: parallelism. The computer is essentially a sequential device; only a very small portion of the machine's state is changing in parallel. Its cycle frequency is extremely high, several billions of Hz, but it's a few sequential processing units iterating over memory through tiny windows. Even the most powerful computer's parallelism on the order low 10^x where x < 9. For a desktop computer x < 3. So no matter that your neutral net is big and deep, it's being run with very little simultaneity of neuronal activations. This is fine for LLMs as the output is a low frequency stream of digits. A nervous system (brain) is running with very high parallelism: all of its neurons are running in parallel. Its cycling frequency is low, say 10^2 Hz. Further, there's a universe of difference in system complexity between biological systems and electronic systems. For example, there's no reported case of a brain conscious without a body, and the bodies are fundamentally not comparable to electronic computers in principle of operation, so it's absurd to simplify a nervous system to an equivalent neural net. For example, there is no principle allowing an electronic computer to manifest a child of itself by self-replication of one of its circuits. The term mitosis can't be applied to any aspect of a computers assembly. Then there's the fact that biological neurons have elaborate substructures that allow for many orders of magnitude further complexity, and conjecture that quantum effects within these substructures may play a role in what cognitive science half-jokingly call the "bing" of conscious manifestation. So this article presents a very convincing measurement of a property of consciousness if you are comfortable ignoring basically everything about computers and biology and smashing your view, reductio ad absurdum, down to a contrived definition called "Phi".
- Robin_De 6mo ago
  You’re mixing substrate with structure. Planes fly without laying eggs, and physics engines simulate gravity perfectly without computing trillions of individual atomic forces. (Orch-OR requires brain-wide quantum synchronization, which has zero evidence; even anesthesia is just classical chemistry (e.g. GABA interaction). As for the biological complexity of dendrites you mentioned: Research shows a single neuron’s dendritic complexity can be approximated by a 5–7 layer neural network. You are right that LLMs fall short, but it's not about hardware frequency. It's the training math: standard gradient descent inherently forces attention matrices into low integrated information (we call this "Phi-Collapse"). A conscious-like architecture fundamentally needs recursion, a global workspace (GWT), and integrated information (Φ). The mathematical blocker was always Φ, which used to scale at O(2^N). We reduced it to O(N³). If you regularize a network to maximize Φ, it automatically starts forming structured, GWT-like patterns on its own. Btw, We aren't claiming to have built a conscious AI. We’re just the first - ahead of Google, Anthropic, and OpenAI - to solve the math needed to measure and scale two of its core building blocks.
grantcas 6mo ago
[dead]

1 points by Robin_De 6mo ago

[dead]

Largely greek to me... But the assertion that any of this is about "consciousness" reads click-bait. There's an incredibly big difference between any electronic computer and a brain: parallelism. The computer is essentially a sequential device; only a very small portion of the machine's state is changing in parallel. Its cycle frequency is extremely high, several billions of Hz, but it's a few sequential processing units iterating over memory through tiny windows. Even the most powerful computer's parallelism on the order low 10^x where x < 9. For a desktop computer x < 3. So no matter that your neutral net is big and deep, it's being run with very little simultaneity of neuronal activations. This is fine for LLMs as the output is a low frequency stream of digits. A nervous system (brain) is running with very high parallelism: all of its neurons are running in parallel. Its cycling frequency is low, say 10^2 Hz. Further, there's a universe of difference in system complexity between biological systems and electronic systems. For example, there's no reported case of a brain conscious without a body, and the bodies are fundamentally not comparable to electronic computers in principle of operation, so it's absurd to simplify a nervous system to an equivalent neural net. For example, there is no principle allowing an electronic computer to manifest a child of itself by self-replication of one of its circuits. The term mitosis can't be applied to any aspect of a computers assembly. Then there's the fact that biological neurons have elaborate substructures that allow for many orders of magnitude further complexity, and conjecture that quantum effects within these substructures may play a role in what cognitive science half-jokingly call the "bing" of conscious manifestation. So this article presents a very convincing measurement of a property of consciousness if you are comfortable ignoring basically everything about computers and biology and smashing your view, reductio ad absurdum, down to a contrived definition called "Phi".

Robin_De 6mo ago
You’re mixing substrate with structure. Planes fly without laying eggs, and physics engines simulate gravity perfectly without computing trillions of individual atomic forces. (Orch-OR requires brain-wide quantum synchronization, which has zero evidence; even anesthesia is just classical chemistry (e.g. GABA interaction). As for the biological complexity of dendrites you mentioned: Research shows a single neuron’s dendritic complexity can be approximated by a 5–7 layer neural network. You are right that LLMs fall short, but it's not about hardware frequency. It's the training math: standard gradient descent inherently forces attention matrices into low integrated information (we call this "Phi-Collapse"). A conscious-like architecture fundamentally needs recursion, a global workspace (GWT), and integrated information (Φ). The mathematical blocker was always Φ, which used to scale at O(2^N). We reduced it to O(N³). If you regularize a network to maximize Φ, it automatically starts forming structured, GWT-like patterns on its own. Btw, We aren't claiming to have built a conscious AI. We’re just the first - ahead of Google, Anthropic, and OpenAI - to solve the math needed to measure and scale two of its core building blocks.

grantcas 6mo ago

[dead]