Two Paradigm Bridges: From LLMs to AGI

Eight pillars of AGI (MACI system-2 scaffold)

Overview

The talk urges new researchers to move past limitations of next‑token prediction. It first critiques the System 1 nature of current LLMs, then proposes two bridges: UCCT for semantic anchoring that shapes a task‑conditioned posterior, and behavior‑aware multi‑agent debate that regulates contention to convert information exchange into progress.

Critique of Current LLMs

System 1 without grounded semantics [02:11]

• LLMs are trained for next‑token prediction and operate as high‑dimensional pattern repositories.
• Patterns are not the same as semantics. Output often reflects frequency rather than meaning.

Patches that do not supply real reasoning

• Chain of Thought is described as a prompt pattern, not a formal reasoning method [03:32].
• RLHF is instance‑based and context‑dependent with a forgetting effect that undermines stability [04:23].

Fixed prior favors popularity over accuracy [13:05]

Training sets a prior over patterns. Without additional control, the model may converge to popular answers when they diverge from accurate ones.

Paradigm Bridge 1: Unified Cognitive Consciousness Theory (UCCT)

UCCT proposes adding System 2 control over System 1 by using semantic anchoring to shift the model from a fixed prior to a task‑conditioned posterior.

Semantic anchoring as a control handle [12:40]

• Provide concrete examples or intents as anchors that steer the distribution the model samples from.
• Anchors act like bait that select compatible latent patterns while suppressing incompatible ones.

Inferencing rather than formal reasoning [14:16]

Unconscious pattern repository [22:04]

• The pre‑trained model stores a vast unconscious of patterns gathered from large corpora.
• Few examples can form an on‑the‑fly classifier by activating relevant regions in this repository, similar to a child learning the concept of cat from a handful of labeled images.

Phase transition behavior [25:50]

• Anchoring exhibits a threshold effect. Once anchoring strength and pattern density exceed a level, behavior shifts sharply.
• Below threshold, outputs drift back to the prior; above threshold, the posterior is stable for the task.

Paradigm Bridge 2: Modulating Behavior in Multi‑Agent Debate

To turn debate into discovery, regulate behavior in addition to exchanging information.

Debate needs behavior control [31:42]

• Key variable is contentiousness. It governs whether agents explore, converge, or stall.

Contentiousness schedule [32:18] [34:04] [34:39]

• Too contentious: agents talk past each other without synthesis.
• Too agreeable: friendly chat lacks novelty and rigor.
• Target schedule: begin with high exploration, then lower contention to enable exploitation and integration.

Emulable tone and emotion [36:04]

• Training data contain stylistic and affective patterns. Models can emulate tone and emotional stance.
• Use this lever to shape debate dynamics and improve planning and reasoning outcomes.

Full summary